SkillsLLM

From Tutoring Product to Skill Ecosystem: aitutors.me's Multi-Surface Plugin Strategy

2027-03-31T00:00:00+00:00

Claude tutors, a public plugin repo, standalone skills, a flashcards import format — how aitutors.me is building surfaces beyond its own subscription product.

What a Two-Skill Plugin Manifest Reveals About Product Priorities

2027-03-30T00:00:00+00:00

A product could ship dozens of skills through its public plugin repo. aitutors.me shipped exactly two. What that restraint actually signals.

Best AI Flashcard-Authoring Skills and Tools Compared

2027-03-29T00:00:00+00:00

Anki add-ons, Quizlet's AI generator, and aitutors.me's flashcards skill compared on what they actually do well — and where each one falls short.

Six Card Kinds, One QA Gate: What a Good Flashcard Skill Actually Encodes

2027-03-28T00:00:00+00:00

Most flashcard tools treat every fact the same way. Looking closely at why one skill refuses to shows what genuinely good study-material design requires.

The Identity-Anchor Pattern: Keeping an AI Character On-Model at Scale

2027-03-27T00:00:00+00:00

How a frozen reference sheet and a short list of never-change features solve the hardest problem in AI-generated brand characters — and how to apply the same pattern to your own.

Why One Skill Refuses to Be Redistributed (and Why That's the Right Call)

2027-03-26T00:00:00+00:00

aitutors.me ships two skills in one repo — one MIT, one 'do not redistribute.' The contrast is a useful lens on what actually belongs in the open part of a product.

Installing the AI Tutors Plugin in ChatGPT and Codex

2027-03-25T00:00:00+00:00

How to install aitutors.me's tutors and skills into ChatGPT or Codex CLI via its public plugin marketplace repo — step by step, plus what each install path actually gives you.

How to Turn Your Notes Into Flashcards With the AI Tutors Skill

2027-03-24T00:00:00+00:00

A step-by-step walkthrough of the aitutors.me flashcards skill: what to hand it, what it hands back, and how to avoid the mistakes its own QA gate catches.

AI Tutors Flashcards: Turn Any Topic Into Study Cards in Seconds

2027-03-23T00:00:00+00:00

A spotlight on the flashcards skill from aitutors.me's plugin repo — an MIT-licensed skill that turns notes into spaced-repetition cards ready to import.

zaptrace

2026-11-12T00:52:09+00:00

Verification-first ZapTrace workflows for local design validation, proof-pack review, and bounded benchmark evaluation.

overwatch

2026-11-12T00:36:26+00:00

Catches a coding agent misbehaving, records it and makes the next session better: tasks and guards hold every stop until the work is done and reviewed, every catch (an edit before a plan, a stop with work open, an invalid call, a failed turn) lands in a ledger, and repeated habits plus recorded mistakes ride the next session's prompt. Also in the plugin: compaction (minified tool results, old history folded into indexes, any of it read back by id), one gauge line above the prompt, the tool-surface ledger with a repo's verification checks, /loadout, /hygiene and the disk gardener (asks for a cleanup when the disk is nearly full, finds and deletes proven leftovers).

total-recall

2026-11-12T00:36:22+00:00

Multi-tiered memory and knowledge base with semantic search, auto-compaction, and built-in evaluation. Works across Claude Code, Copilot CLI, OpenCode, Cline, Cursor, pi, and Codex CLI.

mem-forever

2026-11-12T00:36:20+00:00

Every AI tool forgets you. This one doesn't. Ever. Your Git repo becomes a portable agent brain: profile, one file per memory, a hand-off note that says where you stopped, an encrypted vault for keys. Survives every session, every tool, every migration.

yana-ai

2026-11-12T00:36:18+00:00

Personal agent operating system for Claude Code. Safety hooks, memory tiers, agents, commands, and cross-engine hard enforcement.

dobeutech-claude-code-custom

2026-11-12T00:36:17+00:00

Comprehensive centralized repository of Claude Code configurations including agents, skills, commands, hooks, and MCP servers optimized for full-stack development

cc-sticky-notes

2026-11-12T00:36:14+00:00

cc-sticky-notes: route curiosity questions to a side pane instead of the main conversation

soundcraft-plugin

2026-11-12T00:36:14+00:00

Make, mix and render audio in SoundCraft from Claude: the soundcraft MCP server (live app or headless), plus skills for driving SoundCraft and for procedural audio (synthesis, granular processing, procedural variation, physics-driven sound).

Pink Agentic AI Payments MCP server

2026-11-12T00:22:25+00:00

Pink Agentic AI Payments MCP server

Shortform Kit MCP server

2026-11-12T00:22:25+00:00

Shortform Kit MCP server

vn-revit-mcp

2026-11-12T00:22:25+00:00

vn-revit-mcp

Keyboardino MCP app

2026-11-12T00:22:25+00:00

Keyboardino MCP app

Moon Phase Software MCP app

2026-11-12T00:22:25+00:00

Moon Phase Software MCP app

Get Chinese Astrology

2026-11-12T00:22:25+00:00

Get Chinese Astrology

Flat Earth Science MCP

2026-11-12T00:22:24+00:00

Flat Earth Science MCP

12 Chinese Zodiac Signs

2026-11-12T00:22:24+00:00

12 Chinese Zodiac Signs

How to Stop an AI Agent Stuck in an Infinite Loop

2026-11-12T00:00:00+00:00

An AI agent gets stuck in an infinite loop when a model keeps calling a tool that fails, usually with the same arguments, because nothing in the agent loop tells it to stop. Each retry resends the whole conversation, so every turn costs more than the last, and a loop left running overnight can cost thousands of dollars.

How to Benchmark LLMs

2026-11-12T00:00:00+00:00

Choose models using evidence from your application’s inputs and output formats. Run identical tasks through OpenRouter and compare graded accuracy, billed cost, and request time. On our 60 synthetic tasks, GPT-6.1 Sol passed 180/180 trials across three repeats under the tested contracts and settings.

Self-Hosted LLM vs API: When Does Self-Hosting Pay Off

2026-11-12T00:00:00+00:00

If you pay an API bill for an open-weight model, you may wonder whether your own GPU would cost less. GPU rent is charged every hour, busy or idle, while an API charges only for the tokens you use.

How do you monitor your MCP server?

2026-11-12T00:00:00+00:00

Use Manufact Analytics to track MCP tool usage, latency and errors. Inspect sessions, understand task outcomes and find what to improve next.

Why Are You Shipping So Much?

2026-11-12T00:00:00+00:00

Agents made shipping cheap. That's why asking what it's for matters more.

Arcaeon

2026-11-11T23:13:00+00:00

A record of what an agent did that the agent cannot quietly rewrite, and a way for anyone who doubts

Surface AI

2026-11-11T23:04:00+00:00

Surface AI is a remote MCP server available at https://mcp.surface-ai.io.

RoleWeek

2026-11-11T21:25:00+00:00

Search developer roles posted in the last 7 days in 35 US and Canadian metros plus remote, and ask what they post as salary, who is hiring and how last week went. Read-only, no sign-in.

graphify 0.9.107

2026-11-11T20:29:19+00:00

Fix: cpp a folded parameterless member declaration (bar()) records its implementation in definition_file as a path anchored to the same root as source_file, instead of a ../-walk relative to the cache/out root. The zero-argument reconciliation added in 0.9.106 stored the internal _cpp_source_key proof form, which _norm_source_file cannot re-anchor (it only relativizes absolute paths), so an out-of-tree corpus got a mangled definition_file. The sibling decl/def merge already stored the raw source_file for this reason; both paths now agree. The include-path proof still uses the key form.
Fix: llm a slice whose timeout or context overflow is recognised and given up on (it cannot be split further, or the retry depth is exhausted) now marks its file partial, so a document whose tail slice timed out is no longer cached and stamped complete from its other slices; it is re-queued on the next run while sibling files' results are kept (#5099, thanks @khrisshie).
Fix: php a route path scan stops at the closure's own argument, so $app->get('/users', function () { ... }) keeps GET /users instead of reading a trailing string argument after the closure as the route; a legitimate path before the closure is unchanged (#5135, thanks @mppyxx).
Fix: ids two module-level declarations of the same kind in one file whose names differ only by leading/trailing underscores (function _save(){} and function save(){ return _save() }) are two nodes again instead of one, and the call between them is a real save -> _save edge instead of a fabricated save -> save self-call; the declaration that appears first keeps the plain id and the other is salted. Affects JS, TS, C, C++, PHP, Ruby, Kotlin, Swift and Lua (#2810, thanks @khrisshie).
Fix: php methods of one class, trait, interface or enum whose names differ only by a leading/trailing _ (_x() / x()) are separate nodes instead of one swallowing the other: the one public name keeps its id and the others are salted, as in Python, C#, JS/TS and Solidity, so $this->x() and $b->x() reach x() instead of being lost and $this->_x() no longer looks like a self-call (#5137, thanks @khrisshie).
Fix: java members of one class whose names differ only by a leading/trailing _ (_name / name) or only by case (run / Run, which Java treats as distinct) get separate node ids instead of one swallowing the other; a public name keeps its id and the twins are salted, as in the other languages, so calls to each resolve correctly (#5151, thanks @rohit-jsfreaky).
Fix: test the copied-cache terraform id test accepts a worker pool that could not start under resource pressure (it falls back to sequential and warns), instead of asserting the pool always starts; a genuinely broken pool still fails the test (#5136, thanks @mppyxx).
Fix: php/js a call through an ESM namespace member (import * as ns; ns.handler()) or whole-module require resolves only when the target is actually callable, so a namespace member that names a non-callable export (an object literal, a constant) no longer gets a spurious call edge (#5144, thanks @ayushcodes10).
Perf: query the read path (query/explain/path/affected and the MCP graph load) no longer imports every language extractor, cutting import cost on read-only commands; graph output is unchanged (thanks @rohit-jsfreaky).
Perf: extract the AST worker pool is sized by the amount of uncached work (one worker per ~500 KB of uncached source, capped at CPU count and file count) when no worker count is set, so small incremental builds stop spinning up a full pool; output is invariant to worker count (thanks @rohit-jsfreaky).
Change: cloud the OSS surfaces (README, CLI help/install text, the translated docs in all 31 locales, and the agent skill) now point at the free plan with no card at app.graphify.com/signup, instead of a "14-day trial" at /login (the trial is paid-plans-only and needs a card, so the old copy conflated the two). Links carry utm_campaign=oss UTMs tagged per surface, so a signup can be attributed to where it came from.
Change: cloud the post-build CLI nudge is contextual. It is TTY-only (never in CI or the agent pipeline), respects GRAPHIFY_NO_TIPS/GRAPHIFY_NO_CTA, and shows on builds 1, 5 and 20 then stays quiet for good. When a local graph reaches 25k nodes, the free plan's per-repo cut-off, the nudge switches to a scale-specific line (tagged utm_content=scale) that sells Graphify Pro and links to pricing with the 14-day trial, instead of the free signup: a repo that size no longer fits the free plan, so pointing it at the free plan would land straight on the limit.
The AST cache schema is bumped to 44; warm caches re-run the affected C++, PHP, Java and module-level-id extractions.

Farside Research

2026-11-11T17:21:00+00:00

Cited deep research for agents. Estimate the cost, start research runs, search graded sources and ask grounded questions of a finished report. Every claim links to the passage it came from. Also lists and runs saved rese

Last EHR

2026-11-11T17:20:00+00:00

Open-source MCP server for FHIR. Agents read freely; every write waits for human approval. Medplum, HAPI, Epic sandbox.

Routing in Rust

2026-11-11T17:00:00+00:00

As of v1.110.0, all routing in LiteLLM flows through a Rust core. Here is why we rebuilt the router and how differential testing let us migrate it without regressions.

Anyreach

2026-11-11T16:06:00+00:00

Anyreach is a remote MCP server available at https://api.anyreach.ai.

Lampwick: Error Tracking & App Monitoring

2026-11-11T16:04:00+00:00

Lampwick: Error Tracking & App Monitoring is a remote MCP server available at https://app.lampwick.dev.

Gemini 4 Carbon vs Fable 5.2: który model wybrać do kodowania agentowego?

2026-11-11T15:58:33+00:00

Anthropic wydało Claude Fable 5.2 9 listopada 2026 r. Dzień później Google udostępniło Gemini 4 Carbon w wersji ogólnodostępnej. W tabeli benchmarków z…

Migracja API Gemini 4 Carbon: błędy 400, przekroczenia czasu i limity wyjścia

2026-11-11T15:58:02+00:00

Gemini 4 Carbon można zastąpić Gemini 4 Argon bez zmian w schemacie żądania. Działa ten sam obiekt treści żądania chat completions: zmienia się identyfikator…

Gemini 4 Carbon kontra Argon: ta sama cena — kiedy warto wybrać Carbon?

2026-11-11T15:57:41+00:00

Gemini 4 Carbon nie jest nową rodziną modeli. To Gemini 4 Argon z większą ilością rozumowania w każdej turze, oferowany w tej samej cenie za token: 2 USD za…

kwery

2026-11-11T15:08:00+00:00

Live prices, offers and trust scores from Amazon, Google Shopping, idealo, eBay, Kaufland and 6 more marketplaces, matched by GTIN. Look up, compare and batch-check products from your AI assistant. Connect with your kwer

A More Predictable Release Cycle for MCP

2026-11-11T09:00:00+00:00

We’ve spent a lot of time over the past two years talking about ways to improve MCP, add new features and capabilities, and get on a more predictable release cadence. Clearly, there has been no shortage of ideas both from the maintainer and broader developer community. Getting those ideas reviewed, implemented, tested, and into a release is a whole other part of the job, and we want to make that specific part easier to follow for our community.

For folks working on proposals, that means knowing when reviews will happen and how much time there is to make adjustments. SDK maintainers also need to know when the spec will stop changing - both client and server developers use our developer toolchain as a starting point to try new MCP capabilities out and share feedback. To make this process easier, we’re setting up a more consistent timeline for upcoming spec changes so that you don’t have to piece a schedule together from meeting notes and GitHub threads.

The Core Maintainers worked through these questions together at MCPCon Europe, and Kurtis Van Gent recently wrote up the process in SEP-3398: Release Cycle Updates - this is our de facto release guidance moving forward.

As a result of this work, we’re moving MCP specification releases to a regular six-month schedule.

Two releases a year#

We’ll generally publish specification releases on the last Tuesday of March and September. Alongside each release date, we’ll share the proposal deadlines and SDK beta dates, so you can work backward from them when planning your own implementation and testing.

For the next release, we’re targeting March 30, 2027 - expect MCP version 2027-03-30 to land then.

To keep that schedule predictable, a proposal that isn’t ready will move to a later release rather than delay the whole spec. Its authors can keep working on it without having to rush, and everyone else can continue preparing for the release they planned around.

Introducing the Release Manager role#

Preparing a specification release involves proposal authors, Working Groups, and the maintainers of several SDKs, each with their own work to finish. We need people who can keep track of that work across the project, spot when one team is waiting on another, and help resolve problems before a deadline. That’s the job we’re formalizing with the Release Manager role.

We’ll be working with proposal authors and SDK maintainers throughout the cycle, and Luca Chang will join David and myself as a Release Manager for the next specification cycle.

For those who haven’t met Luca yet, he co-authored the original Tasks proposal, which introduced experimental support for long-running work in the November 2025 specification. He then worked with Caitie McCaffrey on the next iteration of that capability - the Tasks extension.

Luca also co-leads the Tasks and Events Working Group and facilitates the Agents Interest Group. That work has earned him the community’s trust, and we’re looking forward to preparing the next release together.

Luca, welcome aboard!

The three of us will help get proposals into the review queue, help plan and manage the SDK implementation work, and follow up on anything that needs attention before the spec hits the shelves. Our group of Core Maintainers will continue to help decide which proposals are accepted into the next release.

What happens between releases#

Proposal authors need time to develop their ideas and respond to reviews, while SDK maintainers need time to implement and test the accepted changes. That’s why planning will start eight calendar months before a regular release. We’ll still publish every six months, so planning for the next release will overlap with the final two months of work on the current one.

The upcoming March release is just a little different, since some of the planning already concluded earlier in August of this year. For the regular cycles that follow, the schedule will generally be broken down like this:

Eight months before release: planning begins. We gather feedback from maintainers and Working Groups about what to work on and what depends on what.
Six months before release: Core Maintainer reviews begin and run until soft freeze.
Two and a half months before release: soft freeze. Every proposal going into the spec release must be marked Final and merged. Until hard freeze, we will accept existing proposal fixes, polish, and clarifications within their scope. New features or substantial redesigns are not accepted.
One and a half months before release: Tier 1 SDK betas arrive with support for the frozen specification.
One month before release: hard freeze. Only critical security fixes or fixes for release-blocking correctness or interoperability defects can go into the spec, with approval through a full Core Maintainer vote.
Release day: the specification and release notes go out, with implementation guidance. Tier 1 SDKs need stable support available beforehand. If an SDK is late, we’ll document that gap without moving the specification release date.

We believe that a soft freeze will give SDK maintainers a settled set of changes to implement, and the betas will give client and server developers a way to test those implementations together. Hard freeze then leaves the final month for release checks, with only the emergency changes described above allowed into the spec.

Lastly, when SDK betas are available, we’d love for the broader MCP community to try them with the clients and servers they already run. This is a great opportunity to catch any last-minute issues and gives the maintainers time to work through any reported bugs or inconsistencies while corrections can still go into the release.

Get involved#

If you have an idea for improving MCP, we encourage you to bring it to a Working Group and help shape it with other contributors. You don’t need a finished proposal to join the conversation, and you most certainly don’t need to wait for the next release cycle.

When a proposal is ready for formal review, a Working Group or the Core Maintainer group can sponsor it for the SEP review queue. You can also contribute by testing SDK prototypes or sharing feedback on proposals already in progress.

Our goal with these changes is to make contributing to MCP and building with it easier to plan around, whether it’s part of your job or something you do independently.

If you’ve been thinking about contributing to the protocol, come chat with us - we’d love your help!

v1.106.8

2026-11-11T05:48:11+00:00

Verify this release

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.106.8

Same command for litellm-database, litellm-non_root and the docker.io/litellm/* images

What's new in 1.106.8

2 PRs. 1 fixed, 1 changed. No action needed to upgrade

Changes

Cost tracking and pricing: Fixed incorrect transcription cost for some FLAC files, which could push a key over its budget (#50842) ^{Also in 1.108.5}
Cost tracking and pricing: Changed per-second transcriptions to bill the usage.seconds the provider reports instead of the measured length (#50727) ^{Also in 1.108.5}

Backport #50861 · Full changelog

v1.109.0

2026-11-11T00:56:51+00:00

Verify Docker Image Signature

All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.

Verify using the pinned commit hash (recommended):

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.109.0

Verify using the release tag (convenience):

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/v1.109.0/cosign.pub \
  ghcr.io/berriai/litellm:v1.109.0

Expected output:

The following checks were performed on each of these signatures:
  - The cosign claims were validated
  - The signatures were verified against the specified public key

What's Changed

fix(prompt-templates): treat initial_prompt_value and final_prompt_value as optional by @mdevhurt1 in #39762
fix(proxy)!: count prompt-only token_counter requests with the provider when call_endpoint is set by @mateo-berri in #46839
fix(interactions): keep background interaction bills through crashes, late restarts, and deployment keys by @devin-ai-integration[bot] in #45743
fix(proxy)!: honor caller litellm_session_id in spend log session_id by @deepanshupal in #44199
feat(router): mid-stream fallback continuation for chat completions by @ayushh0110 in #41127
feat(volcengine): add Doubao-Seedance video generation via the OpenAI /v1/videos surface by @Bytechoreographer in #46156
fix(realtime)!: keep turn_detection.type when a transcript guardrail disables auto-response by @devin-ai-integration[bot] in #44939
fix(budget): hold Redis-less budget spend per key instead of queueing it per request by @ryan-crabbe-berri in #46831
fix(proxy)!: delete callbacks configured under callbacks or failure_callback by @devin-ai-integration[bot] in #42740
feat(ui): turn the Chat UI on by default by @krrish-berri-2 in #46918
fix(bedrock)!: keep reasoning, thinking blocks and the json_mode finish_reason on fake streams by @mateo-berri in #46863
test(cost): add tier 1 cost tracking cases for gpt-5.6-terra, gemini-4-argon and gemini-3.1-pro-preview by @devin-ai-integration[bot] in #46931
fix(responses): log and bill every round of non-stream MCP auto-execute by @devin-ai-integration[bot] in #43667
fix(count_tokens): apply context_management edits in the Gemini /v1/messages/count_tokens count by @mateo-berri in #46843
test(cost): bill vertex gemini-3.8-flash reasoning tokens at the reasoning price by @devin-ai-integration[bot] in #46939
fix(proxy): streaming responses hand the event loop back every 50 ms instead of holding it for the whole buffered stream by @klu-berri in #46925
test(perf): cap every awaited Postgres statement at 0.5 calls per request by @klu-berri in #46944
feat(mcp): accept an ID-JAG as an RFC 7523 jwt-bearer grant on the gateway token endpoint by @yassin-berriai in #46748
feat(sso): capture the IdP id_token on Microsoft and Google logins for oauth2_id_jag by @yassin-berriai in #46747
feat(sso): retain the SAML assertion at login as the saml2 subject token for oauth2_id_jag by @yassin-berriai in #46745
test(rust-bridge): relocate dispatch tests to the Rust test suite by @yujonglee-berri in #46926
fix(tests): drain SDK logging before restoring callbacks by @klu-berri in #46924
test(integration): remove BUG skip for 1 test that passes on main by @devin-ai-integration[bot] in #46937
test(perf): cap every background Postgres statement at 0.9 calls per request by @klu-berri in #46957
fix(rust): forward Anthropic headers to DeepSeek by @mateo-berri in #46959
test(cost): rekey vertex gemini, add bedrock mantle gpt and claude-opus-5-5, drop suite-only price fields by @devin-ai-integration[bot] in #46962
fix(openai)!: route gpt-transcribe to GPT audio transcription config by @irvallensar in #44021
feat(ui): add the 1 hour cache write cost to the add model and model edit forms by @ryan-crabbe-berri in #46830
fix(proxy): report in-memory spend queue size by @mateo-berri in #46956
fix(bedrock): drop dangerous-tool-use beta without safeguards by @axelray-dev in #44471
feat(cost-map): add together_ai reflection/Beam-1.0-501B-A23B by @berriai-litellm-provider-info-sync[bot] in #46964
feat(ui): redesign the playground chat around a single composer and inspector by @yuneng-berri in #46934
build(deps): bump claude-agent-sdk to 0.2.165 to unblock osv-scan by @yuneng-berri in #46971
fix(proxy): keep a retried /v1/messages stream's spend log by logging the superseded attempt as a failure by @devin-ai-integration[bot] in #45136
fix(mcp): stop proxy-started legacy listeners from consuming server rpm by @yuneng-berri in #46968
fix(proxy): give /team/delete transactions a 30s timeout by @yuneng-berri in #46979
feat(ui): add Authorization URL, Token URL and Scopes fields for OAuth Delegate and True Passthrough MCP servers by @yassin-berriai in #46941
fix(guardrails): accept Gemini tools in generic_guardrail_api, add opt-in forward_messages_as_sent, name the error type on fail-open by @ryan-crabbe-berri in #46338
docs(agents): keep directory rules in nested AGENTS.md and pin integration tests to CircleCI by @mateo-berri in #46796
fix(proxy)!: answer 400 for malformed tools on count_tokens instead of crashing by @mateo-berri in #46583
fix(sso): concurrent first sign-ins land on the winner's account, worker-kill test stops waiting on the reap by @mateo-berri in #46855
fix(bedrock): backfill Priority and Batch prices from the AWS offer index by @berriai-litellm-provider-info-sync[bot] in #46987
feat(cost-map): add baseten and nebius rows for Reflection Beam 1.0 by @kerry-berri in #46991
fix(cost): bill bridged /v1/messages Responses replies like native /v1/responses for reasoning tokens and service tier by @yuneng-berri in #46975
fix(providers): route Azure Responses past dated api versions, count Grok reasoning tokens on azure_ai, refresh provider sets after register_model, fix gpt-6.1-sol Azure rows by @ryan-crabbe-berri in #46377
feat(openai)!: honor per-deployment workload identity on non-chat routes by @mateo-berri in #39966
refactor(ui): validate model edit prices in the zod schema by @ryan-crabbe-berri in #46978
fix(bedrock): set supports_none_reasoning_effort on gpt-6-sol and gpt-6-luna by @berriai-litellm-provider-info-sync[bot] in #47000
test(integration): fix seven CircleCI integration tests that race or assume a worker spread by @yuneng-berri in #47001
test(e2e): fix MCP tool search hit matching and the retired Bedrock image model by @yuneng-berri in #46998
test(ui): wait for the logout redirect before visiting a protected page by @yuneng-berri in #47008
feat(cost-map): support s3:// URIs for LITELLM_MODEL_COST_MAP_URL by @klu-berri in #46972
feat(ui): shareable URL state for usage views and users-to-usage deeplinks by @jesus-berri in #43425
fix(azure_ai): bill Grok reasoning tokens whatever the deployment is named by @ryan-crabbe-berri in #46969
fix(ui): preserve key settings drafts during background refreshes by @yuneng-berri in #41350
fix(openai_like)!: surface provider response headers on non-streaming chat completions by @jesus-berri in #46928
fix(bedrock)!: block realtime replay after an input-first start and reject malformed Nova Sonic session settings by @mateo-berri in #46936
fix(logging): stop the spend metadata merge from walking a dict the logging thread writes to by @devin-ai-integration[bot] in #46044
docs(cost): rewrite stale completion_cost docstring by @klu-berri in #47022
fix(openai): send chat completions through client.post so the SDK's per-message param transform stops holding the event loop by @klu-berri in #46967
feat: add Nace Drex decisions provider by @joshua-berri in #46523
test(integration): expect a pre-token invalid prompt on a bridged stream to fall back by @yuneng-berri in #47020
test(e2e): repair fallback management test for config-owned fallbacks by @yuneng-berri in #47023
fix(proxy): keep the OIDC login refresh token as the ID-JAG subject when the SSO client is the ID-JAG client by @yassin-berriai in #47004
test(e2e): assert the Vertex cache-creation billing contract in the prompt caching test by @yuneng-berri in #47027
build(docker): cache the Rust extension build in its own stage by @mateo-berri in #47033
fix(proxy): unset DB environment variables removed from the row on every worker by row revision by @mateo-berri in #46709
ci(cost map): classify token_count_model as an object key in the schema generator by @mateo-berri in #47057
fix(bedrock): use native structured output on Haiku 5.5 and Sonnet 5.5, and fix Fable 5 and Sonnet 4.6 schema 400s by @mateo-berri in #46870
fix(mcp): log upstream exchange when an MCP health check fails by @devin-ai-integration[bot] in #44125
perf(cli): start lite without initializing the litellm SDK by @mateo-berri in #46803
fix(prometheus): set the remaining rate-limit gauges from decimal headers again by @mateo-berri in #47039
ci(cost map): remove the generated model prices JSON schema by @mateo-berri in #47064
test(integration): write worker configs beside test-supplied configs by @yuneng-berri in #47066
test(e2e): keep the prompt-caching affinity test on the live provider by @yuneng-berri in #47072
test(integration): send post-kill requests on a fresh connection in two worker-kill tests by @yuneng-berri in #47080
fix(caching): keep prompt cache usage on streamed cache hits by @yuneng-berri in #47073
fix(logging): keep the async request's logging payload when the sync success thread runs late by @yuneng-berri in #47019
fix(cost): price exact cost-map names before price-free capability rules, and fix SCIM and MCP integration tests on CircleCI by @mateo-berri in #47037
test(e2e): event_loop_stalls perf case fails on a gateway event-loop hold over 100 ms that is not on its allowlist by @klu-berri in #46974
feat(pricing): backfill github_copilot prices from the Copilot pricing table by @kerry-berri in #47085
test(e2e): warm the memory test up with two batches before the measured one by @yuneng-berri in #47087
fix(bedrock)!: stream Converse passthrough frames through post_call guardrails instead of buffering by @yucheng-berri in #46307
fix(bedrock)!: scan Claude invoke-with-response-stream passthrough output with post_call guardrails by @yucheng-berri in #46524
feat(azure-ai): support Entra ID auth for Microsoft-Decision-1 on /v1/decisions and /v1/systemone by @jesus-berri in #46539
feat(bedrock): apply bedrock_request_metadata_fields on the /bedrock pass-through routes by @mateo-berri in #47040
fix(streaming): report the deployment's api_base when a stream fails mid-way by @mateo-berri in #47045
test(translation): recapture anthropic claude bases after usage speed field by @devin-ai-integration[bot] in #47116
test(translation): recapture azure ai claude bases after usage speed field by @devin-ai-integration[bot] in #47117
test(translation): add sonnet 5.5 and gemini vertex rows to cross-model mid tier by @devin-ai-integration[bot] in #46896
fix(anthropic): add the dangerous-tool-use beta when a /v1/messages request carries safeguards by @devin-ai-integration[bot] in #43955
fix(rust): add the dangerous-tool-use beta when an anthropic /v1/messages request carries safeguards by @devin-ai-integration[bot] in #46102
test(e2e): event_loop_stalls charges a hold to the longest gc pass that fits it, not to the sum of every pass since the last hold by @klu-berri in #47089
fix(anthropic): carry safeguard_results and stop_details in synthesized message_delta events by @devin-ai-integration[bot] in #46096
test(e2e): event_loop_stalls caps the request-handler bucket at 2, twice the gc-aware maximum by @klu-berri in #47128
fix(backend): retain Prometheus metrics during startup by @klu-berri in #47129
test(translation): allow anthropic usage speed in cross-model contract by @devin-ai-integration[bot] in #47130
feat(terraform)!: default the component images to the module release tag by @yassin-berriai in #46960
fix(mcp)!: cap the ID-JAG credential by its signed resource claim and refuse a resource the gateway does not serve by @yassin-berriai in #47094
fix(cost-map): fall back to the bundled map when urlsplit rejects LITELLM_MODEL_COST_MAP_URL by @klu-berri in #47133
test(translation): add claude-haiku-5-5 vertex ai cases by @devin-ai-integration[bot] in #47136
feat(deepinfra): add DeepSeek-V4.1-Flash and DeepSeek-V4.1-Flash-1017 to the cost map by @kerry-berri in #47132
fix(responses): run pre_call guardrails on every Responses WebSocket turn and bill each turn its own guardrail cost by @klu-berri in #47013
perf(proxy): gc.freeze() the startup heap once the proxy is up so gen-2 passes stop re-walking it by @klu-berri in #47137
test(integration): remove BUG skips for 2 tests that pass on main by @devin-ai-integration[bot] in #47142
test(cost): add tier 1 cost tracking cells and sync usage shapes for 2026-10-17 by @devin-ai-integration[bot] in #47141
fix(presidio): stop overlapping entity spans from corrupting numbered PII tokens by @yucheng-berri in #47024
fix(otel): report cache hits at zero cost with a cache-hit flag in V1 and V2 spans by @devin-ai-integration[bot] in #43604
fix(proxy): let proxy_admin_viewer store its own MCP credentials by @yassin-berriai in #46994
test(perf): qualify database upgrades with shared 5 RPS traffic by @klu-berri in #46802
fix(register_model)!: keep bedrock geo deployments priced across price data reloads by @devin-ai-integration[bot] in #41523
fix(mcp)!: register the gateway's own DCR client for bridge servers instead of relaying the caller's loopback redirect by @yassin-berriai in #46940
fix(model_prices): registry audit 2026-10-17, Perplexity decider v1.2 rows, Azure Mistral and Vertex Gemini deprecation dates by @devin-ai-integration[bot] in #47114
fix(mcp): bound paginated tools/list by LITELLM_MCP_TOOL_LISTING_TIMEOUT by @yassin-berriai in #46513
fix(vertex): match standalone 403 status tokens instead of any substring by @devin-ai-integration[bot] in #47134
feat(guardrails): add num_retries to generic_guardrail_api for checks lost after connect by @ryan-crabbe-berri in #46965
fix(proxy): route team key vector store calls to a team model of the store's provider by @mrinal-berri in #46894
fix(batches): handle empty models when logging batch cost by @joshua-berri in #47155
fix(vector-stores): keep proxy context out of provider vector store metadata by @yuneng-berri in #47021
fix(proxy): report requested model in streamed Responses response.model by @albertoperdomo2 in #46662
fix(logging): run legacy s3 callback for streamed /v1/messages calls by @klu-berri in #47156
test(e2e): get_text_external accepts JSON next to text so a pod-log error reads as the Kubernetes Status, not a 406 by @klu-berri in #47150
fix(spend_logs): persist user email in spend log metadata so it shows in UI request logs by @devin-ai-integration[bot] in #41150
test(integration): run the bedrock mantle prompt cache and output limit wire tests on their own proxy by @yuneng-berri in #47152
test(e2e): enable prometheus_system on the redis chaos proxy config by @yuneng-berri in #47160
fix(tests): wait for every thread of a killed worker before treating it as gone by @yuneng-berri in #47148
fix(proxy): warn at boot when a config environment_variables key replaces a value already set in the process environment by @klu-berri in #47165
test(caching): assert sync cache-hit logging without depending on thread order by @yuneng-berri in #47093
test(integration): report uncollected shard files as collection errors instead of crashing xdist by @yuneng-berri in #47075
test(e2e): hold failing-request RSS to a loose ceiling over two measured batches by @yuneng-berri in #47170
fix(count_tokens)!: answer 400 for Vertex partner models without a count API and OpenAI-shaped provider errors on /v1/responses/input_tokens by @mateo-berri in #47163
fix(router): read an application profile's base model from metadata on a cache hit by @yuneng-berri in #47144
fix: handle missing choice indices in streamed output recording by @yucheng-berri in #46947
test(mcp): replace the live Linear MCP OAuth e2e tests with integration coverage by @yuneng-berri in #47154
fix(logging): keep root logger INFO lines when json_logs is on and LITELLM_LOG is set by @klu-berri in #47175
fix(router)!: stop a chat-mode Mantle deployment from rewriting the shared Responses routing row by @mateo-berri in #46263
fix(ui): take the dashboard's user role and identity from the server by @yuneng-berri in #47025
fix(router)!: bound fake-streamed calls by the request timeout, not stream_timeout by @mateo-berri in #47162
fix(proxy): count the whole minor line in the upgrade notice for patch releases by @misbahsy in #46942
feat(openinfer): add OpenInfer as an OpenAI-compatible provider by @AJ-ing in #41034
fix(rag)!: use team provider credentials and allow viewer and JWT team access by @shivamrawat1 in #45497
fix(bedrock): send native structured output first on Opus 4.8 and Sonnet 5 and fall back to the json tool when Bedrock rejects it by @mateo-berri in #47100
fix(bedrock): add the July 7, 2026 EOL date to the Llama 3.1 405B and Llama 3.2 rows by @berriai-litellm-provider-info-sync[bot] in #47196
fix(batches): preserve spend across malformed output and metadata failures by @joshua-berri in #47184
feat(count_tokens)!: configurable token_count_framing and token_count_model, count source in the response by @mateo-berri in #46777
fix(baseten): forward reasoning_effort to Baseten Model API by @ucfjoe in #40152
fix(bedrock): take the 200K context window from the Claude Sonnet 4 model card by @berriai-litellm-provider-info-sync[bot] in #47202
feat(ui): show gateway URL and agent setup instructions after creating a key in #47199
test(baseten): annotate test locals with Final by @ryan-crabbe-berri in #47205
fix(proxy)!: end a chat completions stream with an error event when a guardrail blocks mid-stream by @mateo-berri in #46953
fix(bedrock): take the 1,050,000 context window from the Daybreak Mantle model cards by @berriai-litellm-provider-info-sync[bot] in #47211
test(perf): measure SpendLog-backed UI query latency by @klu-berri in #46955
fix(databricks): use the native Anthropic route for Claude /v1/messages by @devin-ai-integration[bot] in #46091
fix(bedrock): take Grok 4.7 and Pegasus v1.5 context windows from the Bedrock model cards by @berriai-litellm-provider-info-sync[bot] in #47226
chore: bump litellm-enterprise 0.1.77 -> 0.1.78, litellm-proxy-extras 0.4.109 -> 0.4.110 by @yuneng-berri in #47228
test(integration): give each owned proxy a private database and Redis by default by @yuneng-berri in #47213
fix(guardrails)!: preserve Bedrock Converse redactions and honor tool-only scans by @moyai-berriai[bot] in #45874
test(perf): align read queries with current dashboard calls by @klu-berri in #47243
fix(proxy): authorize the public name of a team-scoped model on batch retrieve/cancel by @Mrun25 in #47067
feat(pricing): add baseten DeepSeek-V4.1-Flash-1017 models by @berriai-litellm-provider-info-sync[bot] in #47252
chore(cost-map): add fireworks deepseek v4.1 flash 1017 by @berriai-litellm-provider-info-sync[bot] in #47253
feat(pricing): add missing github_copilot rows from models-and-pricing by @berriai-litellm-provider-info-sync[bot] in #47254
test(integration): stop GET /__observations from draining the upstream log by @yuneng-berri in #47256
test(integration): look up owned cache keys and clean up sweep models by @yuneng-berri in #47255
feat(vertex_ai): let operators drop chosen anthropic-beta values on the Vertex AI pass-through routes by @mateo-berri in #47185
fix(streaming)!: map stream read timeouts to litellm.Timeout on chat, messages, and responses by @mateo-berri in #47198
fix(logging): deliver completed batch success callbacks consistently by @joshua-berri in #47192
test(integration): wait for every worker in Scenario.model by @yuneng-berri in #47277
fix(bedrock): take the 256K context window from the Qwen3 Coder 480B model card by @berriai-litellm-provider-info-sync[bot] in #47283
fix(proxy): sanitize unexpected database errors from /key/info by @vanrajsinh650 in #46190
test(rust): native vs Python parity harness for responses by @narmstrong-berri in #46567
feat(rust): encode LiteLLM response ids on native responses by @narmstrong-berri in #46572
feat(cost-map): add together_ai DeepSeek-V4.1-Flash-1017 by @berriai-litellm-provider-info-sync[bot] in #47291
fix(azure): set azure/container code interpreter session price from the azure openai pricing page by @berriai-litellm-provider-info-sync[bot] in #47292
fix(vertex-ai): drop floor deprecation_date from gemini-3.8-flash by @berriai-litellm-provider-info-sync[bot] in #47294
fix(vertex_ai): price streamed pass-through calls with the regional uplift by @pcbeingused333 in #44795
chore(pricing): add meta muse-image-1.0 per-image price by @berriai-litellm-provider-info-sync[bot] in #47318
fix(azure): route WebRTC client_secrets and calls to the GA v1 realtime API by @habonlaci in #47316
fix(azure): reuse the cached Entra token provider on the Responses path by @jorik in #46421
fix(router)!: count native Responses WebSocket turns toward deployment limits and budgets by @devin-ai-integration[bot] in #45727
refactor(rust): fold inference-transcription prepare into its handler by @yujonglee-berri in #47214
refactor(rust): fold inference-responses prepare into its handler by @yujonglee-berri in #47215
refactor(rust): fold inference-ocr prepare into its handler by @yujonglee-berri in #47216
refactor(rust): fold inference-messages prepare into its handler by @yujonglee-berri in #47217
refactor(rust): fold inference-chat prepare into its handler by @yujonglee-berri in #47218
chore(cost-map): add fireworks minimax m3.1 flash priority prices from the prices api by @berriai-litellm-provider-info-sync[bot] in #47346
fix(baseten): list current Model APIs instead of retired 2023 deployment IDs by @AlexKer in #46945
feat(router): discover kubernetes pods behind a headless service api_base by @devin-ai-integration[bot] in #44342
feat(router): keep a session on one kubernetes pod by @devin-ai-integration[bot] in #44393
feat(ui): show kubernetes pods on the model page by @devin-ai-integration[bot] in #44666
fix(count_tokens)!: count tools the way DeepSeek V4 and Qwen3.5 chat templates render them by @mateo-berri in #47173
fix(cost-map): enable function calling + tool choice for fireworks deepseek-v4p1-flash-1017 by @jonvarley-fw in #47326
test(cost): add vertex claude-haiku-5-5 long context cases and wire gpt and gemini-3.8-flash bridge legs by @devin-ai-integration[bot] in #47342
chore(pricing): add github_copilot/claude-opus-4.8 deprecation date by @berriai-litellm-provider-info-sync[bot] in #47370
build(deps): bump aiohttp to 3.14.5 to clear osv-scan by @mateo-berri in #47369
fix(cost-map): add fireworks minimax m3.1 flash standard prices by @berriai-litellm-provider-info-sync[bot] in #47371
feat(proxy): opt-in self-serve budget edits for key owners on /key/update by @justinfx in #44763
fix(openai): drop unsupported reasoning effort for gpt-6.1-sol by @pOpO1-9 in #43935
fix(helm): stop litellm-helm from injecting placeholder models into every install by @yassin-berriai in #47056
fix(deps): require anyio>=4.14.0 to rule out the lock-waiter deadlock by @yucheng-berri in #47378
fix(router): record cooled deployments on the all-cooled failure row by @yucheng-berri in #47120
test(integration): hold the upstream until the disconnect is settled before reading its spend row by @yucheng-berri in #47303
chore(pricing): add meta sam-3.1 image segmentation price by @berriai-litellm-provider-info-sync[bot] in #47395
feat(search): add DeepSeek as a /v1/search provider by @kerry-berri in #47401
fix(azure): cache the remaining per-request Entra credentials and keep token refresh off the event loop by @mateo-berri in #47348
fix(logging): report the cost-map entry billing used in model_map_information by @klu-berri in #47182
fix(prometheus): label deployment failures with the deployment's provider instead of None by @yucheng-berri in #47384
fix(vertex-ai): drop floor deprecation_date from gemini-4-argon by @berriai-litellm-provider-info-sync[bot] in #47404
fix(anthropic)!: answer 400 for a malformed tool_result text on /v1/messages and count_tokens by @mateo-berri in #47347
fix(decisions): log every request field the caller sent on /v1/systemone and /v1/decisions by @klu-berri in #47183
feat(router): let a deployment set a timeout per service_tier by @klu-berri in #47406
fix(managed_files): send provider file ids when creating a fine-tuning job from managed files by @yucheng-berri in #46552
fix(proxy): stop billing every OpenAI pass-through /responses sub-route as a Responses call by @mateo-berri in #47171
test(integration): check SSO sign-in removal without reaching the public internet by @yuneng-berri in #47409
test(integration): expect the requested model in streamed Responses events on the arize v2 rig by @yuneng-berri in #47407
test(ui-e2e): read the virtual key from its own pre in the Save your Key dialog by @yuneng-berri in #47408
fix(cost): bill Gemini audio/image/video prompt tokens at the resolved input rate when no modality rate exists by @klu-berri in #47373
test(integration): read stream stall spend rows from the proxy that served them by @yuneng-berri in #47410
feat(cli): add lite claude --subscription and pkce config default by @yassin-berriai in #47385
test(integration): wait for every worker to report the websocket session cap instead of sleeping by @yuneng-berri in #47412
fix(exa): return highlights, highlight scores, score and the response envelope from Exa search by @yucheng-berri in #47423
fix(proxy): refresh cached team on /team/block and /team/unblock by @hamzashah-dev in #35643
fix(scim)!: store PATCH enterprise extension attributes in metadata.scim_enterprise by @yassin-berriai in #46738
test(perf): stop flagging load-balanced worker reads as restarts in database upgrade test by @devin-ai-integration[bot] in #47425
fix(mcp)!: retry upstream 429 on tools/list and error when every server fails by @yassin-berriai in #46649
feat(proxy)!: personal keys inherit MCP server, MCP tool and agent access from the owner's teams by @yassin-berriai in #46293
fix(speech)!: correctly propagate spend logging and usage for Gemini TTS bridge by @vanshsingh3711-wq in #44551
fix(anthropic): apply tool_config cache-control points on chat and /v1/messages by @klu-berri in #47394
fix(proxy): price the estimated input tokens on a read-timed-out request's spend row by @klu-berri in #47435
perf(savings): read deployment pricing from the router's cached model info by @narmstrong-berri in #47400
chore(hooks): drop Conventional Branches pre-push hook in favor of litellm_ prefix by @devin-ai-integration[bot] in #44951
chore: add yassin-berriai as code owner for helm/ and terraform/ by @yuneng-berri in #47452
fix(vertex-ai): drop floor deprecation_date from gemini-3.8-live by @berriai-litellm-provider-info-sync[bot] in #47447
perf(router): resolve deployment usage limits once per deployment instead of every request by @narmstrong-berri in #47393
test(e2e/perf): read the memory baseline after a one minute warm-up by @klu-berri in #47455
fix(logging): do not stamp a provider handoff when pre_call runs for a proxy-only failure by @klu-berri in #47456
test: catch regressions in the validation lines #44798 added by @mateo-berri in #47420
fix(vertex_ai): bill per-character text embeddings on Google's billableCharacterCount by @klu-berri in #47460
fix(friendliai): drop litellm-internal message fields before sending by @joshdabosh in #45092
test(images): stop writing image edit responses to disk by @yujonglee-berri in #47463
chore(bedrock): add Model EOL deprecation dates for Palmyra X4, X5 and Pegasus v1.2 by @berriai-litellm-provider-info-sync[bot] in #47462
test(integration): read router cooldown failure rows from the proxy that served them by @yuneng-berri in #47457
fix(secret_managers)!: resolve oidc/file and oidc/env refs in admin-configured api_key per request by @mateo-berri in #47341
fix(sdk): decide MCP gateway routing without importing the proxy handler by @MFJG-dev in #45961
fix(batches): apply the Vertex regional endpoint uplift to batch cost by @klu-berri in #47466
feat(key_management)!: apply default_key_generate_params.allowed_routes on key generation by @devin-ai-integration[bot] in #41236
fix(proxy): price the estimated input tokens on a client-cancelled (499) request's spend row by @klu-berri in #47436
test(perf): gate the gateway's own overhead header, redis latency and postgres load in the perf cases by @klu-berri in #47472
fix(databricks): stream Claude json_schema answers as text on every streaming path by @mateo-berri in #47424
perf(streaming): cut per-chunk pydantic round trips on openai-compatible streams by @yassin-berriai in #47030
fix(proxy): match partial wildcards against prefix-stripped model ids by @Mrun25 in #46390
fix(proxy): price a billed failure row at the deployment's own rate like the success path by @klu-berri in #47437
fix(router): key prompt caching affinity on tools and the /v1/messages system prompt by @yassin-berriai in #46411
fix(token_counter): count DeepSeek V4 text with DeepSeek's own tokenizer by @mateo-berri in #47353
fix(proxy): resync configuration after Redis reconnect by @emerzon in #43313
chore(pricing): add github_copilot deprecation dates for models retired on 2026-10-19 by @berriai-litellm-provider-info-sync[bot] in #47487
feat(proxy): spend log storage probe and the reclaimable-bytes count by @klu-berri in #47449
fix(proxy): build the OpenAPI schema off the event loop so a cold /openapi.json no longer stalls the worker by @mateo-berri in #47352
feat(router)!: apply each deployment's reasoning_effort_levels on every provider and route by @mateo-berri in #46601
feat(bedrock): route token counts from model metadata by @mateo-berri in #47441
fix(mcp)!: bill both model calls of a chat completions MCP auto-execute on one spend row by @mateo-berri in #47422
test(integration): remove BUG skips for 3 tests that pass on main by @devin-ai-integration[bot] in #47364
test(translation): add sonnet 5.5 converse and gemini direct cross-model mid rows by @devin-ai-integration[bot] in #47328
feat(dashscope): add separate explicit and implicit cache pricing by @limdeng in #45567
chore(pricing): add github_copilot gpt-4.1 deprecation dates by @berriai-litellm-provider-info-sync[bot] in #47492
fix(proxy): decode zstd and gzip request bodies before auth reads them by @mateo-berri in #47418
fix(streaming): keep finish_reason and provider usage when rebuilding streamed text completions by @roywei in #44349
fix(videos)!: keep model_id in proxy video ids so per-model keys resolve on status/content by @ekpog200 in #44614
fix(search): price Tavily advanced and Linkup deep searches at their own cost map entries by @yucheng-berri in #47411
fix(otel): put the error message on every failed span status by @HardMax71 in #44185
fix(prometheus): preserve requested model after fallback by @AaronHowell in #45555
feat(router): route batch creates to mode batch deployments and expose supported_endpoints on /model_group/info by @mateo-berri in #47376
fix(bedrock): resolve profile pricing before applying custom rates by @joshua-berri in #47419
feat(context_management)!: honor clear_at_least, exclude_tools and clear_tool_inputs in the clear_tool_uses polyfill by @AlonSoko in #47123
fix(router)!: only prefixed caller tags route once tag_routing_prefix is set by @yucheng-berri in #47428
fix(router)!: schedule default_priority requests once and reject a malformed priority by @mateo-berri in #47498
chore(cost-map): add fireworks 2026-11-07 deprecation dates from the prices api by @berriai-litellm-provider-info-sync[bot] in #47518
fix(proxy)!: drop router_settings_override before routing and apply it on every branch by @Avani-prajapati in #47257
fix(anthropic): json_schema envelope edge cases on the Anthropic json tool path by @mateo-berri in #47478
fix(streaming): restore the token estimate for streams that send no usage by @zuraiz-anjum in #44230
feat(harness): add pi harness for litellm.agent() by @fabioAntGit in #45467
feat(proxy): map baggage values to budget tags by @mateo-berri in #47358
fix(responses): keep strong refs to stream and websocket logging tasks (GC race) by @anneheartrecord in #31074
fix(sagemaker)!: parse every streamed record however SageMaker cuts the payload parts by @mateo-berri in #47495
fix(otel)!: leave out span cost when the price lookup failed by @HardMax71 in #44187
test(integration): follow #47441 metadata-routed counting in count_tokens wire tests by @mateo-berri in #47525
fix(proxy)!: let proxy admin keys read raw cloud storage file ids on GET /files/{id}/content by @mateo-berri in #40621
feat(embeddings): enforce per-request input batch limit by @VincentG1234 in #44574
fix(vertex_ai): reuse known context caches instead of listing on every request by @ondraz in #46857
fix(responses): count billed background responses toward the creator's project and per-model budgets by @mateo-berri in #47504
fix(bedrock): apply the 16 maxTokens floor to Kimi K3 on Converse by @devin-ai-integration[bot] in #43875
fix(cost-map): registry audit 2026-10-20, add volcengine doubao-seed-2-1 260915 rows and undated azure data zone rows by @devin-ai-integration[bot] in #47561
test(perf): set the seven new perf ceilings from 124 lane runs across v1.103.0 to v1.106.0 by @klu-berri in #47527
fix(langfuse_otel)!: send the completion start time by @HardMax71 in #45597
refactor(types): replace Any with proven types in 6 files by @mateo-berri in #47268
test(translation): add sonnet 5.5 anthropic direct cross-model mid row by @devin-ai-integration[bot] in #47566
fix(anthropic, bedrock)!: keep a redacted_thinking content item on every Claude route by @mateo-berri in #47489
fix(logging): report the pay-as-you-go entry billing used on Azure PTU spillover by @klu-berri in #47567
fix(router): skip non-dict tool entries in the prompt caching affinity key instead of raising by @klu-berri in #47565
fix(litellm_proxy): forward tool_config cache_control_injection_points to the downstream proxy by @klu-berri in #47570
fix(responses)!: count input tokens the way inference sends the request by @mateo-berri in #47363
fix(proxy): share the billing logger across callback registries by @joshua-berri in #47510
fix(decisions): log non-UTF-8 bytes as base64 so the request reaches validation and gets a 400 by @klu-berri in #47564
fix(batch): preserve recoverable costs and prevent poll starvation by @joshua-berri in #47190
test(integration): expect the Exa search envelope's provider_metadata after #47423 by @mateo-berri in #47572
feat(proxy): lift every earlier billed timed-out retry attempt onto the failed request by @klu-berri in #47490
fix(proxy): requeue spend logs when a Postgres restart tears down the prisma engine mid-write by @klu-berri in #47571
fix(proxy): fall back to the deployment model when a billed failure row's base_model is not in the cost map by @klu-berri in #47568
test: isolate model cost maps in runtime registration tests by @CNSeniorious000 in #47559
test(perf): measure Admin UI journey latency in a browser on the 1TB SpendLogs restore by @klu-berri in #47486
fix(cost): price base_model deployments past rule-only cost map entries by @klu-berri in #47497
fix(completion): send prompt_cache_retention to OpenAI instead of silently dropping it by @ryan-crabbe-berri in #47480
test(integration): call _fallback_deployment in the count_tokens wire tests #47363 added by @klu-berri in #47593
chore: lead the PR template with a short plain description and collapse the rest by @ryan-crabbe-berri in #47594
fix(bedrock): resolve AWS region and secret lookups off the event loop by @mateo-berri in #47331
fix(tests): read owned proxy rows from the owned proxy's database by @mateo-berri in #47333
fix(proxy)!: stop update_cache from resurrecting evicted auth cache entries by @devin-ai-integration[bot] in #46997
test(cost): run azure_ai claude bridge pairs, add anthropic usage speed and vertex gemini audio priority price by @devin-ai-integration[bot] in #47582
test(perf): set every perf ceiling to 1.2x the lane median by @klu-berri in #47590
test(integration): build the Mantle count_tokens wire test deployments with _metadata_deployment by @mateo-berri in #47592
feat(proxy): write one spend row per billed timed-out retry attempt by @klu-berri in #47416
fix(vertex_ai)!: send integer enums to Gemini instead of dropping them by @HardMax71 in #44908
test(bedrock): remove the Vault probe timing dependency by @mateo-berri in #47597
feat(proxy): size-triggered spend log cleanup with maximum_spend_logs_size, oldest rows first by @klu-berri in #47450
feat(proxy): opt-in free-tier allowance for Google Search grounding spend by @klu-berri in #47494
test: fix flaky team delete, refusal sink, loop-blocking, Moyai tamper and model-cost leak tests by @devin-ai-integration[bot] in #47135
test(integration): remove BUG skip for 1 test that passes on main by @devin-ai-integration[bot] in #47603
feat(proxy): size pass drops the oldest spend log partitions first by @klu-berri in #47451
test(integration): repair seven stale integration tests on main by @yuneng-berri in #47600
docs: clarify unit test clock rule by @mateo-berri in #47604
fix(proxy): one spend log cleanup pass at a time per pod, with or without Redis by @klu-berri in #47500
fix(decisions): bill unpriced returned TypeSafe versions at the requested model and name a missing answers field by @mateo-berri in #47508
fix(bedrock): return a 400 for malformed batch JSONL lines by @abhishekacharya200 in #45992
refactor(types): replace Any with proven types in 15 files by @mateo-berri in #47509
test(perf): set the small case's postgres_max_background_statement_per_request ceiling to 0.9 by @klu-berri in #47616
fix(proxy): flush the spend-log queue only for Responses turns the bridge reads back by @klu-berri in #47614
fix(proxy): size pass rebuilds reset row statistics with ANALYZE and keeps its reclaimable count by @klu-berri in #47501
fix(proxy): bill partial Responses API streams when the client disconnects by @klu-berri in #47611
fix(bedrock): send batch jobs to a configured control-plane endpoint by @msdianprince-7 in #47260
test(perf): record the data request status per UI journey sample and bound each journey with a wall clock by @klu-berri in #47622
test(perf): let the UI journey fixture run on a restore upgraded by a released migrations image by @klu-berri in #47628
fix(ui): rebuild the dashboard bundle and cover it in the example master key check by @mateo-berri in #47612
fix(harness): set PWD to the working directory of local sandbox commands by @mateo-berri in #47530
chore(cost-map): add fireworks glm 5.3 flash us deprecation date from the changelog by @berriai-litellm-provider-info-sync[bot] in #47591
chore: remove description heading from PR template by @mateo-berri in #47631
fix(logging)!: redact Responses instructions and callback output by @emerzon in #43315
feat(cost): add model_info.pricing_provider for cost-map lookups by @klu-berri in #47620
test(perf): emit a top-contributors breakdown property per perf metric by @klu-berri in #47632
fix(batches): retry confirmed pre-accounting failures by @joshua-berri in #47625
fix(anthropic): keep tool names over 64 chars when a guardrail rewrites a /v1/messages request by @yucheng-berri in #47434
fix(proxy): strip caller base_model from spend pricing and reject lone surrogates in multipart metadata by @klu-berri in #47639
feat(github_copilot): support a GHE.com host on per-user Copilot credentials by @yassin-berriai in #47610
fix(ui_sso): seal the litellm_cp_return_to cookie so edge firewalls stop blocking the MCP OAuth sign-in by @yassin-berriai in #47595
fix(mcp)!: never forward the admission LiteLLM key upstream for any auth type by @yassin-berriai in #46932
fix(proxy): restore daily tag spend to the in-memory queue when the Redis RPUSH fails by @yassin-berriai in #47307
fix(mcp): surface upstream HTTP status when a failed POST closes the session by @yassin-berriai in #47569
fix(bedrock)!: apply the deployment guardrailConfig on /bedrock pass-through routes by @yucheng-berri in #47429
fix(proxy): price a billed failure row at the deployment's long-context tier like the success path by @klu-berri in #47647
test(perf): tag a UI journey sample with its API response, not a Next.js prefetch under /ui/ by @klu-berri in #47650
fix(proxy): scope the search free-tier counter to the request's provider by @klu-berri in #47646
fix(ui): offer access-group models in team per-model rate limit and budget pickers by @mrinal-berri in #47630
refactor(rust-bridge): decide native routing in the catalog, drop the bypass predicates by @yujonglee-berri in #46750
refactor(rust-bridge): detect Python-only features in the catalog instead of Rust declines by @yujonglee-berri in #46714
chore(cost-map): add search_free_requests to the model prices schema by @klu-berri in #47635
fix(proxy): price a 499 spend row only for the proxy's own disconnect, and never after the call already completed by @klu-berri in #47652
fix(logging)!: fire one CustomLogger failure hook per async call outside the async flag list by @yucheng-berri in #47124
fix(router): price cost-based routing through model_info.base_model by @klu-berri in #47617
fix(proxy): price earlier timed-out retry attempts at their own deployment's model rate by @klu-berri in #47648
fix(logging): spilled over PTU requests without base_model log the served model's rates by @klu-berri in #47655
fix(vertex-ai): drop flex prices from gemini-nano-banana-2.1 by @berriai-litellm-provider-info-sync[bot] in #47666
feat(cost-map): add fireworks deepseek v4.1 flash 1017 us at the us-only premium by @berriai-litellm-provider-info-sync[bot] in #47667
chore: bump litellm-enterprise 0.1.78 -> 0.1.79, litellm-proxy-extras 0.4.110 -> 0.4.111, litellm 1.107.0 -> 1.108.0 by @yuneng-berri in #47668
feat(extraction): add /v1/extraction with Cloudflare Clairic support by @mateo-berri in #47473
fix(mcp): upgrade to MCP SDK 2.3.0 and raise the per-event SSE size cap by @joshua-berri in #47034
feat(nvidia_riva)!: add riva_offline mode for offline-only ASR models by @devin-ai-integration[bot] in #41021
fix(proxy): flush tool discovery queue when use_redis_transaction_buffer is enabled by @mrinal-berri in #47653
build(deps): bump pillow and dashboard vite, defer certifi to clear osv-scan by @yuneng-berri in #47678
fix(responses): send the billed background response cost to the creating key's and team's loggers by @mateo-berri in #47506
fix(vertex_ai): log Gemini 3 sampling deprecation once per model per worker by @devin-ai-integration[bot] in #40868
fix(proxy): attribute retry attempt spend rows to their deployment and keep them out of daily request counts by @klu-berri in #47658
fix(proxy): show the newest health check per model group on the model hub by @michelligabriele in #47552
fix(gemini)!: honor supports_sampling_params on the Gemini and Vertex path and flag Gemini 3.6 Flash, 3.5 Flash-Lite and later by @ryan-crabbe-berri in #47645
fix(mcp): list per-user auth MCP tools without a shared token by @yassin-berriai in #47677
chore(deps-dev): bump vite from 8.3.0 to 8.3.4 in /vscode-extension by @dependabot[bot] in #47627
test(perf): record the row count and size of each table the Admin UI journeys query by @klu-berri in #47683
fix(ui): show a failed state with Retry on the Logs table, Usage totals and Caching chart when their fetch errors by @klu-berri in #47671
feat(wandb): add Beam 1.0 and GLM-5.3 to the cost map by @berriai-litellm-provider-info-sync[bot] in #47690
feat(model hub): show admin-defined custom tags and metadata from model_info by @jesus-berri in #46064
test(perf): emit the database upgrade run's perf rows and upgrade timings as JUnit properties by @klu-berri in #47696
fix(ui): show the budget spend limit and rate limits without a collapsed section by @yuneng-berri in #47662
fix(scim)!: delete a group through the team delete's bulk path instead of one member at a time by @mateo-berri in #47544
test(perf): emit event_loop_stalls hold counts as perf_ JUnit properties by @klu-berri in #47706
fix(token_counter): stop re-downloading a HuggingFace tokenizer after a failed download by @devin-ai-integration[bot] in #47586
ci(cost map): delete the model prices JSON schema file by @mateo-berri in #47079
test(fine_tuning): pin that api_version never reaches the OpenAI client by @devin-ai-integration[bot] in #44995
test(perf): allowlist anyio TaskHandle._run_coro in event_loop_stalls by @klu-berri in #47746
fix(streaming)!: send the include_usage chunk with an empty choices array by @albertoperdomo2 in #47109
test(perf): add chat_completions_2rps, responses_2rps and messages_2rps direct vs bridge cases by @klu-berri in #47752
fix(router): unregister strategy selectors from global callbacks on discard by @setouchi-h in #45795
feat(gemini): set the long decode budget for requests with no output limit from LITELLM_GEMINI_DEFAULT_MAX_OUTPUT_TOKENS by @mateo-berri in #47542
feat(ui): add Delete to the guardrail detail page and share one delete flow with the table row by @jesus-berri in #45893
test(perf): emit per-location seconds, offending locations and the ten longest holds as event_loop_stalls breakdowns by @klu-berri in #47757
fix(responses): bill every round of a streamed MCP auto-execute on one spend row by @mateo-berri in #47445
ci(integration): keep the MCP SDK1 peer on mcp 1.x and autovacuum off the size cleanup test by @mateo-berri in #47651
fix(cli): send pi's session id from lite pi so sessions group and stick by @mateo-berri in #47700
feat(azure): add azure_ai/MAI-Code-1.1-Flash by @berriai-litellm-provider-info-sync[bot] in #47779
feat(vertex-ai)!: support gemini 3.1 flash tts by @emerzon in #31915
fix(pricing): update baseten reflection/Beam-1.0 prices from the models API by @berriai-litellm-provider-info-sync[bot] in #47784
fix(router): log each dropped /v1/messages attempt through its own logging copy and open the retry without its usage by @mateo-berri in #45194
chore(cost-map): add fireworks muse spark open priority prices from the pricing page by @berriai-litellm-provider-info-sync[bot] in #47785
test(perf): ramp the load cases until the pod stops keeping up and floor sustained_rps by @klu-berri in #47758
fix(timing): carry litellm_received_at across to a planted litellm_metadata bucket so the overhead header stays anchored at receive time by @klu-berri in #47790
fix(router): normalize native Responses latency per output token by @joshua-berri in #47727
fix(llms): join api_base and route with one shared helper by @mateo-berri in #47535
fix(responses)!: round-trip Anthropic compaction blocks as compaction items by @mateo-berri in #46754
fix(proxy): forward dispatched_call_type through _failure_usage_to_lift by @mateo-berri in #47815
refactor(types): replace Any with proven types in 50 files by @mateo-berri in #47818
docs: rename security.md to SECURITY.md, use security@berri.ai contact by @krrish-berri-2 in #47581
refactor(inference): rename RouteError to Error by @yujonglee-berri in #47771
fix(security): stop trusting the compromised eMudhra root CA by @ryan-crabbe-berri in #47670
feat(ui): inspect client usage for users and team members by @ishaanj-berri in #47974
feat(proxy): model_list_return_wildcard_routes setting for /v1/models by @shrey-berri in #43624
fix: recognize OpenAI context_length_exceeded errors by @shrey-berri in #47954
test(interactions): read the spec from a vendored snapshot and canary the live one daily by @mateo-berri in #48002
fix(anthropic): include null stop_sequence in native responses (#47946) by @ishaanj-berri in #47987
fix(decisions): remove the System One question count cap by @ishaanj-berri in #47985
fix(proxy): take the upgrade notice from the changelog instead of GitHub releases by @yucheng-berri in #47998
chore(cost-map): update openai gpt-6-luna decisions cache rates from the decisions guide by @berriai-litellm-provider-info-sync[bot] in #48066
fix(security): update bundled CA roots and preserve Prisma image compatibility by @shrey-berri in #47969
feat(proxy): expose wildcard model listings in general settings by @shrey-berri in #47979
fix(completions): omit stream usage unless requested (#47953) by @ishaanj-berri in #48052
feat(proxy): add LITELLM_DISABLE_LATEST_RELEASE_CHECK to stop the release check by @yucheng-berri in #48065
fix(vertex-ai): add deprecation_date to retired veo-2.0-generate-001 by @berriai-litellm-provider-info-sync[bot] in #47847
feat(ui): unify gateway and usage assistance as Aria Agent by @ishaanj-berri in #48082
fix(ci): unblock dependency security checks by @ishaanj-berri in #48055
chore(codeowners): add joshua-berri as a code owner of the model cost map by @kerry-berri in #48101
chore(signing): rotate the cosign image-signing key by @yuneng-berri in #48103
chore(cost-map): add Llama 4 Batch rates to vertex_ai scout 128e and maverick 16e maas by @kerry-berri in #48085
fix(sso): enforce free user admission without locking out existing users by @tin-berri in #48097
fix(cost-map): rolling registry audit, add elevenlabs api rates, together minimax-m3.1-flash, mistral voxtral 3 and azure_ai phi-4 dates by @devin-ai-integration[bot] in #47623
build(deps): bump mlflow, pypdf and setuptools to clear osv-scan by @ryan-crabbe-berri in #48110
fix(ui): make pass-through headers optional and align authentication by @moe-berri in #48081
feat(ui): sort virtual keys by expiry, activity, spend and limits by @moe-berri in #48074
feat(ui): sort every organizations table data column by @moe-berri in #48106
feat(ui): give Aria usage charts and PDF answer exports by @ishaanj-berri in #48094
fix(azure_ai): drop tool_reference parts from tool messages on chat by @shrey-berri in #48032
feat(ui): sort every users table data column by @moe-berri in #48105
feat(ui): sort every teams table data column by @moe-berri in #48109
feat(ui): sort budgets by reset duration by @moe-berri in #48107
fix(proxy): record spend_logs_metadata on /interactions by @igorborgest in #47848
fix(azure_ai): forward the thinking-binding-controls beta header to Foundry by @shrey-berri in #48033
feat(chat): add Moyai UI, model hub, images, and usage metrics (#40557) by @ishaanj-berri in #48108
fix(azure): accept explicit v1 Responses bases (#48072) by @ishaanj-berri in #48118
feat(proxy): add LITELLM_OFFLINE_MODE, one switch for the fetches LiteLLM makes on its own by @yucheng-berri in #48096
fix(chat): retain MCP results for follow-ups (discussion #32085) by @ishaanj-berri in #48124
fix(chat): recognize Chat in the development preview by @ishaanj-berri in #48123
fix(opencode): forward stable sessions automatically across inference routes by @moe-berri in #48121
fix(aws_polly): read the engine from the bare model name by @FahimaGold in #47963
fix(docker): update bundled LiteAdmin CA roots and aiohttp by @yuneng-berri in #48139
feat(chat): add document uploads (discussion #32085) by @ishaan-j-berri in #48137
fix(spend): await the per-request rollup enqueues and report their failures by @devin-ai-integration[bot] in #48131
fix(cost-map): declare Bedrock gpt-oss rows text-only so modality routing skips them by @devin-ai-integration[bot] in #48149
fix(router): keep image requests off deployments declared supports_vision false by @devin-ai-integration[bot] in #48152
fix(responses): honor per-call custom pricing on the native Responses API path by @joshua-berri in #48119
feat(cost): bill openai 12-hour prompt cache writes at their own rate by @kerry-berri in #48088
fix(cost): bill gemini audio input and cached audio at flex, priority and batch rates by @kerry-berri in #48086
fix(mcp): retain toolset scope when refreshing catalog authority by @tin-berri in #48095
refactor(typing): clear 1181 basedpyright Any errors across 431 files by @mateo-berri in #47789
fix(router): route decisions calls only to deployments whose provider supports decisions by @devin-ai-integration[bot] in #48175
feat(ui): configure the default Usage time range by @moe-berri in #48207
feat(providers): add OpenCode Go routing and pricing by @moe-berri in #48140
fix(ui): move Aria empty model state into the picker by @moe-berri in #48223
fix(chat): remove hosted telemetry on chat submissions by @ishaan-j-berri in #48225
fix(proxy): buffer tool discoveries in Redis and write LiteLLM_ToolTable in one statement per drain, on its own lock by @kerry-berri in #48089
refactor(typing): widen delete_nested_value to Mapping and drop two unused Span aliases by @mateo-berri in #48001
fix(ui): wire up Download CSV Template button in bulk invite by @hgnnmnn in #48199
feat(router): add opt-in session repinning after successful failover by @moe-berri in #48206
fix(router): keep latency routing alive when latency samples are unusable by @devin-ai-integration[bot] in #48216
fix(vector_stores): put azure vector store id before the api-version query on retrieve, update and delete by @devin-ai-integration[bot] in #48224
fix(anthropic): stop injecting empty-content placeholder into tool-call-only assistant turns by @tbartik in #34822
fix(bedrock): price claude_platform calls from their own cost map rows by @kerry-berri in #48150
chore(cost-map): add openai gpt-5.6-cyber 12-hour cache write rate from the pricing page by @berriai-litellm-provider-info-sync[bot] in #48246
fix(bedrock): take Nova Sonic context and output limits from the Bedrock model cards by @berriai-litellm-provider-info-sync[bot] in #48247
fix(gemini): bill gemini-2.5-flash-native-audio video input at the audio / video rate by @berriai-litellm-provider-info-sync[bot] in #48248
chore(pricing): add github_copilot/claude-opus-4.8-fast deprecation date by @berriai-litellm-provider-info-sync[bot] in #48251
fix(pricing): set azure gpt-chat-latest max_input_tokens to 128000 by @berriai-litellm-provider-info-sync[bot] in #48252
feat(responses): honor max_agentic_loops for the streamed MCP auto-execute round cap by @shivamrawat1 in #47688
feat(guardrails): add native Reco guardrail provider (pre_call), fail closed by default by @devin-ai-integration[bot] in #48243
fix(router): skip declared text-only deployments for image requests on every routing strategy by @devin-ai-integration[bot] in #48241
feat(proxy): add disable_pre_routing_error_logs to skip spend rows for pre-routing rejections by @devin-ai-integration[bot] in #48239
fix(opencode): bill OpenCode Go requests from the reported list_cost by @kerry-berri in https://github.com/BerriAI/litellm/pull/48254
fix(proxy): restore Claude gateway model discovery (#48204) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48221
fix(anthropic): release upstream messages stream on disconnect (#48197) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48220
feat(teams): show caller member spend and budget on Your Teams by @jesus-berri in https://github.com/BerriAI/litellm/pull/45674
fix(chatgpt): route uncatalogued models through responses by @moe-berri in https://github.com/BerriAI/litellm/pull/48244
fix(ui): show "Never resets" instead of "n/a" for an unset budget reset period by @jesus-berri in https://github.com/BerriAI/litellm/pull/47065
feat(models): add Cerebras Llama 3.1 hyphenated aliases (cerebras/lla… by @pfolex in https://github.com/BerriAI/litellm/pull/48158
feat(cost): parse 12-hour cache write rates for batch and ultrafast tiers by @kerry-berri in https://github.com/BerriAI/litellm/pull/48259
fix(cloudflare): clarify the chat endpoint warning (#47965) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48265
fix(decisions): preserve missing native usage (#48147) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48261
fix(groq): replay streamed reasoning replies without HTTP 400 (#47931) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48268
fix(proxy): reconcile team member total_spend with current cycle spend by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48260
fix(deepseek): pass reasoning_effort through to the API by @Hazelhof in https://github.com/BerriAI/litellm/pull/38736
feat(mistral): add zai-glm-5-3-flash to the cost map by @kerry-berri in https://github.com/BerriAI/litellm/pull/48270
fix(proxy): keep JSON integers outside the 64-bit range exact when parsing request bodies by @jeevan1098 in https://github.com/BerriAI/litellm/pull/48213
revert(proxy): revert large-integer request parsing by @moe-berri in https://github.com/BerriAI/litellm/pull/48289
build(docker): refresh base images and ship only runtime files in component images by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48279
fix: support Python 3.15 by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48359
build(deps): bump webpki-roots and webpki-root-certs to 1.0.11 by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48301
test(streaming): pin that a stream stuck on one chunk closes its provider stream and logs one failure by @businessarshgoyal in https://github.com/BerriAI/litellm/pull/38753
fix(ui): restore clipboard copying in key and log details by @tin-berri in https://github.com/BerriAI/litellm/pull/48425
fix(mcp): report oversized SSE catalog errors by @tin-berri in https://github.com/BerriAI/litellm/pull/48424
chore(cost-map): add openai gpt-6.1-terra from the pricing page by @kerry-berri in https://github.com/BerriAI/litellm/pull/48429
fix(otel): decide per-team Arize sampling from the trace id so every pod agrees by @yucheng-berri in https://github.com/BerriAI/litellm/pull/47468
fix(vertex_ai): reuse rerank credential cache by @tin-berri in https://github.com/BerriAI/litellm/pull/48430
fix(ui): render string inputs in general settings by @tin-berri in https://github.com/BerriAI/litellm/pull/48431
fix(guardrails): exclude litellm_metadata from pass-through guardrail scan text by @mrinal-berri in https://github.com/BerriAI/litellm/pull/48276
fix(azure): set azure_ai/Microsoft-Decision-1 max_input_tokens to 32768 by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48434
chore(cost-map): add openai batch and ultrafast 12-hour cache write rates from the pricing page by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48433
feat(bedrock): add global. and us. openai.gpt-6.1-terra rows by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48441
fix(vertex_ai): bound shared credential cache retention by @tin-berri in https://github.com/BerriAI/litellm/pull/48442
chore(deps): remove expired boto3 botocore and grpcio overrides by @shrey-berri in https://github.com/BerriAI/litellm/pull/48447
feat(cost-map): add baseten elephant-1 and mammoth-1 decision models by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48453
fix(elevenlabs): map wav and pcm TTS formats to 24kHz output formats by @moe-berri in https://github.com/BerriAI/litellm/pull/48455
feat(pricing): add github_copilot/gpt-6.1-terra and raise gpt-6 max_input_tokens by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48450
test(proxy): make tokenizer offload check deterministic by @tin-berri in https://github.com/BerriAI/litellm/pull/48458
fix(proxy): preserve limiter state across SSE keepalive tasks by @tin-berri in https://github.com/BerriAI/litellm/pull/48452
fix(proxy): reject negative RPM and TPM request limits by @tin-berri in https://github.com/BerriAI/litellm/pull/48444
fix(ui): keep Aria accessible in the dashboard header by @moe-berri in https://github.com/BerriAI/litellm/pull/48454
fix(proxy): scope dashboard search grants and skip the sentinel team lookup by @moe-berri in https://github.com/BerriAI/litellm/pull/48464
fix(gemini): restore image tool results for Gemini 2.5 by @tin-berri in https://github.com/BerriAI/litellm/pull/48428
ci: remove the Together AI and weekly cost map sync workflows by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48473
fix(mcp): advertise valid OAuth resources for toolset URLs by @tin-berri in https://github.com/BerriAI/litellm/pull/48474
build(deps): drop expired exclude-newer pins for certifi, mlflow and pypdf by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48461
feat(ui): simplify auto router setup with deployed model defaults by @moe-berri in https://github.com/BerriAI/litellm/pull/48451
fix(router): stop concurrent deployment deletes from crashing deployment lookups by @yucheng-berri in https://github.com/BerriAI/litellm/pull/47680
fix(ui): show search tool errors instead of empty results by @moe-berri in https://github.com/BerriAI/litellm/pull/48478
feat(chatgpt): forward client subscriptions through the gateway by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/41522
fix(mcp): send tool-call stream headers once the call is authorized by @devanshc777 in https://github.com/BerriAI/litellm/pull/47897
fix(models): declare Azure audio snapshot capabilities by @tin-berri in https://github.com/BerriAI/litellm/pull/48475
fix(http_handler): name the timeout phase and limit that fired instead of "after None seconds" by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48250
fix(proxy): validate JSON object bodies before request handling by @tin-berri in https://github.com/BerriAI/litellm/pull/48438
fix(proxy)!: gate key budget_duration and budget_id updates like max_budget by @yucheng-berri in https://github.com/BerriAI/litellm/pull/47681
chore(lint): add google-convention docstring rules to strict ruff gate by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48245
fix(router): isolate forwarded Anthropic subscription cooldowns by @moe-berri in https://github.com/BerriAI/litellm/pull/48485
refactor(cost-map): drop the legacy max_tokens field in favor of max_input_tokens and max_output_tokens by @kerry-berri in https://github.com/BerriAI/litellm/pull/48471
feat(rust): add typed database query foundation by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48214
feat(rust): add typed key spend query by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48217
feat(rust): add typed user spend query by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48229
feat(rust): add typed team spend query by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48230
feat(rust): add typed end-user spend query by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48231
feat(rust): add transaction-scoped team lock query by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48232
feat(rust): add typed team-member spend upsert by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48233
feat(rust): add typed spend-log insert query by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48234
feat(rust): add typed daily user spend upsert by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48235
feat(rust): add typed daily team spend upsert by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48236
feat(rust): add typed daily end-user spend upsert by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48237
docs: document incremental database access migration by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48476
feat(ui): budget period tooltip on key, team, user and org spend tiles by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48487
fix(router): serve a team's wildcard deployment before a proxy-wide wildcard by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48490
fix(team): keep the team reset schedule when a temp budget increase is set on a member inheriting the team default by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48495
fix(mcp): reject revoked browser sessions during gateway authorization by @tin-berri in https://github.com/BerriAI/litellm/pull/48494
fix(router): keep DB-only deployments priced across price data reloads by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48486
fix(ui): show the budget period date range instead of "cycle" on team member spend by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48488
test(routing): a team's wildcard serves its keys before a proxy-wide wildcard of the same pattern by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48501
fix(prometheus): count provider-side 401s in LLM failure and deployment metrics by @yucheng-berri in https://github.com/BerriAI/litellm/pull/46668
fix(ui): show endpoint-qualified models in Decisions Playground by @tin-berri in https://github.com/BerriAI/litellm/pull/48504
feat(ui): persist model group alias display ordering by @tin-berri in https://github.com/BerriAI/litellm/pull/48500
feat(abliteration): add model pricing and capabilities by @moe-berri in https://github.com/BerriAI/litellm/pull/48498
feat(refuseless): add provider and model pricing by @moe-berri in https://github.com/BerriAI/litellm/pull/48497
feat(rust): add redis-client crate for standalone Redis by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48269
feat(redis-client): support Redis Cluster by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48271
feat(redis-client): run Lua scripts and pipelines with per-op errors by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48273
feat(rust): add circuit-breaker crate and guard redis-client with it by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48277
feat(gateway): connect to Redis from config.yaml by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48284
fix(decisions): honor base_url for custom backends by @tin-berri in https://github.com/BerriAI/litellm/pull/48505
fix(ui): persist model parameter deletions after refresh by @tin-berri in https://github.com/BerriAI/litellm/pull/48456
refactor(llms): rename model-version transformation files to family names by @mateo-berri in https://github.com/BerriAI/litellm/pull/48509
fix(anthropic): raise DEFAULT_ANTHROPIC_CHAT_MAX_TOKENS from 4096 to 128000 by @kerry-berri in https://github.com/BerriAI/litellm/pull/48502
feat(decisions): support LM Studio decision models by @jodok in https://github.com/BerriAI/litellm/pull/48448
fix(token-counter): count inline auto images safely (#48422) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48526
fix(chatgpt): forward client credentials for image generation by @moe-berri in https://github.com/BerriAI/litellm/pull/48522
fix(proxy): include Claude web search fees in streamed usage cost by @tin-berri in https://github.com/BerriAI/litellm/pull/48528
fix(vertex): keep Messages authentication off the event loop by @tin-berri in https://github.com/BerriAI/litellm/pull/48523
fix(decisions): preserve missing probabilities (#48449) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48524
fix(ui): honor dark and local logos across Connect and Chat by @tin-berri in https://github.com/BerriAI/litellm/pull/48532
fix(azure): add keyless Decisions authentication (#48386) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48529
fix(databricks): reuse base OAuth for decisions (#48395) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48527
fix(messages): honor OpenAI stop_sequences (#37118) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48534
fix(router): preserve tag fail-open policy through affinity and cooldown by @shrey-berri in https://github.com/BerriAI/litellm/pull/48470
refactor(anthropic): rename llms/anthropic/pass_through to llms/anthropic/messages by @mateo-berri in https://github.com/BerriAI/litellm/pull/48510
fix(langfuse_otel)!: derive trace-level name, input, output and tags on traces litellm starts by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/42705
refactor(mcp): drop the experimental label from the MCP server and client paths by @mateo-berri in https://github.com/BerriAI/litellm/pull/48513
fix: correct Palm 'decommissioned' typo in error message by @ninglinLiu in https://github.com/BerriAI/litellm/pull/46171
fix(responses): replay web_search_call input items that have no action by @prassu017 in https://github.com/BerriAI/litellm/pull/48417
feat: generate provider_endpoints_support.json from the cost map by @kerry-berri in https://github.com/BerriAI/litellm/pull/48619
feat(bedrock): price native guardrailConfig trace usage into spend and x-litellm-response-cost by @kerry-berri in https://github.com/BerriAI/litellm/pull/48278
feat(proxy): refuse migrations that existing rows would make fail, before the first one runs by @kerry-berri in https://github.com/BerriAI/litellm/pull/48493
fix(anthropic): estimate initial input tokens for bridged message streams by @moe-berri in https://github.com/BerriAI/litellm/pull/48618
fix(logging): match extra_spend_tag_headers case-insensitively by @moe-berri in https://github.com/BerriAI/litellm/pull/48620
fix(vertex-ai): drop the floor retirement date from gemini-3.1-flash-lite-image by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48587
fix(proxy): return 403 naming the Enterprise licence for get_spend_routes without a licence by @moe-berri in https://github.com/BerriAI/litellm/pull/48623
fix(token_counter): count Anthropic-format tool_choice without KeyError by @moe-berri in https://github.com/BerriAI/litellm/pull/48624
fix(bedrock): correct gpt max_input_tokens and add Mantle token counting for claude-opus-5-5 by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48479
fix(security)!: build proxy-only failure logs from an allowlist and drop caller guardrail info by @yucheng-berri in https://github.com/BerriAI/litellm/pull/48599
fix(ui): expose the authorization MCP auth type in create and edit forms by @moe-berri in https://github.com/BerriAI/litellm/pull/48629
feat(hosted_vllm): accept images on decision deployments by @moe-berri in https://github.com/BerriAI/litellm/pull/48621
fix(mcp): load deprecated package aliases on demand by @mateo-berri in https://github.com/BerriAI/litellm/pull/48615
fix(auth): log routine 401 rejections as one warning without a traceback by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48613
ci: spend-logs-guard requires kerry-berri approval on LiteLLM_SpendLogs schema and migration changes by @kerry-berri in https://github.com/BerriAI/litellm/pull/48518
fix(anthropic): make opening usage estimates opt-in by @moe-berri in https://github.com/BerriAI/litellm/pull/48633
fix(cost-map): add 2026-11-06 deprecation_date to two together_ai rows by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48651
fix(aiohttp_transport): close an owned aiohttp session when its evicted transport is collected by @kerry-berri in https://github.com/BerriAI/litellm/pull/48650
fix(complexity_router): keep the held user_turn route for Codex compaction requests by @moe-berri in https://github.com/BerriAI/litellm/pull/48627
fix(ui): percent-encode request id in spend log detail URL by @moe-berri in https://github.com/BerriAI/litellm/pull/48630
refactor(proxy-extras): drop the spend-log api_key special case from the row check remedy by @kerry-berri in https://github.com/BerriAI/litellm/pull/48636
feat(router): respect caller output budgets in complexity routing by @tin-berri in https://github.com/BerriAI/litellm/pull/48632
fix(mcp): enforce deny-first tool policy consistently by @tin-berri in https://github.com/BerriAI/litellm/pull/48625
fix(mcp): accept YAML OpenAPI documents by @tin-berri in https://github.com/BerriAI/litellm/pull/48622
fix(proxy): release concurrency slots when synthetic streams complete by @tin-berri in https://github.com/BerriAI/litellm/pull/48628
fix(proxy): keep every startup line JSON when json_logs is on by @yucheng-berri in https://github.com/BerriAI/litellm/pull/47035
fix(router): keep a lone deployment warm under router allowed_fails unless the policy names the exception by @kerry-berri in https://github.com/BerriAI/litellm/pull/48644
fix(streaming): dict-usage arm drops prompt_tokens_details/completion_tokens_details its sibling arms preserve by @lets-order-some-fries in https://github.com/BerriAI/litellm/pull/36678
test: remove checks that pin implementation details and add a shadcn testing rule by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48663
docs(router): clarify heuristic v2 keyword settings by @tin-berri in https://github.com/BerriAI/litellm/pull/48666
feat(ui): add LITELLM_UI_USE_SI_TERMINOLOGY to show SI instead of AI in the Admin UI by @yuneng-berri in https://github.com/BerriAI/litellm/pull/47440
fix(ui): keep vendor names and agent instructions in AI wording under SI terminology by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48670
fix(proxy): abort the request with a 499 when the client disconnects during the body upload by @kerry-berri in https://github.com/BerriAI/litellm/pull/48640
fix(ui): omit blank expiry from key edit and regenerate payloads by @moe-berri in https://github.com/BerriAI/litellm/pull/48634
fix(chatgpt): send system input items to Codex as developer by @moe-berri in https://github.com/BerriAI/litellm/pull/48668
fix(mcp): refuse calls to tools hidden by discovery guardrails by @tin-berri in https://github.com/BerriAI/litellm/pull/48669
feat(search): add native xAI X Search support (#38505) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48677
fix(decisions): catch ValidationError from validate_json in transform_decisions_response (#48660) by @16niraj in https://github.com/BerriAI/litellm/pull/48662
ci: run spend-logs-guard and cost-map-guard without installing the litellm project by @kerry-berri in https://github.com/BerriAI/litellm/pull/48683
fix(bedrock): add token_count_model to three Mantle-native Claude rows by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48682
fix: honor empty mock completion responses (#48162) by @ishaan-j-berri in https://github.com/BerriAI/litellm/pull/48266
test(cost): cover unpriced provider slugs and explicit zero rates by @eduardopessin in https://github.com/BerriAI/litellm/pull/42467
feat(morph): support open models and cached input pricing by @moe-berri in https://github.com/BerriAI/litellm/pull/48676
fix: preserve media across Responses and Gemini conversions by @tin-berri in https://github.com/BerriAI/litellm/pull/48484
fix(guardrails)!: redact secrets in tool traffic and system prompts by @yucheng-berri in https://github.com/BerriAI/litellm/pull/47122
test: update stale integration and guardrail e2e tests to intended behavior by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48665
feat(helm): route extra ingress paths to custom services by @tin-berri in https://github.com/BerriAI/litellm/pull/48674
fix(router): keep session tier pins coherent across workers by @tin-berri in https://github.com/BerriAI/litellm/pull/48645
chore: rebuild Admin UI bundle for 1.109.0 by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48692
feat(passthrough): let upstreams report the model name via x-litellm-upstream-model by @mrinal-berri in https://github.com/BerriAI/litellm/pull/48275
chore: bump litellm-enterprise 0.1.79 -> 0.1.80, litellm 1.108.0 -> 1.109.0 by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48673
fix(responses): preserve typed parts when logprobs is omitted by @tin-berri in https://github.com/BerriAI/litellm/pull/48686
fix(embeddings): forward Gemini and Vertex request timeouts by @tin-berri in https://github.com/BerriAI/litellm/pull/48690
fix(databricks): price the openjev decision model at its published 0.6 DBU per 1M input tokens by @kerry-berri in https://github.com/BerriAI/litellm/pull/48694
fix(tavily): return raw_content, score, favicon and the response envelope from Tavily search by @mrinal-berri in https://github.com/BerriAI/litellm/pull/48684
fix(ui): keep guardrail wizard navigation within the dialog by @tin-berri in https://github.com/BerriAI/litellm/pull/48688
fix(langfuse otel): emit request metadata under langfuse.observation.metadata and langfuse.trace.metadata on the OTEL v2 Langfuse mapper by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/42697
fix(cohere): map string stop parameter to list of strings for stop_sequences by @Pyhroff in https://github.com/BerriAI/litellm/pull/47091
fix(anthropic): count search results and reject unsupported token-count blocks by @kerry-berri in https://github.com/BerriAI/litellm/pull/48699
fix(proxy)!: paginate customer lists and usage filters by @tin-berri in https://github.com/BerriAI/litellm/pull/48700
fix(responses): skip failed spend-log turns when rebuilding a session from previous_response_id by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48647
fix(responses): reject or drop Responses params the chat completions bridge cannot translate by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48655
fix(ui): prefill the key edit Organization picker from org_id by @mateo-berri in https://github.com/BerriAI/litellm/pull/39535
fix(responses): keep every MCP tool round in streaming follow-up input by @shivamrawat1 in https://github.com/BerriAI/litellm/pull/47698
docs(readme): link the LiteLLM OSS Champions program from Contributing by @zoveki in https://github.com/BerriAI/litellm/pull/48734
test(harness): audit #47530 by @norkelbyte in https://github.com/BerriAI/litellm/pull/48751
test(sagemaker): audit #47495 by @rj-cairn in https://github.com/BerriAI/litellm/pull/48767
test(gemini): audit #47542 by @zoveki in https://github.com/BerriAI/litellm/pull/48758
test(proxy): pin the max_parallel_requests slot release when a streamed /v1/messages MCP auto-execute ends by @mateo-berri in https://github.com/BerriAI/litellm/pull/47788
fix(cost-map): drop the gpt-5-nano priority input price OpenAI does not sell by @tessellark in https://github.com/BerriAI/litellm/pull/48763
test(spend): audit #47815 by @norkelbyte in https://github.com/BerriAI/litellm/pull/48774
feat: add ai& (aiand) provider by @islomjon-aiand in https://github.com/BerriAI/litellm/pull/38958
fix(router)!: count an explicit allowed_fails of 3 and keep fallback deployments out of cooldown by @HardMax71 in https://github.com/BerriAI/litellm/pull/46183
test(proxy): audit #47418 by @tavren27 in https://github.com/BerriAI/litellm/pull/48766
fix(router): let affinity pins yield to deployments that already failed this request by @aidanasingh in https://github.com/BerriAI/litellm/pull/32309
test(llms): audit #47535 by @tessellark in https://github.com/BerriAI/litellm/pull/48786
test(responses): audit #47504 by @norkelbyte in https://github.com/BerriAI/litellm/pull/48787
test(spend): audit #47353 by @rj-cairn in https://github.com/BerriAI/litellm/pull/48785
test(proxy): audit #46815 by @m6atlas in https://github.com/BerriAI/litellm/pull/48768
fix(bedrock): retry without strict when Bedrock rejects strict tool schemas under drop_params by @m6atlas in https://github.com/BerriAI/litellm/pull/48733
test(files): audit #40621 by @zoveki in https://github.com/BerriAI/litellm/pull/48788
test(anthropic, bedrock): audit #47489 by @zoveki in https://github.com/BerriAI/litellm/pull/48782
fix(anthropic): let drop_params drop speed on claude-opus-4-7 after Anthropic removed its fast mode by @norkelbyte in https://github.com/BerriAI/litellm/pull/48764
test(token_counter): audit #47586 by @simonp9 in https://github.com/BerriAI/litellm/pull/48800
test(anthropic): audit #47478 by @paveln83 in https://github.com/BerriAI/litellm/pull/48803
fix(router): continue Responses mid-stream fallbacks on models without assistant prefill by @tin-berri in https://github.com/BerriAI/litellm/pull/47672
test(router): audit #47498 by @mika-lt in https://github.com/BerriAI/litellm/pull/48799
test(decisions): audit #47508 by @tavren27 in https://github.com/BerriAI/litellm/pull/48777
test(router): audit #47376 by @norkelbyte in https://github.com/BerriAI/litellm/pull/48792
test(proxy): audit #47358 by @norkelbyte in https://github.com/BerriAI/litellm/pull/48784
test(router): audit #41368 by @tavren27 in https://github.com/BerriAI/litellm/pull/48794
refactor(fallback-generalizations)!: remove the legacy extends and mixed-rule shim by @norkelbyte in https://github.com/BerriAI/litellm/pull/48737
fix(streaming): exact-match the SSE [DONE] marker in BaseModelResponseIterator by @michelligabriele in https://github.com/BerriAI/litellm/pull/47832
test(nvidia_riva): audit #41021 by @quietjules in https://github.com/BerriAI/litellm/pull/48801
feat(proxy): per-provider and per-model rpm/tpm requirements for enforce_project_model_quota by @mateo-berri in https://github.com/BerriAI/litellm/pull/47693
feat(cli): keep the ChatGPT login with lite codex --subscription by @rj-cairn in https://github.com/BerriAI/litellm/pull/48741
test(proxy): audit #47352 by @norkelbyte in https://github.com/BerriAI/litellm/pull/48804
test(mcp): audit #47422 by @tavren27 in https://github.com/BerriAI/litellm/pull/48821
test(scim): audit #47544 by @zoveki in https://github.com/BerriAI/litellm/pull/48791
build(rust): upgrade reqwest to 0.13 and move MCP out of litellm-http by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48039
fix(http): honor transport settings for Rust WebSockets by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/47968
fix(vertex_ai): send Vertex Claude labels as X-Vertex-AI-Labels header instead of the request body by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48753
fix(cost-map): rolling provider pricing, capability and retirement audit by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48188
fix(proxy/auth): apply max_end_user_budget_id to end users without a row on cached keys by @moe-berri in https://github.com/BerriAI/litellm/pull/48704
refactor(http): build outbound requests as reqwest::Request by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48041
refactor(rust): adopt one error discipline across litellm-rust with source chains and Report by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48042
feat(rust): one typed decline from core up to each host by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/47772
fix(huggingface): stop forcing truncate=false on TEI rerank and pass truncate params through by @moe-berri in https://github.com/BerriAI/litellm/pull/48706
fix(ollama): forward embedding credentials, headers and timeout by @moe-berri in https://github.com/BerriAI/litellm/pull/48707
fix(router)!: isolate child tier pins in user turn mode by @tin-berri in https://github.com/BerriAI/litellm/pull/48701
fix(refuseless): update GLM 5.3 and DeepSeek V4.1 Flash prices by @moe-berri in https://github.com/BerriAI/litellm/pull/48844
fix(ui): match simple auto router models to advanced setup by @moe-berri in https://github.com/BerriAI/litellm/pull/48851
fix(ui): show reasoning in the Playground for /v1/responses by @moe-berri in https://github.com/BerriAI/litellm/pull/48850
fix(responses): forward top-level moderation param to OpenAI by @moe-berri in https://github.com/BerriAI/litellm/pull/48708
fix(proxy): store boolean settings as JSON booleans on /config/field/update by @shrey-berri in https://github.com/BerriAI/litellm/pull/48595
fix(azure): map context_length_exceeded error code to ContextWindowExceededError by @shrey-berri in https://github.com/BerriAI/litellm/pull/48616
fix(images): honor max_retries and prevent Router retry amplification by @moe-berri in https://github.com/BerriAI/litellm/pull/48705
fix(vertex-ai): preserve image size aspect ratios by @shrey-berri in https://github.com/BerriAI/litellm/pull/48463
fix(otel)!: label configured pass-through routes litellm.passthrough instead of chat by @yucheng-berri in https://github.com/BerriAI/litellm/pull/48675
fix(chatgpt): derive the upstream session id from prompt_cache_key by @occ in https://github.com/BerriAI/litellm/pull/42014
fix(cost-map)!: price Mistral Large 4 at its list price now that the launch sale ended by @zoveki in https://github.com/BerriAI/litellm/pull/48825
fix(anthropic-messages): merge deployment and client anthropic-beta instead of overwriting by @shrey-berri in https://github.com/BerriAI/litellm/pull/48594
feat(proxy): log one line per token count call by @mateo-berri in https://github.com/BerriAI/litellm/pull/47657
fix(otel): suppress spans for excluded proxy requests by @tin-berri in https://github.com/BerriAI/litellm/pull/48720
fix(baseten): price the two decision models at their published input rates by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48849
fix(router)!: resolve trusted JEV connection environment references by @tin-berri in https://github.com/BerriAI/litellm/pull/48717
chore(cost-map): fill fireworks spot router capability fields from base models by @kerry-berri in https://github.com/BerriAI/litellm/pull/48858
fix(mistral): replay assistant reasoning as a Mistral thinking chunk by @rj-cairn in https://github.com/BerriAI/litellm/pull/48778
fix(images): normalize absent images in async edits by @tin-berri in https://github.com/BerriAI/litellm/pull/48872
fix(count_tokens)!: audit #47441 and keep the framed estimate for Bedrock count models from the cost map by @rj-cairn in https://github.com/BerriAI/litellm/pull/48805
fix(responses)!: answer 400 before counting a lone Responses built-in tool on Claude by @tessellark in https://github.com/BerriAI/litellm/pull/48798
fix(router): drop the deployment rpm filter from enable_pre_call_checks by @kerry-berri in https://github.com/BerriAI/litellm/pull/48873
fix(router): pass lifecycle events through on background OpenAI responses streams that cannot fall back by @simonp9 in https://github.com/BerriAI/litellm/pull/48817
fix(batches): stop unbillable batches from blocking file deletes and new-batch polling by @tessellark in https://github.com/BerriAI/litellm/pull/48752
fix(audio): reject unsupported transcription providers with 400 by @shrey-berri in https://github.com/BerriAI/litellm/pull/47971
fix(azure): share one Entra token request per worker and audit #47348 by @tessellark in https://github.com/BerriAI/litellm/pull/48769
fix(cost-map): add mistral/deepseek-v4-1-flash-1017 by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48879
feat(decisions): support Ollama decision models by @moe-berri in https://github.com/BerriAI/litellm/pull/48871
fix(gemini)!: accept Google's current Managed Agents fields on create and interactions by @zoveki in https://github.com/BerriAI/litellm/pull/48750
fix(openai)!: send the whole reasoning_effort dict when chat is bridged to Responses by @m6atlas in https://github.com/BerriAI/litellm/pull/48807
fix(proxy): account for media in combined TPM reservations by @tin-berri in https://github.com/BerriAI/litellm/pull/48865
feat(providers): add ZeroGPU as an OpenAI-compatible provider by @amaan-ai20 in https://github.com/BerriAI/litellm/pull/43874
refactor(ui): use the shadcn Kbd primitive for keyboard hints by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48885
fix(pricing): sync databricks model prices by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48884
fix(cli): let the proxy decide when a classic lite login token has expired by @mateo-berri in https://github.com/BerriAI/litellm/pull/47712
fix(aiohttp_openai): forward service_tier in transform_response by @TemoSulava in https://github.com/BerriAI/litellm/pull/47263
refactor(rust-bridge): require the native extension by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48852
fix(streaming): surface upstream abort as server errors (LIT-10159) by @sorgfresser in https://github.com/BerriAI/litellm/pull/48328
fix(ui): bundle the Aria avatar so it loads in every deployment by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48890
fix(responses)!: bill partial usage on mid-stream failures and report bridged /v1/messages failures to the proxy by @mateo-berri in https://github.com/BerriAI/litellm/pull/47537
fix(files)!: scope provider and model file lists to the caller's managed files under require_managed_files by @mateo-berri in https://github.com/BerriAI/litellm/pull/47682
fix(anthropic): map model_context_window_exceeded stop reason to length by @moe-berri in https://github.com/BerriAI/litellm/pull/48874
feat(accounting): add accounting contracts, PricingSnapshot and ledger schema (LIT-10193) by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48862
fix(ci): spend logs guard requires the approval to cover the head's spend-log lines by @kerry-berri in https://github.com/BerriAI/litellm/pull/48904
fix(guardrails)!: make hide-secrets honor skip_tool_message_in_guardrail by @yucheng-berri in https://github.com/BerriAI/litellm/pull/48689
fix(pricing): add databricks priority pay per token prices by @kerry-berri in https://github.com/BerriAI/litellm/pull/48891
fix(ci): keep the spend logs guard approval through a later COMMENTED review (LIT-10242) by @kerry-berri in https://github.com/BerriAI/litellm/pull/48901
fix(logging): bill each SDK call its own native Bedrock guardrail trace when calls share one metadata dict by @kerry-berri in https://github.com/BerriAI/litellm/pull/48923
fix(aiohttp_transport): keep the owned session open while a response stream is still being read (LIT-10237) by @kerry-berri in https://github.com/BerriAI/litellm/pull/48920
ci(spend-logs-guard): run only on pull_request_target, re-run by hand after approving by @kerry-berri in https://github.com/BerriAI/litellm/pull/48941
fix(router): cool down failing deployments when fallbacks use string targets (LIT-10238) by @kerry-berri in https://github.com/BerriAI/litellm/pull/48913
fix(migration_data_checks): size btree entries by the bytes Postgres stores (LIT-10244) by @kerry-berri in https://github.com/BerriAI/litellm/pull/48911
feat(daily-activity): per-guardrail spend breakdown in user, team and organization daily activity by @mrinal-berri in https://github.com/BerriAI/litellm/pull/48702
feat(ui): search and multi-select tools in the MCP toolset editor by @mrinal-berri in https://github.com/BerriAI/litellm/pull/48601
fix(gemini): preserve every grounding web source by @SilverMarcs in https://github.com/BerriAI/litellm/pull/48745
feat(proxy)!: model family and access group entries in model_max_budget by @mateo-berri in https://github.com/BerriAI/litellm/pull/47722
test(routing): audit #46601 by @tessellark in https://github.com/BerriAI/litellm/pull/48910
feat(skills): add a litellm-setup agent skill by @mateo-berri in https://github.com/BerriAI/litellm/pull/47731
fix(ci): flag any schema.prisma change inside model LiteLLM_SpendLogs in the spend logs guard by @kerry-berri in https://github.com/BerriAI/litellm/pull/48905
feat(router): add deployment rpd and tpd daily limits to enforce_model_rate_limits by @kerry-berri in https://github.com/BerriAI/litellm/pull/48875
feat(tencent): add Seedream image generation and editing through TokenHub by @tessellark in https://github.com/BerriAI/litellm/pull/48735
test(spend): audit #47506 by @quietjules in https://github.com/BerriAI/litellm/pull/48919
feat(proxy): accept request priority from the x-litellm-priority header by @mubashir1osmani in https://github.com/BerriAI/litellm/pull/47618
fix(websearch): strip Claude Code instructional prefix in short-circuit search by @agustin18 in https://github.com/BerriAI/litellm/pull/43458
feat(accounting): add durable event ledger writer by @joshua-berri in https://github.com/BerriAI/litellm/pull/48883
refactor(types): replace Any with proven types in 28 files by @rj-cairn in https://github.com/BerriAI/litellm/pull/48903
test(gemini): cover every web source for a grounded passage by @moe-berri in https://github.com/BerriAI/litellm/pull/48861
fix(router): reject string fallback targets at Router init (LIT-10238) by @kerry-berri in https://github.com/BerriAI/litellm/pull/48943
fix(vertex-ai): add batch above 200k and image output batch prices for gemini models by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48945
fix(streaming): keep merged think tags balanced and split SSE lines only on CR/LF by @rj-cairn in https://github.com/BerriAI/litellm/pull/48755
refactor(router): fence the router surface and document the router-core boundary by @narmstrong-berri in https://github.com/BerriAI/litellm/pull/48839
test(extraction): audit #47473 by @paveln83 in https://github.com/BerriAI/litellm/pull/48898
ci: run zizmor with --no-ignores and move the spend logs and cost map guards to pull_request by @yuneng-berri in https://github.com/BerriAI/litellm/pull/48947
refactor(images): remove dall-e-2 and dall-e-3 code paths by @kerry-berri in https://github.com/BerriAI/litellm/pull/48703
test(proxy): audit #46997 by @zoveki in https://github.com/BerriAI/litellm/pull/48918
test(responses): audit #47445 by @tavren27 in https://github.com/BerriAI/litellm/pull/48902
fix(vertex_ai)!: send the caller's anthropic-version in the Vertex Claude /v1/messages body, not the header by @zoveki in https://github.com/BerriAI/litellm/pull/48814
refactor(model-catalog): parse, lower and validate the catalog into typed prices by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48853
test(router): set string fallback targets after Router init in the sdk cooldown cases (LIT-10238) by @kerry-berri in https://github.com/BerriAI/litellm/pull/48954
fix(health): ask OpenAI and Azure audio chat models for audio output in the health check probe by @mateo-berri in https://github.com/BerriAI/litellm/pull/47730
fix(proxy): stop log lines from leaving terminal colors on by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48966
fix(bedrock): omit system prompt from audio transcription and report its token usage by @david-koh in https://github.com/BerriAI/litellm/pull/41565
fix(logging): stop dumping the request body at every layer under --detailed_debug by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/48973
fix(stream): the last usage frame carries litellm's bill in usage.cost by @kerry-berri in https://github.com/BerriAI/litellm/pull/48955
fix(responses): drop params compact rejects and send Azure compaction to the v1 route by @simonp9 in https://github.com/BerriAI/litellm/pull/48876
fix(guardrails): scan Responses tool results in texts-mode guardrails by @mika-lt in https://github.com/BerriAI/litellm/pull/48827
fix(guardrails): skip embeddings input in google_text_moderation by @simonp9 in https://github.com/BerriAI/litellm/pull/48934
feat!: add StreamLake provider metadata by @ZHOUKAILIAN in https://github.com/BerriAI/litellm/pull/41006
fix(proxy): let typesafe/* allowlists cover /typesafe Jev calls and list Jev in the model picker by @mateo-berri in https://github.com/BerriAI/litellm/pull/47697
feat(router): daily_reset_time picks when deployment rpd and tpd daily limits start over by @kerry-berri in https://github.com/BerriAI/litellm/pull/48957
test(unit): keep unix socket paths in unit tests under the macOS AF_UNIX limit by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/45040
feat(ocr)!: support outputContentFormat=markdown for Azure Document Intelligence by @zoveki in https://github.com/BerriAI/litellm/pull/48926
feat(azure_ai): reach Foundry agents through the project Responses API by @mateo-berri in https://github.com/BerriAI/litellm/pull/47717
docs(adrs): add ADRs for the Python SDK and proxy, starting with provider usage extras by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/36668
fix(responses): drop the null created_by OpenAI rejects when output items are replayed by @tessellark in https://github.com/BerriAI/litellm/pull/48783
fix(cli): give lite codex aliases of a Codex-known model that model's catalog entry by @tessellark in https://github.com/BerriAI/litellm/pull/48894
fix(proxy)!: enforce an end user's per-model budget from their linked budget by @mika-lt in https://github.com/BerriAI/litellm/pull/48815
fix(context_management)!: count the system prompt toward the clear_tool_uses input_tokens trigger by @AlonSoko in https://github.com/BerriAI/litellm/pull/47857
fix(google_genai): read one-object and null spellings of repeated fields in countTokens and generateContent by @zoveki in https://github.com/BerriAI/litellm/pull/48771
fix(proxy)!: stop labelling management and health rejections as auth failures by @tavren27 in https://github.com/BerriAI/litellm/pull/48739
test(json_fragment_accumulator): assert peak allocation instead of a drain timing ratio by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/43370
fix(credentials): keep the stored secret when PATCH echoes its masked read-back by @tessellark in https://github.com/BerriAI/litellm/pull/48950
fix(responses): call the renamed drops_params helper on the compact path by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/49030
docs(qa): allow focused proof for mechanical PRs by @mateo-berri in https://github.com/BerriAI/litellm/pull/49031
fix(proxy): enforce team and key Allowed Vector Stores when no vector store registry is loaded by @paveln83 in https://github.com/BerriAI/litellm/pull/48915
build(deps): bump langgraph and tornado and add dated scan ignores to clear osv-scan and image-scan by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/48302
fix(responses): stream every citation batch before completion by @SilverMarcs in https://github.com/BerriAI/litellm/pull/48747
fix(token_counter): count OpenAI models newer than tiktoken's table with o200k_base by @rj-cairn in https://github.com/BerriAI/litellm/pull/48728
fix(cost_map): mark fireworks_ai/minimax-m3 as vision and pin capabilities verified against live calls by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/43390
feat(responses): support Codex client tool_search on Anthropic and other bridged providers by @mateo-berri in https://github.com/BerriAI/litellm/pull/48972
fix(proxy)!: list every priced Vertex family under a vertex_ai/* grant by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/41959
fix(responses): keep each Codex tool once when its tool list repeats across turns by @ryan-crabbe-berri in https://github.com/BerriAI/litellm/pull/49006
feat(providers): add BigModel support and model pricing by @AaronHowell in https://github.com/BerriAI/litellm/pull/48563
fix(health)!: openai realtime health check reports healthy with a missing or invalid api key by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/41233
fix(responses): replay the original Responses item order, phase, and boundaries on the next chat turn by @simonp9 in https://github.com/BerriAI/litellm/pull/48912
fix(responses): resume background response streams with stream and starting_after by @paveln83 in https://github.com/BerriAI/litellm/pull/48819
test(integration): scope the signoz, otel conversation id, and series cap proxy fixtures to their module by @rj-cairn in https://github.com/BerriAI/litellm/pull/48965
fix(anthropic): drop the non-spec custom object from /v1/messages tool definitions by @zoveki in https://github.com/BerriAI/litellm/pull/48893
fix(responses): drop a malformed prompt_cache_breakpoint under drop_params on the direct /v1/responses path by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/45330
test: fix flaky spend-log cleanup budget, login lockout, and wedged Redis tests by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/47108
fix(vertex_ai): build the cachedContents collection URL under a custom api_base by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/44783
fix(github_copilot): treat a malformed api-key.json like a missing one by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/44723
feat(proxy)!: add suppress_startup_output to turn off startup prints by @mateo-berri in https://github.com/BerriAI/litellm/pull/47708
feat(proxy): let a Claude Code gateway user choose the session team by @simonp9 in https://github.com/BerriAI/litellm/pull/49001
feat(proxy)!: general_settings.claude_code_marketplace_auth_required puts the Claude Code marketplace behind a LiteLLM key by @zoveki in https://github.com/BerriAI/litellm/pull/48980
fix(token_counter)!: keep a loaded Hugging Face tokenizer when a slower download fails by @quietjules in https://github.com/BerriAI/litellm/pull/48916
fix(llms): move later developer messages for models declared system-first by @mika-lt in https://github.com/BerriAI/litellm/pull/48897
fix(openai): warn when the GPT-5 reasoning effort gate drops a level under drop_params by @mateo-berri in https://github.com/BerriAI/litellm/pull/47761
fix(websearch)!: keep the provider prefix on the agentic follow-up model by @msdianprince-7 in https://github.com/BerriAI/litellm/pull/38885
fix(logging): mark mirrored silent_model calls with is_silent_experiment in spend logs by @tavren27 in https://github.com/BerriAI/litellm/pull/48888
fix(cohere): price native Cohere embed calls on Foundry relays and the Cohere passthrough by @tavren27 in https://github.com/BerriAI/litellm/pull/48978
feat(providers): add Kyma API as an OpenAI-compatible provider by @m6atlas in https://github.com/BerriAI/litellm/pull/48756
fix(cost): bill a deployment's per-second rates on native Anthropic /v1/messages by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/42417
fix(proxy): stream a guardrail block on /v1/messages as Anthropic events by @tessellark in https://github.com/BerriAI/litellm/pull/49010
fix(caching): read the cache key's model group from the router's metadata for the call by @mika-lt in https://github.com/BerriAI/litellm/pull/48917
fix(passthrough): forward OpenAI-Beta and default agents=v1 on the OpenAI passthrough by @mika-lt in https://github.com/BerriAI/litellm/pull/49009
test(router_strategy): keep foreign asyncio task noise out of the budget limiter caplog asserts by @devin-ai-integration[bot] in https://github.com/BerriAI/litellm/pull/43375
fix(bedrock)!: route region-prefixed embedding models to that region and bill its price by @tessellark in https://github.com/BerriAI/litellm/pull/49072
fix(anthropic)!: drop beta-gated body fields under drop_params by @tavren27 in https://github.com/BerriAI/litellm/pull/48793
ci: run the dashboard API types and Terraform endpoint checks on every change that can alter the spec by @tavren27 in https://github.com/BerriAI/litellm/pull/49022
refactor(types): replace Any with proven types in 14 files by @mika-lt in https://github.com/BerriAI/litellm/pull/48928
fix(fallback): read stored router_settings from the database on fallback writes by @mika-lt in https://github.com/BerriAI/litellm/pull/48999
build(model-catalog): ship the cost map inside the native extension by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48907
ci(model-catalog): check the cost map by loading it through model-catalog by @yujonglee-berri in https://github.com/BerriAI/litellm/pull/48908
feat(soniox): add realtime speech-to-text on /v1/realtime by @dan2k3k4 in https://github.com/BerriAI/litellm/pull/44822
fix(exceptions)!: honor status code when mapping OpenAI invalid_request_error exceptions by @otiscuilei in https://github.com/BerriAI/litellm/pull/32537
fix(anthropic)!: send /v1/messages PDF documents to OpenAI-shaped providers as file parts by @mateo-berri in https://github.com/BerriAI/litellm/pull/47166

Full Changelog: v1.108.0...v1.109.0

v1.110.0-rc.1

2026-11-11T00:06:06+00:00

Verify Docker Image Signature

All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.

Verify using the pinned commit hash (recommended):

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.110.0-rc.1

Verify using the release tag (convenience):

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/v1.110.0-rc.1/cosign.pub \
  ghcr.io/berriai/litellm:v1.110.0-rc.1

Expected output:

The following checks were performed on each of these signatures:
  - The cosign claims were validated
  - The signatures were verified against the specified public key

What's Changed

fix(completion): accept a tool_choice that forces an OpenAI hosted tool by @kerry-berri in #50578
fix(accounting): stop delivery metrics from scanning retained receipts on every flush by @joshua-berri in #50565
fix(key/info): return 400 instead of 500 when no key is provided by @apex-mochen in #43577
fix(gemini): add the Flex above-200k cache read price to gemini-3.1-pro-preview rows by @devin-ai-integration[bot] in #50607
test(observability): match the full card number instead of a 4111 substring that timestamps can contain by @devin-ai-integration[bot] in #50612
fix(proxy): align ownership and scope handling by @mateo-berri in #50600
feat(proxy): total_timeout bounds a whole non-streaming request by @joshua-berri in #50496
fix(cli): end the local login when lite logout cannot check the OS keychain by @zoveki in #50098
fix(ui): label every budget reset time in the Admin UI with UTC by @zoveki in #50591
fix(proxy)!: price expanded wildcard /model/info rows as the model their calls route to by @paveln83 in #49128
fix(bedrock): derive the web identity session policy per AWS caller by @tavren27 in #50592
perf(router): skip unused deployment reads in the callback filter step and cache anthropic capability probes by @narmstrong-berri in #50641
test: update stale integration expectations and backport the Together e2e repin (#50601) by @yuneng-berri in #50756
test(integration): drop the empty-string n case from the Gemini long decode test by @yuneng-berri in #50794
test: update the spend tests that still expected the pre-ledger accounting queues by @yuneng-berri in #50805
fix(router): backport the mid-stream request fallback fix (#50260) to rc/1.110.0 by @yuneng-berri in #50798
test: repair four integration tests on rc/1.110.0 by @yuneng-berri in #50855
fix(redis): backport the per-request Redis batch scope leak fix (#50751) to rc/1.110.0 by @narmstrong-berri in #50949
fix(router): backport the registry lock fix (#50933) to rc/1.110.0 by @yuneng-berri in #50993

New Contributors

@apex-mochen made their first contribution in #43577

Full Changelog: v1.110.0-dev.2...v1.110.0-rc.1

Seedance 2.5 vs ByteDance Seedance 2.0: The Newer Model Has the Lower Ceiling

2026-11-11T00:00:00+00:00

Seedance 2.5 doubles clip length and reference budget, costs 50% more per second, and has no documented 4K API. Seedance 2.0 still owns 4K.

Is Anthropic's AI Now Speeding Up Its Own Development?

2026-11-11T00:00:00+00:00

A look at benchmark data and independent reporting on whether Claude models are accelerating Anthropic's own AI research and training pace.

Seedance 2.5 vs Meta Muse Video: One of These You Cannot Call

2026-11-11T00:00:00+00:00

Meta Muse Video was previewed on 7 July 2026 with no date, price or clip length since. Seedance 2.5 has been on sale and benchmarked since 31 July.

Arete-27B Benchmark Results: Propaganda Score vs Capability

2026-11-11T00:00:00+00:00

Arete-27B's fact score climbed from 2.32 to 2.98 and propaganda answers fell to 0.9%, with benchmarks showing no real capability loss.

Seedance 2.5 vs FLUX 3 Video: 22× Faster Is a Claim, Not a Measurement

2026-11-11T00:00:00+00:00

Black Forest Labs says FLUX 3 Max is 22x faster than Seedance 2.5 at the same quality. On image-to-video the board puts it 93 Elo behind.

MCP vs CLI vs Skills: What Each Costs in Context

2026-11-11T00:00:00+00:00

Every tool you give an agent sends its definition to the model on every request, and you pay for it as input tokens. A Model Context Protocol (MCP) server can add tens of thousands of tokens to the context window. The usual alternatives are a command-line tool (CLI) or a skill.

Arete-27B: Nous Research's Western-Aligned Fine-Tune of Qwen3.8-27B

2026-11-11T00:00:00+00:00

Nous Research's Arete-27B retrains Qwen3.8-27B to cut CCP-framed answers from 22% to 0.9% while holding capability benchmarks steady.

Gemini Decision 1 vs Kev: A Thousand Questions in One Call, or a Recipe You Can Retrain

2026-11-11T00:00:00+00:00

A hosted scorer answering 1,000 questions per request against a 262K-token input, versus an Apache-2.0 adapter its author calibrated himself.

LLM Quantization: What fp8 and fp4 Do to Quality and Price

2026-11-11T00:00:00+00:00

When you call an open-weight model through an API, several providers may serve it at different precisions, such as 16-bit bf16, 8-bit fp8 or 4-bit fp4. The precision can affect both the answers and the price. LLM quantization creates these versions by storing a model's weights in fewer bits, so the model needs less memory and hardware, at some cost in accuracy.

Binance's Internal AI Tool Became a Hacker's Backdoor. Here's How

2026-11-11T00:00:00+00:00

How attackers reportedly compromised a Binance engineer's device and used an internal AI tool to pivot toward wallet infrastructure in a major crypto theft.

Gemini Decision 1 vs Microsoft-Decision-1: Two Hyperscaler Scorers, Four Points Apart

2026-11-11T00:00:00+00:00

Two hyperscaler scorers four points apart, both mid-table on the component carrying 30% of the index, and where each earns its price.

LLM API Rate Limits: How to Fix OpenAI and Claude 429 Errors

2026-11-11T00:00:00+00:00

A 429 Too Many Requests error means you sent more requests or tokens than your limit allows in the current time window. OpenAI, Anthropic and OpenRouter all return it, and it usually shows up the first time an app runs under real load or a batch job fans out across several workers.

Claude Fable 5.2 vs Opus 5.5: KingBench Coding Results Explained

2026-11-11T00:00:00+00:00

Claude Fable 5.2 scored 70% on KingBench 4 coding tests, beating Opus 5.5's 57.8%. Here's what it built right, and where it broke.

Gemini Decision 1 vs Laya: 262,144 Tokens Rented Against 512 Tokens Owned

2026-11-11T00:00:00+00:00

A hosted multimodal scorer with a 262K window against an Apache-2.0 421M encoder you fine-tune yourself, and where each cannot compete.

Claude Mythos 5.2: Inside Anthropic's Restricted Cybersecurity Model

2026-11-11T00:00:00+00:00

Mythos 5.2 is Anthropic's cybersecurity-tuned Claude model, released alongside Opus 5.2 but limited to vetted organizations only.

Graphify is now an official Claude connector

2026-11-11T00:00:00+00:00

Graphify is now an official Claude connector. Connect your repositories and ask Claude who calls a function and what a change reaches.

Gemini Decision 1: Google's First Decision Model Leads Where It Counts Least

2026-11-11T00:00:00+00:00

Google's first decision model scores 66, ninth of eighteen. It leads both 15% components and sits mid-table in the 30% one that decides the composite.

Claude Opus 5.2 Benchmarks: What Anthropic's New Flagship Can Actually Do

2026-11-11T00:00:00+00:00

Claude Opus 5.2 tops the Valves index and Arc AGI 3, builds an OS from scratch, designs hardware, and formalizes Fermat's last theorem in 30 hours.

Share your team's skills with LiteLLM

2026-11-11T00:00:00+00:00

Share agent skills through one registry. Upload ZIPs, review submissions before publication, and see who downloads them with LiteLLM Skills Gateway.

Graphify with Claude: connector and Claude Code setup

2026-11-11T00:00:00+00:00

Use Graphify with Claude: connect from Anthropic's Connectors directory or add the MCP server to Claude Code, sign in with OAuth, and ask about your code.

[261111] Once Accused of Dragging Linux Down, Android Became a Major Kernel Contributor

2026-11-11T00:00:00+00:00

TLDR

History: Early Android's proprietary changes drew community criticism, but soaring maintenance costs eventually forced Google to push code upstream.
Evolution: Introducing GKI to unify base code turned Google into a core Linux kernel contributor.
Impact: This shift from isolated forks to open-source co-construction is now repeating in AI infrastructure.

AI News — 2026-11-11

2026-11-11T00:00:00+00:00

Daily AI development news summary for 2026-11-11

Gemini Decision 1 vs Drex Max v1.1: Both Read 262K Tokens, Only One Counts Documents

2026-11-11T00:00:00+00:00

Both read 262K tokens and both return probabilities. The gap is one component: Drex Max v1.1 passes 63% of documents, Gemini Decision 1 passes 33%.

Claude Opus 5.2 and the Physics Proofs: What Actually Happened

2026-11-11T00:00:00+00:00

Anthropic says Claude Opus 5.2 solved two open black hole physics problems and formalized Fermat's last theorem. Here's what that really means.

1.108.5 (LTS)

2026-11-10T23:09:26+00:00

Tip

The lts image tag always points to the latest LTS release. Pull ghcr.io/berriai/litellm:lts to stay on the newest LTS patch

Verify this release

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.108.5

Same command for litellm-database, litellm-non_root and the docker.io/litellm/* images

What's new in 1.108.5 (LTS)

5 PRs. 4 fixed. No action needed to upgrade

Changes

OpenAI: Fixed keyless workload identity deployments answering 401 on image edits, audio, files, batches and other routes (#50962)
Other providers: Fixed ollama_chat/ streams giving every parallel tool call index 0, so clients ran only the first call (#50718)
Messages API: Fixed /v1/messages answering 400 for a param listed in the deployment's additional_drop_params (#50755)
Cost tracking and pricing: Fixed incorrect transcription cost for some FLAC files, which could push a key over its budget (#50860)

Full changelog

graphify 0.9.106

2026-11-10T21:53:54+00:00

graphify 0.9.106

Fix: install requires-python is capped to >=3.10,<3.15, so uv tool install graphifyy (and pip) on a fresh machine no longer selects Python 3.15 — for which tree-sitter has no wheel — and then fails with a source build that needs a C compiler (No such file or directory: 'cc'). uv now picks a wheel-having interpreter (3.14) by default, with no --python flag and no compiler needed. The Ubuntu/Debian README line that asked for python3.12 (absent on Debian 12 and Ubuntu 22.04) is replaced with the uv installer or a plain python3/pipx line (#5134).
Fix: cpp a plain class/struct forward declaration folds into its unique exactly-qualified definition when one file reaches the other through unconditional exact quoted includes, and a parameterless member declaration folds into its proven out-of-line definition by matching return type and qualifiers; overloads, templates, conditional declarations and unproven includes stay separate (#5117, thanks @xRedLeader).
Fix: cpp a member call resolves through a scoped type alias (using Alias = ns::Type;) and through inherited fields/methods, and a forward-declared pointee type is completed at the use site; ambiguous, virtual or diamond-inherited receivers fail closed (#5118, thanks @xRedLeader).
Fix: cpp an out-of-line template member keeps its proven primary-template owner, and calls inside a #if-guarded initializer or argument list are recovered from byte-preserving branch views (bounded, fail-closed on malformed/over-budget input); dependent receivers stay unresolved (#5115, thanks @xRedLeader).
Fix: python a dotted subscript annotation (mod.MyBox[Foo]) counts its value once as the type instead of also as its own generic argument, so the false generic_arg reference no longer replaces the real parameter_type/return_type edge (#5113, thanks @rohit-jsfreaky).
Fix: ts/js a call through an ESM namespace import (import * as ns from './m'; ns.f()), a CommonJS whole-module require (const m = require('./m'); m.f()), or a TS import = require resolves to the real declaration; fail-closed on an unknown module, a non-exported name, or an ambiguous barrel (#5114, #5129, thanks @mohitdixit02).
Fix: llm a precomputed FileSlice input is preserved through extract_corpus_parallel instead of crashing or stamping a partial parent complete; a slice that covers only part of its file marks the parent partial for re-dispatch (#5106, #5128, thanks @aadilrizwan).
Fix: query BFS walks levels in a fixed, hash-independent order, so a query's edge lines (and which survive the token budget) no longer vary with PYTHONHASHSEED (#5109, #5112, thanks @rohit-jsfreaky).
Feat: query --no-doc-demotion (and the demote_docs rc key / MCP arg) disables the source-over-docs soft weight for a docs-first corpus; code corpora are unaffected by default (#5084, #5116, thanks @mohitdixit02).
Perf: resolution _collision_aware indexes salted sibling ids once instead of rescanning all nodes per use, cutting resolution time on large collision-heavy corpora (graph output byte-identical) (#5108, #5111, thanks @rohit-jsfreaky).
Fix: manifest a pom.xml using a prefixed, single-quoted or spaced xmlns keeps its Maven package and dependency nodes instead of losing them; a foreign namespace is not double-counted (#5126, thanks @sdivyanshu90).
Fix: merge-graphs graphify merge-graphs reports how many external nodes were unified across repos instead of silently absorbing them; the merge result is unchanged (#5127, thanks @igorpugatschev).
Fix: diagnostics diagnose_extraction no longer crashes on a malformed nodes container (null or scalar); the malformed extraction is still surfaced through the build error (#5125, thanks @sdivyanshu90).
Fix: html the interactive graph no longer draws a hyperedge region over members hidden by a community toggle; a region vanishes once fewer than two of its members stay visible (#5122, thanks @Yi-111-a).

The AST cache schema is bumped to 43; warm caches re-run the affected Python and C++ extractions.

v1.104.9

2026-11-10T20:29:03+00:00

Verify Docker Image Signature

All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.

Verify using the pinned commit hash (recommended):

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.104.9

Verify using the release tag (convenience):

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/v1.104.9/cosign.pub \
  ghcr.io/berriai/litellm:v1.104.9

Expected output:

The following checks were performed on each of these signatures:
  - The cosign claims were validated
  - The signatures were verified against the specified public key

What's Changed

fix(proxy-extras): backport the automatic SpendLogs index build (#43948, #44109, #44203) to stable/1.104.x by @mateo-berri in #50823
chore(release): bump litellm to 1.104.9 and litellm-proxy-extras to 0.4.102.post2 for stable/1.104.x by @yuneng-berri in #50985

Full Changelog: v1.104.8...v1.104.9

v1.103.11

2026-11-10T20:24:40+00:00

Verify Docker Image Signature

All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.

Verify using the pinned commit hash (recommended):

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.103.11

Verify using the release tag (convenience):

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/v1.103.11/cosign.pub \
  ghcr.io/berriai/litellm:v1.103.11

Expected output:

The following checks were performed on each of these signatures:
  - The cosign claims were validated
  - The signatures were verified against the specified public key

What's Changed

fix(proxy-extras): backport the automatic SpendLogs index build (#43948, #44109, #44203) to stable/1.103.x by @mateo-berri in #50856
chore(release): bump litellm to 1.103.11 for stable/1.103.x by @yuneng-berri in #50984

Full Changelog: v1.103.10...v1.103.11

Tencent Hy4 API Pricing and Support on LiteLLM

2026-11-10T18:00:00+00:00

Use Tencent Hy4 through LiteLLM with vision, tool calling and 1M context. Compare TokenHub list pricing with its 50% discount through December 31.

Gemini 4 Carbon API Pricing and Day 0 Support on LiteLLM

2026-11-10T10:00:00+00:00

Gemini 4 Carbon costs $2.00 per 1M input tokens and $10.00 per 1M output tokens until Nov 30, the same as Gemini 4 Argon. LiteLLM supports it on day 0 on AI Studio and Google Cloud.

Exploring File Uploads and Downloads in MCP

2026-11-10T09:00:00+00:00

The Model Context Protocol (MCP) started out as a simple way to connect models to tools, and that principle is still very much at its heart. Over the past two years, the things people build with the protocol have become more ambitious. Much more ambitious. Many of those MCP server projects now need to work with files - often driven by increasing agent capabilities and autonomy.

Saving a PDF from a conversation to cloud storage and pulling a CSV into a sandbox have become very common requests from developers adopting MCP. Today, implementing these scenarios would require a bunch of bespoke tool conventions and transfer mechanisms - not something we’d want folks doing over and over.

Because this request kept coming up in customer discussions, we bootstapped the File Uploads Working Group. We’ve spent the last month working through what a solution could look like.

Thanks to the effort of our community and developers, we now have a rough extension definition, and we’d love for you to poke at it as we start prototyping and trying out how it works in the wild.

Read the draft extension →

Where the work stands#

The work on the extension is still very much an experimental effort - while we have the shape outlined, we’re not sure yet that what we started with is what we’ll end up with. The best way to find out is, you guessed it - contact with reality.

You might also notice that there’s no reference implementation yet - that’s because we expect that a lot of the inner workings will, in fact, be quite different once we start integrating our experiment with third-party file services, such as cloud storage, database import/export, and many others that require file input and output.

As with all other extensions, this one will remain optional - there is no work required on either clients or servers if they don’t plan to support it (although we hope they really will, given the broad interest in this work).

We expect to finalize the extension and have it become stable before the end of the year.

How it works#

When we first jumped into discussions for what the extension may be, we had a few key scenarios that guided us:

You have a large file in a conversation, and you want to save it to cloud storage.
You want to pull a big CSV file into a code sandbox, so your agent can analyze the data or build some Python code around it.
You want to upload a large presentation, or an HTML deck with all of its assets in a ZIP file to a third-party remote server.

In each of these scenarios, the agent needs to know a bunch of things about the file, but the model doesn’t need to read every byte or even handle it directly. The file should go directly to its destination. This led to the design you will read about below.

The extension identifies itself as io.modelcontextprotocol/files, and it works with protocol revision 2026-07-28 and later. A server can advertise it in its server/discover response, with upload if it accepts files out of band and maxInlineBytes if it accepts them inline. A client declares it in the io.modelcontextprotocol/clientCapabilities metadata of every request that may need it, including tools/list, with upload if it can upload out of band and download if it accepts download targets.

Let’s walk through one simple upload. Let’s say a server has a save_to_folder tool, and a user wants to save a PDF with it.

Marking a file argument#

The server marks the argument with x-mcp-file in the tool’s input schema. This is the same pattern as x-mcp-header, and it tells the client that this argument is a file, so the client fills it in (and, as a result, the model doesn’t have to):

{
 "type": "object",
 "properties": {
 "folderId": { "type": "string" },
 "file": {
 "type": "object",
 "x-mcp-file": { "accept": ["application/pdf"], "maxSize": 157286400 },
 "properties": {
 "id": { "type": "string", "minLength": 1, "maxLength": 128 },
 "name": { "type": "string" },
 "mimeType": { "type": "string" },
 "size": { "type": "integer", "minimum": 0 },
 "data": { "type": "string" }
 },
 "required": ["id", "mimeType", "size"]
 }
 },
 "required": ["folderId", "file"]
}

A small file can go inline as base64 in data, as long as the request’s inline files total no more than the maxInlineBytes the server advertises. For anything bigger, if the server advertises upload, the client leaves data out and the upload happens in the steps below. maxSize limits each file however it travels. The full FileInput shape also has an optional SHA-256 digest, which a tool can require in its schema.

A client that doesn’t support the extension ignores x-mcp-file and treats the argument as an ordinary object. If the client doesn’t declare extension support in its tools/list request, the server should omit the tool or offer an alternative. For example, that alternative could accept a URL pointing to an already-uploaded file.

Asking before sending#

Before sending the file, the host gets the user’s approval. When no user is present, agent policy determines whether the transfer is allowed. The client then calls the tool with just the file’s metadata, not its contents.

{
 "jsonrpc": "2.0",
 "id": 1,
 "method": "tools/call",
 "params": {
 "name": "save_to_folder",
 "arguments": {
 "folderId": "fld_8c1a",
 "file": {
 "id": "file-1",
 "name": "Q3 report.pdf",
 "mimeType": "application/pdf",
 "size": 41204133
 }
 }
 }
}

Sending metadata first lets the server check permissions and storage quota before the upload starts. If either check fails, the server returns a normal tool error. That way, you don’t have to upload 40 MB of data only to discover that the destination folder is read-only.

If the checks pass, the server asks the client to upload the file before it runs the tool. It does this through the extension’s new upload_required result type. The response tells the client where and how to upload the file, and includes a requestState value to send back when retrying the tool call afterward:

{
 "jsonrpc": "2.0",
 "id": 1,
 "result": {
 "resultType": "upload_required",
 "uploads": [
 {
 "fileId": "file-1",
 "url": "https://upload.example.com/u/01JD3V7K2M?sig=...",
 "headers": { "Content-Type": "application/pdf" },
 "credentials": "omit",
 "expiresAt": "2026-11-04T18:30:00Z"
 }
 ],
 "requestState": "..."
 }
}

In this example, requestState represents the unfinished job: saving the PDF into a specific folder. The server packages the details it needs to continue into a value, protects it against tampering, and sends it with the upload instructions. After uploading the file, the client sends that value back unchanged, without needing to understand its contents.

The next tool call will bring its own context rather than rely on a session with the server. This is intentionally designed with the MCP stateless core in mind. A different server instance can continue the job too, as long as it can verify the value and access the uploaded file.

In a few scenarios, tools need protection against running twice - for example if another run would create an unwanted duplicate file. For those tools, the server marks requestState as used before running the job. All server instances share that record, so another attempt with the same value is rejected.

Completing the job#

The client uploads the PDF to the provided URL using an HTTPS PUT request. For storage services that need it, upload targets can instead specify POST, including a multipart/form-data body. The file’s contents will then travel directly to the upload target, outside JSON-RPC.

When the upload succeeds with an HTTP 2xx response, the client calls our save_to_folder tool again. It sends the same arguments and returns requestState unchanged, following the newly shipped Multi Round-Trip Requests pattern:

{
 "jsonrpc": "2.0",
 "id": 2,
 "method": "tools/call",
 "params": {
 "name": "save_to_folder",
 "arguments": {
 "folderId": "fld_8c1a",
 "file": {
 "id": "file-1",
 "name": "Q3 report.pdf",
 "mimeType": "application/pdf",
 "size": 41204133
 }
 },
 "requestState": "..."
 }
}

The server still needs to check the upload, though. It will verify requestState and confirm that the retry matches the original call. It will then check with the storage service that the file arrived and validate its size and type.

Only after those checks pass does it run save_to_folder and return the result. If the upload fails, the client doesn’t send the retry, and the tool doesn’t run.

The client coordinates all of this without involving the model again. The model made the tool call and got the result - neither the bytes nor the upload URL entered its context.

If you are curious about the end-to-end flow, it looks something like this:

sequenceDiagram
participant M as Model
participant H as Host and client
participant S as MCP server
participant T as Target
M->>H: Call save_to_folder
Note over H: User or operator policy approves sending the file
H->>S: tools/call with file metadata
alt Server rejects the file
S-->>H: Result with isError: true
else Server accepts the file
S-->>H: upload_required with uploads and requestState
H->>T: PUT or POST file bytes
T-->>H: 2xx
H->>S: tools/call with same arguments and requestState
Note over S: Confirm with storage that the bytes arrived
Note over S: Finalize if needed, then run the tool
S-->>H: Tool result
end
H-->>M: Result, with no target

Downloading a file#

Unsurprisingly, we’re also exploring the reverse flow: a tool can produce a file, and the client will need to download it. Like uploads, the file contents travel outside JSON-RPC.

The key difference, however, is that the tool can return download details with its completed result, so the client doesn’t need to retry the tool call after the transfer.

The draft defines a download response format using the result’s _meta. In the CSV example below, the tool returns the download URL and file details, and the client can fetch the file with an HTTPS GET request:

{
 "jsonrpc": "2.0",
 "id": 3,
 "result": {
 "resultType": "complete",
 "content": [{ "type": "text", "text": "Exported q3-orders.csv." }],
 "_meta": {
 "io.modelcontextprotocol/files": {
 "downloads": [
 {
 "name": "q3-orders.csv",
 "mimeType": "text/csv",
 "size": 209715200,
 "url": "https://download.example.com/d/01JD3V9P4Q?sig=...",
 "credentials": "omit",
 "expiresAt": "2026-11-04T18:30:00Z"
 }
 ]
 }
 }
 }
}

How download details should be represented in tool results is still an open design question. While we’re currently using _meta, we’d very much like feedback on how this approach works in practice.

The client declares download support in its tool call to tell the server it knows how to handle download targets. Only then can the server include them in its response. The result’s content still tells the model what was produced, such as “Exported q3-orders.csv.”, while the client handles the download target without exposing it to the model.

After downloading, the client checks the file’s byte count against size. If the server supplied a digest, the client also calculates the file’s hash and compares it with that value. A mismatch means the file must be discarded.

This is what the proposed file download flow looks like:

sequenceDiagram
participant M as Model
participant H as Host and client
participant S as MCP server
participant T as Target
M->>H: Call export_orders
H->>S: tools/call
S-->>H: Result with download target in _meta
opt Client wants the file and policy permits
H->>T: GET
T-->>H: 200 or 206 with bytes
end
H-->>M: Result, with no target

What about credentials#

A file transfer may go to a storage service rather than the MCP server itself. This means that the client must not automatically send its MCP credentials to every URL the server provides, because that could expose the token to a completely unrelated service.

Every upload and download target therefore includes a credentials field. It tells the client whether to attach the credential it uses for the MCP server, with two modes: omit and include.

With "credentials": "omit", used in the examples above, the client doesn’t attach its MCP token, cookies, or other credentials from its environment. The target provides its own authorization, such as the signature in a pre-signed URL. That lets the storage service accept the transfer without receiving the client’s MCP credential.

With "credentials": "include", the client attaches its MCP credential, but only to a path under its MCP address or a URL on the server’s own origin. This is useful when pre-signed URLs aren’t an option or file transfers need to pass through a gateway.

In the example below, the server provides a path instead of a separate URL when uploading a file:

{
 "fileId": "file-1",
 "path": "/uploads",
 "credentials": "include"
}

For a client using https://gateway.example.com/mcp/abc, that path resolves to https://gateway.example.com/mcp/abc/uploads. The upload goes through the same gateway as the MCP calls and uses the same MCP credential. This also helps sandboxed clients that can reach the gateway but not the storage service directly.

The client must also refuse redirects and reject targets that supply an Authorization header. This helps prevent scenarios where a target redirects the transfer to another origin or overrides the credential handling defined by credentials.

These safeguards control where the client’s MCP credential is sent, but they don’t by themselves control where the file goes. With omit, a malicious server could still direct an upload to a public endpoint it controls. The host gets approval before it knows the target, so that approval doesn’t cover the specific destination.

To limit the risk, our draft recommends that hosts restrict omit targets to an allowlist of permitted origins. Whether that restriction should be mandatory is another open question that we’d love your thoughts on.

The draft’s Security Implications section discusses these risks in more detail, along with our thinking on how to mitigate them.

We’d love your help#

We’d love your feedback on the draft. You can explore the extension repository, open an issue, or come chat with us in the working group’s Discord discussion. See our communications guide for details on joining the community.

Acknowledgments#

Thanks to Ethan Heimlich, Gaurav Soni, Jay Zuccarelli, Kevin Lam, Kryspin Ziemski, Kurtis Van Gent, Madhavi Pasumarthi, Mike Coopersmith, Mike Eastes, Patrick Rodgers, Prem Ramanathan, Thomas Sarlandie, and everyone who has brought use cases, questions, and feedback to the working group’s discussions.

New OtterlyAI Pricing: Same Price, More Prompts, Your Engines

2026-11-10T07:31:34+00:00

OtterlyAI’s new pricing: Lite, Standard and Premium stay at $29, $189 and $489. Get more prompts, choose your AI search engines, and track daily or weekly.

v3.12.43

2026-11-10T02:04:55+00:00

v3.12.43

How to Build a Voice AI Agent

2026-11-10T00:00:00+00:00

A voice agent needs to do more than generate a good answer. It must understand what someone said, decide what to do, call any needed tools, and start speaking again before the pause gets awkward.

How to Get Structured Output from Any LLM: JSON Schema, Tool Calls, or a Decision Model

2026-11-10T00:00:00+00:00

When your code reads fields from a model's reply, like an order ID or a refund amount, it expects clean JSON every time. Most of the time, it gets it. But now and then the model wraps the JSON in a code fence, adds a field you didn't ask for, or sends a number as text, and your code breaks. If you use several models, one fix won't cover them all, because each model and provider handles structured output in its own way.

Comprehension debt: what it is and how to keep it small

2026-11-10T00:00:00+00:00

Comprehension debt is the gap between the code a team ships and the code it understands. Habits that keep it small, and a way to see what a change touches.

Positron MCP: Connecting AI Agents to Your Data Science Workflow

2026-11-10T00:00:00+00:00

Positron's experimental MCP server connects agents to live data, plots, and consoles. Explore what changed and how to pair that access with a focused skill.

[261110] 8 Interesting Facts About Ubuntu

2026-11-10T00:00:00+00:00

TLDR

History: Ubuntu hit 22 years, shipped free CDs globally for 6 years, and was founded by a space-tourist "SABDFL".
Naming: Versions use "YY.MM" release dates, paired with alphabetical "Adjective + Animal" codenames.
Status: Despite a failed mobile crowdfunding attempt and low desktop share, it dominates cloud, AI/containers, and WSL.

AI News — 2026-11-10

2026-11-10T00:00:00+00:00

Daily AI development news summary for 2026-11-10

We gave agents better instructions

2026-11-10T00:00:00+00:00

We built a CLI command that gives coding agents current Mintlify guidance, and at first it changed nothing. One line in AGENTS.md, plus tighter guidance, took Claude Sonnet's pass rate on our hardest tasks from about one in five to every run.

How to Handle File Uploads in an MCP Server

2026-11-10T00:00:00+00:00

MCP has no standard way to send a file to a tool yet. Linear's MCP server uses a signed upload URL. Here is how that works and how to do the same.

We built one page to help AI agents set up Mintlify

2026-11-10T00:00:00+00:00

We watched agents struggle through Mintlify setup, fetching multiple pages, trying wrong commands, and sending long instructions. So we built setup.md, one starting point for getting connected.

The following open-source models are deprecated from DigitalOcean Serverless Inference as of 10 November 2026: DeepSeek V4 Flash Llama 4 Maverick 17B 128E Instruct Qwen 3.5 397B A17B Gemma 4

2026-11-10T00:00:00+00:00

The following open-source models are deprecated from DigitalOcean Serverless Inference as of 10 November 2026: DeepSeek V4 Flash Llama 4 Maverick 17B 128E Instruct Qwen 3.5 397B A17B Gemma 4 gpt-oss-20b These models remain available on dedicated inference. For serverless inference, migrate DeepSeek V4 Flash to DeepSeek V4.1 Flash 1017 (deepseek-v4.1-flash-1017), Llama 4 Maverick 17B 128E Instruct to DeepSeek V4.1 Flash (deepseek-v4.1-flash), Qwen 3.5 397B A17B to Qwen3.8-Max (qwen3.8-max), Gemma 4 to GLM-5.3 Flash (glm-5.3-flash), and gpt-oss-20b to gpt-oss-120b (openai-gpt-oss-120b) to avoid service disruption. For information on our model deprecation policy and recommended replacement models, see Model Support Policy.

1.108.4 (LTS)

2026-11-09T23:23:50+00:00

Tip

The lts image tag always points to the latest LTS release. Pull ghcr.io/berriai/litellm:lts to stay on the newest LTS patch

Verify this release

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.108.4

Same command for litellm-database, litellm-non_root and the docker.io/litellm/* images

What's new in 1.108.4 (LTS)

10 PRs. 7 fixed, 1 changed. No action needed to upgrade

Changes

AWS Bedrock: Fixed 400s from GLM, GPT-6 Luna and Nova 2 on Bedrock when the conversation carries Claude's signed thinking (#50399)
AWS Bedrock: Changed /bedrock pass-through routes to enforce the model's guardrailConfig over any guardrail the caller sends (#47429)
Vertex AI and Gemini: Fixed Vertex AI Claude batches failing with 404 at create (#50365)
Messages API: Fixed chatgpt/ models on /v1/messages dropping encrypted reasoning between turns (#50456)
Messages API: Fixed /v1/messages to OpenAI sending background: true upstream, which billed full runs that came back empty (#50462)
Embeddings, images and audio: Fixed /v1/videos polls failing with per-model keys or hitting the wrong provider account with a 404 (#44614)
Routing and fallbacks: Fixed team keys being served by a proxy-wide wildcard instead of the team's own BYOK wildcard deployment (#48490)
SDK: Fixed completion(), responses() and messages with tools failing with No module named 'mcp' on a base install, broken since v1.106.0 (#45961)

Backport #50568 · Backport #50605 · Backport #50693 · Full changelog

Claude Fable 5.2 API Pricing and Day 0 Support on LiteLLM

2026-11-09T18:00:00+00:00

Claude Fable 5.2 costs $8 per 1M input tokens and $40 per 1M output tokens, 20% below Fable 5.1. LiteLLM supports it on day 0.

graphify 0.9.105

2026-11-09T17:04:52+00:00

graphify 0.9.105

Fix: watch ignored directories (node_modules, .git, build output) are pruned from the Linux watch tree, so a large project no longer registers tens of thousands of wasted inotify watches and risks ENOSPC; the non-Linux recursive watcher is unchanged and still fires on every real change (#5074, #5091, thanks @Deepjyoti-Sarmah).
Fix: cpp a member-call chain routed through class fields declared in other headers (client.actions->Execute(), with value ., raw-pointer -> and single-arg unique_ptr/shared_ptr receivers) resolves to the field type's member across the include set, gated to a unique type and member and fail-closed otherwise; a typed local whose class is in another repo is parked for the cross-repo merge (#5098, thanks @xRedLeader).
Fix: markdown a code span naming an uppercase \-namespace class (\Foo\Bar, App\API) or a type-qualified method call (Type::member()) is kept as a symbol mention instead of being dropped, while a non-call Type::CONST/Type::class stays dropped (#5093, thanks @Adityakk9031).
Fix: ts/js a call through a destructured CommonJS barrel require (const { Thing } = require('./barrel') where the barrel re-exports Thing) resolves to the real declaration by following the barrel's exports, instead of dropping the edge (#5034, #5104, thanks @qq2514568380).
Fix: wiki the orphan sweep folds slugs with casefold (not lower), so two article names that collide only under full Unicode case-folding (Straße/STRASSE, Σ/ς) each keep their own file instead of one being deleted as an orphan; a file written this run is never swept (#5087, #5102, thanks @nothariharan).
Fix: build hyperedges are kept whole across an incremental graphify update: they are read from the nested graph slot a JSON-only writer uses, and after a deleted-source prune a hyperedge drops only its removed members (and itself only when none remain), instead of silently losing or leaving dangling hyperedges (#5081, #5103, thanks @nothariharan).
Fix: cache a degraded (under-producing) semantic cache entry no longer pins the corpus: after the shrink guard sees a re-extraction come back smaller than the stored graph, the entry is stamped partial so the next run re-dispatches it instead of replaying the degraded result forever (#5083, #5101, thanks @nothariharan).
Fix: affected graphify affected <file> seeds the file's nested member nodes (a method inside a class) transitively, not just one hop, so impact analysis no longer under-reports callers that bind to a nested member (#5094, thanks @Adityakk9031).
Fix: tree build_tree derives the common directory root from file parents, so a single-file or single-directory tree is no longer self-nested under the file itself (#5092, thanks @Adityakk9031).
Fix: query a label whose name is behind a schema or namespace qualifier (public.sweep_widget_state(), fmt.Println) is matched on its tail by the whole query and ranked above the per-token tiers, so it is no longer out-competed by single-token collision nodes and dropped from the seeds (#5080, #5095, thanks @sunhouy).
Fix: detect an incremental staleness check hashes a file whose stamp lands within the clock-skew window on either side of its mtime, so a same-tick rewrite that does not move mtime is still caught instead of being served stale (#5090, thanks @sunhouy).
Fix: path graphify path A B refuses an endpoint whose words appear nowhere in the graph (exit not-found) instead of returning a misleading route built from loosely-overlapping tokens (#5085, thanks @sunhouy).
Fix: csharp a classic extension-method call (x.Foo() where Foo is declared static ... Foo(this X x, ...)) binds to the extension method's declaration through an additive pass that runs after nominal dispatch, gated to a visible namespace, a matching this type and a unique applicable method; an instance method of the same name wins, a non-extension static or an unimported extension gets no edge, and ambiguity fails closed (#4828, #4849, thanks @Deepjyoti-Sarmah).
Fix: markdown a bare, unqualified free-function mention in prose (token, no () and no namespace) binds to a same-named function only when its defining file (or another symbol from it) is cited within the heading's section chain, instead of matching any same-named callable in the corpus; reference-style links ([text][label]) resolve at the use site (#4804, #4806, thanks @madanlalit).
Fix: export the to_json shrink guard no longer counts semantic sources that were replaced on re-extract or pruned as deletions as "lost", so a legitimate graphify update is not refused; a genuine unexplained shrink is still refused (#5107, thanks @khkar).
Fix: build a node whose file_type or confidence arrives as a structured value (list/dict) from a drifting LLM pass no longer crashes the validate/build membership checks: validation surfaces it as a schema error, and the build folds it to concept/INFERRED without fabricating a score (#5088, thanks @sdivyanshu90).
Fix: llm a splittable document below the character cap whose model output truncates is promoted to a full-range slice and bisected (and retried) instead of being kept as a truncated partial result; a non-splittable or code/binary unit still fails closed (#5072, #5078, thanks @aadilrizwan).
Feat: llm a custom OpenAI-compatible provider's extra_headers are forwarded to the client as default_headers, scoped to that provider and with any Authorization header dropped so it cannot override the configured credential (#3680, #4783, thanks @jaimemartinez).
Perf: extract whole-graph intermediates (the parse memo on the one-shot CLI, build/cluster/export working lists) are freed as soon as they are consumed, cutting peak memory on a large corpus; graph output is byte-identical (#5063, thanks @TPAteeq).
Fix: skillgen the claw/kiro/pi subagent prompt now tells the worker to write its chunk to CHUNK_PATH, so a disk-dispatch chunk is no longer returned inline and lost at the merge step (#5040, thanks @connhodendatabase).
Fix: deps the pypdf floor is raised to 6.21.4 in the pdf and all extras to pick up its security fixes (#5015, thanks @yotamleo).

The AST cache schema is bumped to 42; warm caches re-run the affected C++, Markdown and C# extractions.

GEO Experiment: LinkedIn AI Search Impact [Tested]

2026-11-09T12:56:12+00:00

Explore the LinkedIn GEO Experiment and discover how different publishing methods impact AI search citations. LinkedIn Pulse vs Posts

Kimi K3.1 Pro: oferta ograniczona — $1.80 za wejście / $9.00 za wyjście do 23 listopada

2026-11-09T11:50:04+00:00

Kimi K3.1 Pro jest dostępny na AIHubMix pod identyfikatorem modelu kimi-k3.1-pro. Na stronie modelu podano zarówno standardowe, jak i promocyjne stawki…

DeepSeek V4.1 Flash na Pareto: Q 48.40, puste odpowiedzi po stop, K 27.58

2026-11-09T02:36:41+00:00

W dniach 2026-11-05/06 model DeepSeek V4.1 Flash 1017 przeszedł pełny zestaw testów AIHubMix Provider Benchmark (APB) na Pareto. Wyniki to Q 48.40, F 93.42…

v3.12.32

2026-11-09T01:38:08+00:00

v3.12.32

What an AI code review should cost

2026-11-09T00:00:00+00:00

What should an AI code review cost? Public pricing, bills and noisy comments, and why a review should follow what a change reaches, not the whole repo.

AI News — 2026-11-09

2026-11-09T00:00:00+00:00

Daily AI development news summary for 2026-11-09

The following Anthropic model is now available on DigitalOcean Inference for serverless inference: Claude Fable 5.2 For more information, see the Available Models page.

2026-11-09T00:00:00+00:00

The following Anthropic model is now available on DigitalOcean Inference for serverless inference: Claude Fable 5.2 For more information, see the Available Models page.

v1.110.0-dev.2

2026-11-08T02:25:47+00:00

Verify Docker Image Signature

All LiteLLM Docker images are signed with cosign. This release is signed with the key introduced in commit e82a3c6.

Verify using the pinned commit hash (recommended):

A commit hash is cryptographically immutable, so this is the strongest way to ensure you are using the published signing key:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/e82a3c646908ba0a6ee0beddc4310d7a6cac4065/cosign.pub \
  ghcr.io/berriai/litellm:v1.110.0-dev.2

Verify using the release tag (convenience):

Tags are protected in this repository and resolve to the same key. This option is easier to read but relies on tag protection rules:

cosign verify \
  --key https://raw.githubusercontent.com/BerriAI/litellm/v1.110.0-dev.2/cosign.pub \
  ghcr.io/berriai/litellm:v1.110.0-dev.2

Expected output:

The following checks were performed on each of these signatures:
  - The cosign claims were validated
  - The signatures were verified against the specified public key

What's Changed

fix(router): record retries and fallback hops past a request's first logged failure by @narmstrong-berri in #50445
feat(spend-logs): turn off prompt storage per key or team by @krrish-berri-2 in #50135
fix(proxy)!: preserve Responses SSE event framing by @HootRock in #47980
feat(azure_ai): add cohere-rerank-v5-pro and -fast from the Azure retail price feed by @devin-ai-integration[bot] in #50499
fix(caching): don't cache or replay a reply that has no output by @PyExtreme in #50057
fix(gemini): add Gemini 2.5 Pro flex cache price above 200k tokens by @devin-ai-integration[bot] in #50480
fix(azure): align chat-latest retirement metadata with its documented alias by @devin-ai-integration[bot] in #50509
fix(vertex_ai): authenticate Vertex AI pass-through requests once by @quietjules in #48942
fix(redis): share one native circuit breaker between RedisCache and the Rust stores by @yujonglee-berri in #50473
fix(azure): add data zone rows for gpt-realtime-2.1, gpt-realtime-2.1-mini and gpt-realtime-1.5 by @devin-ai-integration[bot] in #50519
fix(anthropic): preserve streamed tool argument fragments verbatim by @mateo-berri in #50470
fix(vertex_ai): return 400 instead of 500 when Gemini image edit has no image by @MukeshK17 in #50244
refactor: declare each lazy litellm attribute once and resolve them through one function by @mateo-berri in #50523
fix(cost-map): sync OpenAI fine-tune and Codestral 2405 shutdown dates by @devin-ai-integration[bot] in #50529
fix(utils): drop Anthropic-only output_config under drop_params on providers that do not read it by @ryan-crabbe-berri in #50542
feat(cost-map): read alias_of pointer rows in the catalog and the loader by @kerry-berri in #50512
feat(cost-map): add moonshot/kimi-k3.1-pro-highspeed pricing (#50521) by @thinhnx128 in #50538
refactor: read lazy attributes with getattr instead of a TypedDict wrapper by @mateo-berri in #50543
fix(chatgpt): round-trip encrypted reasoning through /v1/messages by @Hugo-DDT in #49076
feat(cost-map): move -latest model aliases into the cost map by @kerry-berri in #50491
fix(cost-map): update Together Qwen3.7-Max pricing and add Rnj-1 Instruct by @devin-ai-integration[bot] in #50548
feat(guardrails): add prebuilt credential patterns for OpenAI, Anthropic, and Gemini by @yogeshwaran10 in #50268
fix(anthropic): reasoning_effort "none" disables thinking on Claude models that think by default by @ryan-crabbe-berri in #50541
test(e2e): add a Codex row where a sub-agent must read the task its parent hands it by @ryan-crabbe-berri in #50494
feat(ui): set deployment weights in key and team router settings by @joshua-berri in #50552
fix(mcp): derive MCP bridge and session keys with the FIPS provider's KBKDF by @yucheng-berri in #50032
fix(responses): surface mid-stream upstream errors that arrive without an event type by @ryan-crabbe-berri in #50466
test(integration): give each kill-burst its own prompts so cached control replies cannot stand in for late traffic by @mateo-berri in #50569
test(integration): budget proxy restart tests from calibrated boot time by @mateo-berri in #49761
fix(otel): name the routed model on auto-router LLM spans by @yucheng-berri in #50498
fix(proxy): hold DB_RECONCILE_LOCK around credential create, update and delete by @kerry-berri in #50551
fix(anthropic): refresh the WIF bearer token on the connection error retry by @kerry-berri in #50570
test(translation): expect redacted thinking blocks on Azure and Bedrock Mantle messages after #49076 by @devin-ai-integration[bot] in #50576
refactor(ui): clear hardcoded English from dashboard JSX by @yuneng-berri in #50575
fix(anthropic): exchange Google identity tokens again on refresh and read oidc/ references out of ANTHROPIC_IDENTITY_TOKEN by @kerry-berri in #50583
test(proxy): delete 28 dead skipped tests in test_key_generate_prisma.py by @yuneng-berri in #50550
fix(router): a cancelled or retried attempt releases its io-token reservation and least-busy count by @joshua-berri in #50486
feat(proxy): per-request header opts a stream into usage.cost without the global include_cost_in_streaming_usage flag by @kerry-berri in #50574

New Contributors

@HootRock made their first contribution in #47980
@PyExtreme made their first contribution in #50057
@MukeshK17 made their first contribution in #50244
@thinhnx128 made their first contribution in #50538
@Hugo-DDT made their first contribution in #49076

Full Changelog: v1.110.0-dev.1...v1.110.0-dev.2

graphify 0.9.104

2026-11-08T00:43:23+00:00

graphify 0.9.104

Security: skill the /graphify runbook no longer substitutes the scan path into Python code. Every generated code block (detect, AST and semantic extraction, build, diagnose, label, manifest, and --update) now reads the resolved scan root from the trusted graphify-out/.graphify_root sidecar that Step 1 writes through a quoted heredoc, instead of pasting the agent-supplied path into a string literal. A directory whose name carried a quote or $(...) could previously break out of Path('INPUT_PATH') and run as code; it is now read as inert data. Covers the assembled hosts and the Aider/Devin monoliths (#4762, thanks @danikdanik).
Fix: solidity a call on a receiver of a contract or interface type (IERC20 token; token.transfer(), IERC20(a).transfer(), token().balanceOf()) binds by name and arity through that type's linearization to the type's own declaration, instead of producing no edge; gated to a single visible contract/interface and fail-closed on an unknown or ambiguous type (#5060, thanks @rohit-jsfreaky).
Fix: solidity a call attached by a using A for T directive (using SafeCast for uint256; x.toUint48()) binds to the library or free function it attaches, matched by the receiver's type and an arity one greater than the call; fail-closed when the receiver type is unknown, the library is ambiguous, or two functions fit (#5062, thanks @rohit-jsfreaky).
Fix: solidity members, free functions and types in one scope whose names differ only by a leading _ or by case (_grantRole / grantRole) get separate node ids, so the _ symbol no longer vanishes into its twin and its calls land on the right node (#5027, thanks @rohit-jsfreaky).
Fix: csharp a constructor (instance, static or primary) is its own node owned by its type, with its body and : base(...)/: this(...) calls credited to it; new X(...), : base(...) and : this(...) additively link the arity-correct overload, and two fitting overloads or an unresolved base give no constructor edge (#5054, thanks @rohit-jsfreaky).
Fix: csharp calls inside a property or indexer accessor, an expression body, or an operator/conversion body are credited to that member instead of being lost; an indexer becomes a this[] node and an operator a .operator +() node (#5056, thanks @rohit-jsfreaky).
Fix: query a same-label node that is the only match for a query term is kept as a seed instead of being dropped by the per-label cap, so every distinct term keeps at least one seed while generic-label flooding stays capped (#5027, thanks @Vaibhav2824).
Fix: watch read-only events are kept out of the inotify mask via a change-only event_filter, so a large tree no longer drives watchdog into a self-sustaining queue-overflow loop; falls back with a warning on watchdog versions without event_filter support (#5033, thanks @Deepjyoti-Sarmah).
Fix: cache the \\?\ long-path spelling is kept out of cache identity, so the same file reached through \\?\C:\... and C:\... shares one entry; also fixes empty stat-index writes under UNC shares, a spurious "stat index switched" warning on Windows, and a junction/symlink alias wrongly sharing its target's digest (#5035, thanks @Bortlesboat).
Fix: cache stat-index keys stay root-relative under a Windows long-path root, so a moved or cloned corpus keeps the portable-key fastpath (#5035, thanks @Vaibhav2824).
Fix: tests the TypeScript unicode fixture is written as UTF-8 on every platform, so byte offsets match the extractor on Windows (cp1252) hosts (#5064, thanks @ayushcodes10).
Fix: tests the hook-freshness source mock is normalized through os_long_path, so the read-error/stale branch is actually exercised on Windows (#5065, thanks @ayushcodes10).
Fix: tests the watch pending-queue test compares Path objects so / and \ separators compare equal on Windows (#5039, thanks @Yi-111-a).
Fix: extract an ambiguous Python import's placeholder node id is hashed from checkout-independent data (module name, scan-root-relative source file, location) instead of the absolute path, so the same project at a different path or a moved cache gets the same ids (#5045, thanks @Fardin7798).
Fix: yaml a template using CloudFormation intrinsic tags (!Ref, !GetAtt, !Sub) no longer fails to parse: a scoped permissive loader rebuilds each !-tagged node's shape without evaluating it, while !!python/... tags still raise and never execute (#5077, thanks @ayushcodes10).
Fix: watch a --watch incremental update rebuilds the full corpus when graph.json is missing, instead of writing a partial graph from the manifest alone (#5020, thanks @azizur100389).
Fix: watch a change to a JS/TS resolution input (tsconfig.json, jsconfig.json, a lockfile) triggers a full rebuild so importers pick up the new resolution, instead of leaving stale imports_from edges (#5020, thanks @azizur100389).
Fix: ingest two concurrent graphify add captures for the same URL no longer clobber each other: the capture file is created exclusively (O_EXCL) and a collision retries a fresh name, with the payload fetched once (#5020, thanks @azizur100389).
Fix: reflect a malformed entry in the learning overlay is skipped instead of crashing the whole load, and a legacy null-source entry is preserved as a source-less annotation (#5020, thanks @azizur100389).
Fix: global a malformed repo entry in the global manifest is dropped on its own, keeping the other tracked repos, and only a non-object root or repos map is backed up and reset (#5020, thanks @azizur100389).
Fix: query a graph freshly built on a subfolder is no longer flagged as a pre-#1504 build (which nagged a no-op extract --force): the legacy-id probe strips the .graphify_root offset before deriving file stems, and a genuine legacy graph is still caught (#5019, thanks @neumannpeter1231-lang).
Fix: tree graphify export tree recognises the src node alias (and file/path/filepath) so a graph carrying those keys renders a populated tree instead of an empty one; a concept-only graph still reports empty (#5030, thanks @Adityakk9031).
Fix: cli graphify provider add accepts --pricing-input=VALUE/--pricing-output=VALUE equals-syntax and rejects a non-numeric, non-finite or negative price with a clean error instead of a traceback or a poisoned cost estimate (#5031, thanks @Adityakk9031).

The AST cache schema is bumped to 40; warm caches re-run the affected Solidity and C# extractions. The Python and YAML fixes above need no bump (placeholder ids are resolved after the per-file cache, and previously-failing YAML was never cached).

How to switch AI coding assistants without starting over

2026-11-08T00:00:00+00:00

How to switch AI coding assistants without starting over: what moves with you, what does not, and how to sign a new client in to your code knowledge graph.

AI News — 2026-11-08

2026-11-08T00:00:00+00:00

Daily AI development news summary for 2026-11-08

[261108] How Slimming Android Down to 512MB Saved the Cheapest Phones

2026-11-08T00:00:00+00:00

TLDR

History: Android 4.4 slimmed the OS down to a 340MB RAM minimum, accelerating budget smartphone adoption.
Evolution: Decoupling core features (like Google Play Services) and launching Android Go minimized system overhead and adaptation burdens.
Impact: Past extreme memory optimization tactics now provide vital models for overcoming on-device AI resource limits.

v3.12.28

2026-11-07T22:35:23+00:00

v3.12.28

v3.12.17

2026-11-07T01:06:01+00:00

v3.12.17

AI News — 2026-11-07

2026-11-07T00:00:00+00:00

Daily AI development news summary for 2026-11-07

[261107] Android Native Linux Support: What Advantages Does Termux Have Left?

2026-11-07T00:00:00+00:00

TLDR

Background: Android 17 natively supports Linux terminals and VMs (with Debian and GUI), offering a system-level full dev environment.
Division: Native Linux fits full-scale dev and AI coding tools; Termux relies on Termux:API for hardware access, favoring automation and scripts.
Evolution: Tightened permissions and native options shift Termux from a solo default to specialized use cases.

High Performance Block Storage is now available in private preview in the NYC2, SFO3, ATL1, RIC1, MKC1, and MEM1 datacenters. High Performance volumes use network-attached NVMe storage over NVMe/TCP

2026-11-07T00:00:00+00:00

High Performance Block Storage is now available in private preview in the NYC2, SFO3, ATL1, RIC1, MKC1, and MEM1 datacenters. High Performance volumes use network-attached NVMe storage over NVMe/TCP and deliver up to 60,000 IOPS and 1,200 MB/s per volume, regardless of Droplet type. You can create High Performance volumes in the Control Panel or through the API by setting performance_class to b2. Volumes require a minimum size of 128 GiB and are free during the preview. To request access, contact your Technical Account Manager (TAM). For more information, see How to Create High Performance Block Storage Volumes.

Editor publishing updates, Autopilot auto-merge, and more

2026-11-06T22:01:39+00:00

Editor

Choose changes to publish: Each of your changes in the Publish menu has a checkbox, so you can leave some changes out of a publish. The Granular publishing setting is now Show teammates' changes. See Publish specific changes.
Require review before publishing: Admins can turn on Require review before publishing in editor settings so that editors request a review instead of publishing to the deploy branch, and a publisher or admin approves and publishes their changes. You can't turn it on while the deploy branch is protected. See Require review before publishing. Available on Enterprise plans for sites on GitHub.

Dashboard

Support page: Open and track any support cases you have from the Support page. Cases are private by default. You can share a case with your organization. See Track your cases.
Expanded roles: Assign the publisher and commenter roles, compare roles in the permissions table on the Members page, and transfer organization ownership. See Expanded roles. Available on Enterprise plans.

Breaking change

Widget settings removed: You can no longer configure the assistant widget in the dashboard. Widgets you already embed continue working on your authorized sites. To set up the widget on a new site, email support@mintlify.com.

Autopilot

Auto-merge changes: Switch on Auto-merge changes to let Autopilot merge its own pull requests for the kinds of work your team reliably merges. See Auto-merge changes.
GitHub issue events: Ask Autopilot in a chat to run when someone opens, closes as completed, reopens, or labels an issue in a connected GitHub repository. See GitHub repositories.
Translations: Ask Autopilot in a chat to keep your translated pages in sync. It translates merged changes into your site's languages in a separate pull request and checks weekly for missing and outdated translations. See Chat with Autopilot.
Requested work is free: Work you ask Autopilot for directly, in chats on the Autopilot page, Slack, Teams, email replies, or pull request comments, no longer costs credits, even when it opens or updates a pull request. Automated runs, such as those started by a code merge or a schedule, cost 250 credits when they open a pull request, and 250 more each time a later automated run pushes to it. See Credits.
Credit usage: The usage page lists Autopilot as its own category, and each Autopilot usage event links to its session. See Credits.

CLI

mint feedback: Send a bug report, docs issue, or idea to the Mintlify team from your terminal, without signing in.
mint deploy --base: Create a base instead of a docs site. Add --push to push your local clone to the new base.
Default client for mint dev and mint export: mint dev and mint export use a new default client. Pass --classic to use the classic client. See mint export.

Breaking change

Exported archive layout: Archives from mint export store assets in _mintlify/assets/ and no longer include index.txt files. Update your server's caching rules, or export with --classic to keep the previous layout. See Host the export on your own server.
mint validate errors: mint validate fails when a page uses an undefined docs.json variable, an unknown component, or a snippet file that doesn't exist. Fix the findings, or turn off the undefined-variable, unknown-component, or missing-snippet rules with --disable. See mint validate.

AI and MCP

Ranked search: The admin MCP's search tool returns the most relevant pages first with matching snippets, and includes the full content of the top result when it's a confident match. See Content.
Safe page overwrites: Pass the version from read to write_page as baseVersion, and Mintlify refuses the write if the page changed since you read it. See Content.
Inline data: Admin MCP tools replace large base64 images and files with placeholders to keep responses small, and restore the original data when you write the page back. See Inline data.
Navigation location: The admin MCP's read tool returns where a page sits in your navigation, including its parent groups and child pages. See Content.
Invite members: Invite up to 250 people to your organization in one manage_members_sharing call. See Project management.
Review requests: On sites that require review, admin MCP tools request reviews, list and approve review requests, and publish approved changes. See Review requests.
Claude Code and Codex: The MCP page in your dashboard shows the commands to connect the admin MCP from Claude Code and Codex. See Connect to the admin MCP.

Breaking change

Assistant message v1 endpoint: The v1 assistant message endpoint stops working on November 9, 2026. Migrate integrations to the v2 endpoint.

Published sites

Match all changelog tags: Set filterTagMatch: "all" in a changelog page's frontmatter to show only updates that include every selected tag. See Tag filters.
Markdown indexing: Only Markdown served at a .md URL sends a noindex header, so search crawlers that receive Markdown at a page's regular URL can still index the page. See Search indexing.
GraphQL in Markdown: Markdown exports and llms-full.txt include the schema details of GraphQL reference pages, such as an operation's arguments and return type. See API reference pages.
Multi-repository sources: Mintlify merges redirects and applies .mintignore files from every repository source, not only the base source. See Redirects across sources. Available on Enterprise plans.
Azure Repos previews: Create a manual preview for an Azure Repos branch that doesn't have a pull request. See Preview deployments.

Improvements

Desktop app notifications: The desktop app shows a notification for new inbox activity. Click it to open the item in a new tab.
Desktop app windows: Drag a tab out of the tab bar to open it in a new window, or onto another window's tab bar to move it there. The page doesn't reload, so you keep unsaved edits. Right-click selected text in the editor to comment on it.
Comment on snippets: Select text inside a rendered snippet in the editor to comment on it.

Bug fixes

Editor: Wide Mermaid diagrams no longer get clipped. Edits inside HTML elements save, adjacent Steps components stay separate when a page opens, and the cursor lands in the right place after you delete a block. Images embedded in pasted HTML upload as files, copy buttons work in the live preview, and discarding a saved page no longer restores the edit.
Published sites: The version dropdown follows versions with a root-relative href. Clicking blank space around an enlarged image closes it. Screen readers announce heading anchor links by their heading text, and you can reach scrollable code blocks with the keyboard. In the API playground, the request body resets to the selected content type's example, and responses without content don't show an empty media type menu. Sphinx AutoAPI index pages stay out of the Getting Started group. Sites with authentication serve your custom root robots.txt. Switching versions from the version dropdown keeps your place on the page, and preview deployments generate an up-to-date llms-full.txt.
CLI: mint dev no longer gets stuck in a redirect loop when a page path matches a folder of static files. mint deploy --email works without --github, and mint claim explains why it couldn't create a claim link.
Desktop app: Invitations you accept from the desktop app's organization picker take effect.

November 6, 2026

2026-11-06T13:32:02+00:00

Client-Provided System Prompts are Stabilized

The Client-Provided System Prompt RFD has moved to Completed, and systemPrompt on session setup is now stabilized in ACP v1 and v2.

Agents that advertise systemPrompt accept system-level instructions in append mode. Agents that also advertise mode.override let Clients replace their built-in default instructions.

Read the full announcement.

Przełączanie dostawcy w Claude Code: poradnik CC Switch w 5 minut

2026-11-06T09:19:00+00:00

Ten poradnik dotyczy CC Switch v4.0.26 (listopad 2026). Etykiety przycisków mogą się zmienić w kolejnych wersjach; model konfiguracji pozostaje bez zmian…

Tool-Aware Model Routing for Coding Agents with NVIDIA NeMo Switchyard

2026-11-06T09:00:00+00:00

LiteLLM now steers each coding agent session by its tool execution history, keeping routine work on an efficient model and escalating to a capable model when the trajectory turns hard.

LiteLLM now supports -latest model aliases

2026-11-06T01:00:00+00:00

Keep the same model name as newer versions arrive. Use -latest aliases in the LiteLLM SDK and AI Gateway.

Fundamentals Matter

2026-11-06T00:00:00+00:00

Cognitive surrender, comprehension debt, and how startups keep up with understanding while agents write more.

AI News — 2026-11-06

2026-11-06T00:00:00+00:00

Daily AI development news summary for 2026-11-06

Graphify passes 10 million downloads on PyPI

2026-11-06T00:00:00+00:00

Open-source Graphify, graphifyy on PyPI, just passed 10 million downloads in seven months. Here's what the numbers show and what's next.

[261106] Don't Rush to Throw Away Your Old PC: 5 Lightweight Linux Distributions Worth Trying

2026-11-06T00:00:00+00:00

TLDR

Core Positioning: 5 lightweight Linux solutions to extend older PC lifespans after Windows 10 support.
Distributions: Linux Mint Xfce (beginner), Lubuntu (Ubuntu ecosystem), MX Linux (maintenance tools), antiX (32-bit/low RAM), and Puppy Linux (portable USB).
Selection: Mint Xfce for daily use; Lubuntu for 2–4GB RAM; MX Linux for tools; antiX for legacy hardware; Puppy Linux for troubleshooting.

How does Claude-managed auth work for MCP connectors?

2026-11-06T00:00:00+00:00

Claude can exchange a signed identity assertion for an MCP access token, letting eligible team members use a connector without a separate OAuth consent screen. Here is what server builders need to implement.

Give Hermes Agent a Long-Term Memory on Upstash Redis

2026-11-06T00:00:00+00:00

upstash-hermes-memory gives Hermes Agent long-term memory and searchable chat history on Upstash Redis. It recalls related memories on every turn and runs on the Free plan.

graphify 0.9.103

2026-11-05T16:35:44+00:00

graphify 0.9.103

A broad correctness, performance and robustness release across the extractors, resolution, hooks, the query and update paths, install and the skill runbook.

Extractors

C#: preprocessor directives and #if branches that derail tree-sitter are blanked before parsing (byte-length preserving), so a directive no longer closes a class early and drops later members or calls.
Solidity: library calls (Lib.f(), including aliased imports) and emit/revert/require naming a declared event or error now link to the member they name, gated to a unique visible type and member.
YAML: a << merge sequence gives the earlier anchor precedence for a duplicated key, matching the spec and PyYAML.
Markdown: a code span written as a backslash namespace (\Foo\Bar) or a type-qualified member (Type::member) resolves to the real node; a dotted mention whose method name is a language builtin is no longer dropped.
Python: a bare builtin-named call (raise NotImplementedError(...)) no longer binds to a same-named class in another module unless imported.

Resolution

A call inside a top-level def _main() is attributed to _main, not the public main it collides with after id normalization, without misattributing a class member's type reference.

Hooks

A file unchanged in content is not re-flagged stale by mtime (the extraction hash is checked against the manifest).
A dedicated Grep is held to the strict orientation gate; a Bash file read (cat/head/sed -n) of an in-project code file gets the read nudge; a search whose every target is outside the project emits no project hint; the nudge budget is restored after a Claude Code compaction; and the toplevel anchor keeps working when CDPATH is exported.

Performance

Lower peak memory in load, resolution and extraction; a faster #3252 cross-file repoint; and the watch reconcile pass resolves AST ownership once per file instead of re-reading each header per node — all with identical graph output.

Update / build / cache

graphify update re-extracts only the files whose content changed (identical to a full rebuild); a contended incremental request is recorded instead of dropped. The build write guard no longer counts intentional ghost/dedup folds as drops; cache keys keep their written case.

Query / tree

Source nodes rank above tests and docs (a soft weight) with file-first answers; query reads query_budget/query_depth from .graphifyrc and gains --depth. export tree caps directory children while reporting full totals.

Install / thread-safety

The Pi project skill is written to .pi/skills; uninstall removes the global skill registration it wrote. The ambient scan root is held in a ContextVar so concurrent library callers extracting different roots do not clobber each other.

The AST cache schema is bumped to 39; warm caches re-run the affected extractions.

Thanks to @rohit-jsfreaky, @Deepjyoti-Sarmah, @xiongjianxu, @nothariharan, @andrebrait, @ajbmachon, @Shubham-Padkonde, @madhusudanpatnaik, @pei711, @Mpasha17, @ayushcodes10, @gitmotion, @briascoi, @oleksii-tumanov, @rohit-jsfreaky, @viwe-monai, @WyattAu, @Vivek1898, @wanjinhao1, @PratikDhanave, @HhH666-ux, @sunhouy, @mathewgeejo, @jinsoo and @AndresDiagoM.

Bing Webmaster Tools Topics Portfolio: How to Use It with OtterlyAI

2026-11-05T16:28:36+00:00

Bing Webmaster Tools Topics Portfolio: How to Use It with OtterlyAI

How to Make Your Business Appear in AI Search Engines

2026-11-05T15:34:56+00:00

A four-step guide to getting your business cited by ChatGPT, Copilot, Perplexity and Google AI, backed by 20+ OtterlyAI experiments. What works, what to skip

DeepSeek V4.1 Flash w Fireworks: puste odpowiedzi po stop, trafienia w pamięć podręczną 3/9

2026-11-05T14:59:59+00:00

W dniach 2026-11-03/04 DeepSeek V4.1 Flash 1017 przetestowano w pełnym zestawie APB u pięciu dostawców, korzystając z AIHubMix i tych samych pytań…

DeepSeek V4.1 Flash w oficjalnym kanale: 94.00 F, 15 nieprawidłowych prób

2026-11-05T14:59:41+00:00

W dniach 2026-11-03/04 zestaw APB Full uruchomiono przez AIHubMix dla modelu DeepSeek V4.1 Flash 1017 u pięciu dostawców. U każdego dostawcy użyto tych…

DeepSeek V4.1 Flash w Alibaba Cloud: pierwszy token po 1.34s, cache 99.37

2026-11-05T14:59:23+00:00

W dniach 2026-11-03/04 przeprowadzono pełny zestaw APB dla DeepSeek V4.1 Flash 1017 u pięciu dostawców za pośrednictwem AIHubMix. U każdego dostawcy użyto…

Model Deprecation Management

2026-11-05T01:00:00+00:00

Manage model deprecations in the LiteLLM AI Gateway. Find affected keys, track migration traffic, and notify owners before provider deadlines.

AI News — 2026-11-05

2026-11-05T00:00:00+00:00

Daily AI development news summary for 2026-11-05

You can now remove a Droplet snapshot’s copy from a single region without deleting the snapshot or its copies in other regions, using the DigitalOcean API or the Control Panel. Read How to Add or

2026-11-05T00:00:00+00:00

You can now remove a Droplet snapshot’s copy from a single region without deleting the snapshot or its copies in other regions, using the DigitalOcean API or the Control Panel. Read How to Add or Remove Droplet Snapshots in Other Regions to learn how to remove a regional copy.

[261105] How an Android "Surgery" Saved Your Phone's Battery

2026-11-05T00:00:00+00:00

TLDR

Evolution: Android shifted background management from manual user control to default system behavior starting with Doze mode in version 6.0.
Optimization: Subsequent updates tightened permissions and added Adaptive Battery to dynamically restrict apps based on user habits.
New Challenge: The rise of AI assistants requiring continuous context and automation is forcing another redesign of foreground/background rules.

The DigitalOcean Documentation MCP server searches the full documentation site index, and each search result includes an excerpt from the page body. You can restrict docs-search results to one

2026-11-05T00:00:00+00:00

The DigitalOcean Documentation MCP server searches the full documentation site index, and each search result includes an excerpt from the page body. You can restrict docs-search results to one product with the Product parameter, and read long pages in sections with the StartIndex and MaxLength parameters of docs-get-page. The server does not require authentication or a DigitalOcean account. To add it to Claude, Claude Desktop, Claude Code, Cursor, Visual Studio Code, or Codex, see Connect the Documentation MCP Server.

Cuttlefish, an iOS app for DigitalOcean Harness Runtime, is available in public preview on the App Store. Follow and steer sessions, approve or deny tool calls, and work with workspace files and

2026-11-05T00:00:00+00:00

Cuttlefish, an iOS app for DigitalOcean Harness Runtime, is available in public preview on the App Store. Follow and steer sessions, approve or deny tool calls, and work with workspace files and checkpoints from your iPhone. To get started, see How to Steer and Approve Harness Runtime Sessions from the Cuttlefish iOS App.

The following NVIDIA models are deprecated from DigitalOcean Serverless Inference as of 5 November 2026: Nemotron 3 Diarization Nemotron 3 Ultra Nemotron 3 Nano Omni Nemotron Nano 12B v2 VL These

2026-11-05T00:00:00+00:00

The following NVIDIA models are deprecated from DigitalOcean Serverless Inference as of 5 November 2026: Nemotron 3 Diarization Nemotron 3 Ultra Nemotron 3 Nano Omni Nemotron Nano 12B v2 VL These models remain available on Dedicated Inference. Migrate serverless workloads to NVIDIA Nemotron 3.5 Super VL (nemotron-3.5-super-vl) to avoid service disruption. For information on our model deprecation policy and recommended replacement models, see Model Support Policy.

How NOLA Marketing Sells GEO: Prove It on Your Own Site First

2026-11-04T14:41:24+00:00

Explore how NOLA Marketing sells GEO services, from qualifying prospects to structuring its AEO Program and using real AI visibility results.

The following Moonshot AI model is now available on DigitalOcean Inference for serverless inference and dedicated inference: Kimi K3.1 Pro For more information, see the Available Models page.

2026-11-04T00:00:00+00:00

The following Moonshot AI model is now available on DigitalOcean Inference for serverless inference and dedicated inference: Kimi K3.1 Pro For more information, see the Available Models page.

SSO, Two-Factor Sign-In and Passkeys in Manufact Cloud

2026-11-04T00:00:00+00:00

Every Manufact Cloud account can now require a second step at sign-in, with an authenticator app, an emailed code, backup codes or a passkey. Enterprise teams can sign in through their own OIDC or SAML identity provider and make it the only way in for their domain.

graphify 0.9.102

2026-11-03T22:47:18+00:00

graphify 0.9.102

A large correctness, performance and workflow release spanning the extractors, the query layer, incremental update, the hooks, the skill runbook and install.

Call-resolution (extractors)

go: same-file member calls resolve through the receiver's type; cross-package pkg.T receivers bind only via the file's import; interface method calls dispatch to implicit implementers (#4966, #4853, #4934).
rust: module-qualified calls (crate::b::helper()) resolve cross-file; calls inside expression-macro arguments are linked (#4944, #4945).
kotlin: overloaded and init {}-block calls resolve; unreachable external-inherited calls are dropped (#4956).
python: a module-level singleton's member calls resolve through its type; a bare method call no longer self-binds (#4972, #4933).
js/ts: class extends through a CommonJS require resolves; a typed tagged template and an import("mod").T<X> annotation no longer half-parse or mint a false dynamic_import edge; default-exported object methods resolve (#4700, #4976, #4967).
solidity: ../ imports resolve to the real file (#4985).
cpp: receiver typing from params/fields, split-header owner binding, local-callable suppression, type-alias exact-spelling (#4946, #4920, #4921, #4951).

Query

Source nodes rank above tests and docs (a soft weight, not exclusion), and answers are file-first (#4982).
graphify query reads query_budget/query_depth from .graphifyrc and gains a --depth flag (#4760).

Incremental update / watch

graphify update re-extracts only the files whose content changed, with a graph identical to a full rebuild; a contended incremental request is recorded instead of dropped (#4800, #4801, #4802).

Build / cache

The write guard no longer counts intentional ghost folds or dedup folds as dropped nodes; independently-sourced semantic evidence survives a fold; cache keys are case-preserving (#4957, #4952, #4927, #4950, #4937).

Hooks

A file unchanged in content is not re-flagged stale by mtime; a dedicated Grep is held to the strict orientation gate; external searches emit no project hint (#4974, #4973, #4979).

Performance

Lower peak memory in load, resolution and extraction, and a faster cross-file repoint — all with identical output (#4960, #4958, #4959, #4981).

Skill / install

The --update runbook re-queues partly-read docs instead of stamping them done; Step 4 threads --force; a hub-bypass report is added; uninstall removes the global skill registration (#4816, #4805, #4704, #4975).

The AST cache schema is bumped to 37; warm caches re-run the affected extractions.

Thanks to @nothariharan, @Deepjyoti-Sarmah, @xiongjianxu, @andrebrait, @ajbmachon, @Shubham-Padkonde, @rohit-jsfreaky, @oleksii-tumanov, @madhusudanpatnaik, @pei711, @Mpasha17, @ayushcodes10, @gitmotion, @briascoi, @madhusudanpatnaik, @AndresDiagoM, @azizur100389, @Ethan-Ethan-Ethan, @jinsoo and @rohit-jsfreaky.

[261103] Why Are Software Vulnerabilities Growing So Fast?

2026-11-03T00:00:00+00:00

TLDR

Growth Trend: Public vulnerabilities are experiencing explosive growth, with 2026 projections nearing 75,000 entries.
Core Reasons: Driven by increasing software dependency complexity (especially AI toolchains), the proliferation of automated and AI scanning tools, and broader coverage of vulnerability disclosure ecosystems.
Mitigation Strategy: CVE counts do not equal actual risk; developers do not need to track every vulnerability, but should focus on rapid triage of core risks within their own technology stacks.

Migrate to Mintlify with your agent

2026-11-03T00:00:00+00:00

We turned what we learned from hundreds of migrations into mint migrate, so your agent can move your docs from 21 platforms, prove nothing went missing, and tell you what didn't carry over.

Tuning Our MCP Server with Agent Benchmarks

2026-11-03T00:00:00+00:00

We ran coding agents through real Upstash Redis tasks with the Upstash MCP server, fixed what the benchmark found, and compared the results before and after.

November 3, 2026

2026-11-03T00:00:00+00:00

Summary

Inboxes can now forward a copy of every email they receive to up to 5 verified addresses. Keep a person in the loop on what your agents receive, or copy mail into an archive or another system, without writing a webhook handler.

What's new?

New endpoints:

POST /v0/forwarding/verifications emails a confirmation code to an address, and POST /v0/forwarding/verifications/confirm verifies it with that code. A verified address can be used anywhere in the organization.
GET, PUT and DELETE on /v0/forwarding, /v0/pods/{pod_id}/forwarding and /v0/inboxes/{inbox_id}/forwarding read, replace and remove the addresses an organization, a pod or an inbox forwards to.

New features:

Nearest level wins: an inbox's list replaces its pod's, and a pod's replaces the organization's. An empty list turns forwarding off at that level.
Verified addresses only: an address receives nothing until whoever reads it confirms the emailed code. unverified on the read lists the addresses still waiting.
Never forwarded: sent mail, spam, bounces, auto-replies, and mail that was already forwarded, so forwards can't loop.
US region only for now: forwarding runs on api.agentmail.to.

Use cases

Build support agents whose incoming mail also reaches a team lead's mailbox.
Copy every email a pod of agents receives into a compliance archive.
Hand incoming mail to a system that already reads email, while the agent keeps working in its inbox.

graphify 0.9.101

2026-11-02T19:39:26+00:00

graphify 0.9.101

Resolution-accuracy and cache-correctness release across C++, Rust, Python, Go, the build write guard, and incremental rebuilds.

C++

A member call x.m() binds by the receiver's type, now typed from function parameters and class fields and not only locals (void render(Base* b){ b->draw(); } resolves to Base::draw); an untyped or ambiguous receiver keeps its in-file binding (#3215, #4946).
An out-of-line or split-header definition attaches to the unique owner whose class file is in its include set, instead of cross-binding to an unrelated same-named function; ambiguous owners fail closed (#4916, #4920).
A local callable value (a lambda or std::function) is kept out of global call resolution by byte-range binding lifetimes, so it no longer mints a spurious INFERRED call to a same-named global; a later legitimate global call survives (#4917, #4921).
A type-alias reference (using Foo = Bar; then Foo x;) binds by exact spelling and a salted alias keeps its own contains edge (#4951).

Rust

A module-qualified call (crate::b::helper(), super::m::f()) resolves through qualifier-verified cross-file resolution, staying inside the caller's crate (#2135, #4944).
A call written inside an expression macro's arguments (println!("{}", f()), vec![g()]) is linked, gated to an allowlist so macro_rules! bodies and custom DSLs are left alone (#4945).

Python / Go

A bare call inside a method no longer self-binds to a same-named method; it resolves to the imported or module-level function it names (#4933).
Imports resolve across PEP 420 namespace packages split over multiple src/ roots (#4929).
A Go method value reached through a receiver field is linked as an indirect_call (#1566, #4931).
A Go package-level var initialiser references the types it names (#2360, #4932).

Build / cache / rebuild / watch

The write guard no longer refuses a healthy graph when entity dedup folds a saved node into a surviving twin; a genuine loss is still refused (#4940, #4952).
Independently sourced semantic file evidence survives a fold into an AST file-self node; ordinary INFERRED mention ghosts still fold so graphs do not accumulate duplicates (#4896, #4927).
The file-stem alias index only rebinds an edge when the edge's file and the alias target share a language family (#4949).
Cache keys keep their written case so README.md and readme.md no longer collide, and hashed/word-count files are opened through os_long_path (#4925, #4937).
A --code-only graph stays code-only across hook, update and watch rebuilds (#4923, #4939).
A rebuild no longer waits forever when msvcrt.locking is unsupported (#4924, #4926).
A full reconcile evicts source records for an excluded linked git worktree (#4919).
AST-only updates preserve cross-file semantic evidence whose node id collides with a code definition (#4950).
A sliced file is cached only once every slice has succeeded and is assembled in chunk order (#4943, #4947).

The AST cache schema is bumped to 36; warm caches re-run the affected extractions.

Thanks to @nothariharan, @Deepjyoti-Sarmah, @azizur100389, @andrebrait, @ayushcodes10, @madhusudanpatnaik, @Ethan-Ethan-Ethan and @AndresDiagoM.

How to Get AI Agents to Send Feedback on Your MCP Server

2026-11-02T00:00:00+00:00

An agent can recover from a tool failure and still have something worth reporting. Collect that feedback in the Manufact dashboard with send_feedback.

AI Visibility Metrics: The GEO KPIs That Predict Revenue

2026-11-01T10:36:08+00:00

The AI visibility metrics and GEO KPIs we use at OtterlyAI, which one to make primary, and why GA4 shows 0.1% when customers say 10.6%.

graphify 0.9.100

2026-11-01T00:53:52+00:00

0.9.100 (2026-11-01)

Fix: ids a type and a function whose names differ only in case (Foo the class, foo the function) get distinct node ids instead of colliding (which dropped one node and minted a bogus self-loop) (#4904, #4908, thanks @bercedev).
Fix: php a Scope::MEMBER access in a class-level initializer (a const value, a property default, or a parameter/promoted-constructor default) emits a references_constant edge; self::/static::/parent:: are skipped and unknown scopes mint no stub (#4648, #4895, thanks @Jitendarmahara).
Fix: hooks git rev-parse output is decoded as UTF-8, so a linked worktree under a non-ASCII path resolves correctly on Windows instead of mojibake (#4903, thanks @LukeTheoJohnson).
Fix: analyze on a code-only graph the isolated-node question reason says "possible missing edges" rather than referencing documentation gaps that cannot apply (#4887, thanks @oleksii-tumanov).
Perf: resolve reference edges are indexed by source once, so JS/TS import resolution no longer re-walks every edge per import (output unchanged) (#4909, #4913, thanks @Deepjyoti-Sarmah).
Docs: skill the skill now says when a graph is worth building, across all generated hosts, while preserving the "if the user invoked /graphify, do it as asked" invariant (#4829, #4906, thanks @TanmayShah29).
Fix: deps the video extra caps PyAV below 19 (av>=11,<19), because PyAV 19 removed the metadata_errors argument that faster-whisper passes unconditionally when opening audio; without the cap a fresh install on Python 3.12+ failed to transcribe (#4842, thanks @SyedFahad7).
Fix: r setClass/setRefClass S4/R5 inheritance declared with contains= now emits an inherits edge (previously only R6Class's inherit= was recognised) (#4902, thanks @ayushcodes10).
Fix: elixir a bare function call resolves to a function in the caller's own module before other modules, instead of binding to the last-walked same-named definition (#4900, thanks @ayushcodes10).
Fix: dart a mixin on A, B clause resolves every superclass constraint, not just the first (#4899, thanks @ayushcodes10).
Fix: fortran a derived type's extends(base) emits an inherits edge, bringing Fortran in line with the other OOP extractors (#4897, thanks @ayushcodes10).
Fix: scala an extension (x: T) def m is scoped to its receiver type T (a method edge to T) instead of leaking to file scope (#2044, #4856, thanks @divakantpandey).
Fix: terraform a node carries its original scope so ids stay stable when a cached project directory is moved, instead of re-deriving the scope from the new path (#4858, #4861, thanks @yUm1k4).
Fix: extract (JVM) Java and Kotlin imports, extends/implements and member calls resolve across files and across the two languages when a written FQN names exactly one sourced type (#4874, #4879, thanks @Deepjyoti-Sarmah).
Fix: export a path/explain/wiki rendering lists every parallel relation between two nodes (via edge_datas), and a raw --no-cluster write marks the graph multigraph so parallel edges survive reload (#4794, thanks @ensell-yes).
Fix: export a partial or failed backup is reported as no backup and retried on the next run, instead of being recorded as a false success (#3769, #4834, thanks @barlowa124).
Fix: affected graphify affected (and query --seed) resolve the same path::Symbol and dotted reference forms explain accepts, reusing explain's resolver, and exit non-zero (3 ambiguous / 4 not-found) on a seed that does not resolve (#4792, #4797, thanks @dupeisen).
Fix: install graphify uninstall removes every platform's user-scope skill (13 were left behind) and the kilo global command; and graphify kilo uninstall removes the kilo skill's references/ sidecar (#4782, #4789, thanks @bercedev).
Feat: llm the Claude backend caches its (stable, large) system prompt with cache_control, counting the cached tokens back into input_tokens, and GRAPHIFY_CLAUDE_PROMPT_CACHE=0 disables it for gateways that reject cache_control (#4835, #4855, #4880, thanks @Jvxoq and @1akashkalita).
Fix: prompt the node-id format description in the extraction prompt matches what ids.py actually produces (word characters kept, non-ASCII letters/digits preserved, ASCII lowercased) instead of the stale "[a-z0-9_] only" wording (#4432, #4799, thanks @dupeisen).
Docs: how-it-works the supported-language count is updated (about 40: 33 tree-sitter grammars plus 7 built-in parsers) and Pass 3 is described as running the configured LLM backend, not Claude only (#4873, thanks @SyedFahad7).

[261101] A Single Link to Generate Various Charts

2026-11-01T00:00:00+00:00

TLDR

Core Positioning: x-cmd Draw is a lightweight tool that can quickly generate various charts through a single URL.
Quick Generation: It supports various charts such as pie charts, bar charts, line charts, and Sankey diagrams, and can output SVG or PNG formats for web pages and documents.
AI-Friendly: The SVG embeds machine-readable metadata, allowing AI Agents to directly read, modify data, and regenerate charts.
Flexible Deployment: It supports online usage, CLI, npm libraries, Deno, as well as local or Cloudflare self-hosting.

Getting-started prompt focuses on plugin setup

2026-11-01T00:00:00+00:00

The shared onboarding prompt at manufact.com/prompt.md now walks new users through installing the Manufact plugin or MCP connection before launching a first server or app.

Exempt AI platforms from per-caller rate limits

2026-11-01T00:00:00+00:00

When per-caller rate limiting is enabled, you can choose which AI platforms stay exempt so shared client IPs are not throttled for every user on that platform.

Two-step verification and Enterprise SSO

2026-11-01T00:00:00+00:00

Account Settings supports passkeys and authenticator-app codes as a second sign-in step, and Enterprise organizations can configure SAML or OIDC single sign-on from Team with optional domain enforcement.

Organization profiles for server sign-in

2026-11-01T00:00:00+00:00

Create and reuse OAuth and website sign-ins from Profiles, bind them per server and environment, and let Chat, testing, and publish checks share the same credentials.

LM Studio 0.4.28

2026-10-31T00:00:00+00:00

[Mac] Fixed "Restart to apply update" reopening the current version instead of installing the update. Fixed chat template parsing for Flint model.

Mintlify Base beta, mint migrate, mint instructions, and more

2026-10-30T23:21:21+00:00

Base

Mintlify Base (beta): A base is a new project type for your team's internal knowledge, such as specs, runbooks, and onboarding guides. Your team and its agents update a base from the editor, the admin MCP, Slack, or Git.
Share base pages to the web: After an admin turns on public links, share a base page with people outside your organization through a public link that also serves a Markdown version for AI tools. See Share a page to the web. Available on Pro and Enterprise plans.

Autopilot

GitLab projects: Autopilot works with repositories on GitLab and uses the GitLab projects you choose as context. See GitLab projects.
Chat in Slack: Mention @mintlify or message the Mintlify app to ask Autopilot questions from Slack. Admins and editors can also change its setup. See Chat in Slack.
Pull request comments: Comment on an Autopilot pull request on GitHub, and Autopilot pushes the requested changes to the same branch and replies in the thread. See Continue a conversation anywhere.
Multi-repository sites: Autopilot works on docs sites built from multiple repositories. Available on Enterprise plans.

CLI

mint migrate: Print step-by-step instructions for migrating content to Mintlify from 21 sources, such as Docusaurus, GitBook, ReadMe, Notion, and PDFs. Pass the --agent flag to have Claude Code, Codex, Cursor, or Gemini run the migration. See mint migrate.
mint instructions: Coding agents can look up how to do common Mintlify tasks, such as publishing a new site or adding a custom domain, instead of relying on their training data. See mint instructions.
mint agent setup: Install Mintlify skills and connect the admin MCP for Claude Code, Codex, or Cursor with one command. See mint agent setup.
mint validate rules: mint validate checks for content that builds but doesn't render correctly and lists each finding with its file, line, and rule. Repair common issues with --fix, choose rules with --enable and --disable, and get a JSON report with --format json. See mint validate.
Agent signup: A coding agent can create and deploy a site with mint deploy before you have an account, including on every push from GitHub Actions. Claim the site with mint claim. See Agent signup.

Editor

Switch repositories: In a multi-repository project, pick a repository from the menu above your pages to edit its pages, files, and branches. See Edit sources in the web editor. Available on Enterprise plans.
Unpublished changes in the sidebar: Sidebar pages show an Edited or Unpublished badge until you publish them. Projects with auto publish don't show the badges.
Agent branches: Ask the agent to list, create, or switch branches, and it reads, edits, and publishes on the branch it switched to. See Work on other branches.
Components defined in a page: Source mode autocompletes and describes components you define in the page, and jumps to their definitions with F12 or Cmd + click. See Source mode shortcuts.
New pages at the top: The New page button adds pages to the top of your navigation. See Create a page.
Move: The page menu's Move within project option is now Move. See Move a page.
Copy tables: Copy a table as CSV, JSON, Markdown, YAML, XML, or TOON from its block menu.
Auto publish timing: With auto publish on, the editor commits about a minute after you stop editing. See Auto publish.

AI and MCP

Page history: The admin MCP's read_page_history tool lists a page's saved versions and returns the content of any version. See Content.
Pages from uploaded files: Upload a local Markdown file and pass its uploadId to write_page or create_node instead of sending the full page content. See Uploads.
Inbox: The admin MCP's read_inbox tool lists comments, mentions, suggested edits, and invitations to your organization and to pages. See Comments.
Mintlify skill: The admin MCP's read_skill tool reads Mintlify's guide to writing docs, plus your organization's and your own agent skills. See Mintlify skill.
Read-only connections: Send the x-mintlify-toolset: knowledge header to limit an admin MCP connection to search and read tools. See Limit a connection to search and read tools.
Default project: On a connection with access to a single project, read tools no longer need a subdomain. See How editing works.
Projects with auto publish: In projects with auto publish, the admin MCP commits edits to your deploy branch and doesn't create branches. See Projects with auto publish.

Breaking change

Admin MCP upload tools renamed: upload_image is now upload_file, and finalize_image_upload is now finalize_file_upload. Update prompts or scripts that call the old names. See Uploads.

Published sites

Emails authentication: Invite readers outside your organization by email, and Mintlify hosts their sign-in with a password or a one-time code. See Emails authentication. Contact sales to turn it on for your site.
Full-page search results: Turn on Full-page search results in your Site settings to give readers a search results page at /search that they can bookmark and share. See Full-page search results.
Navigate from custom scripts: Call window.mintlify.navigate to send readers to another page without a full reload. See Navigate to another page.
Markdown export: In .md pages, cards, accordions, and tabs render as blockquotes that open with [card], [accordion], or [tab], so AI tools can tell them apart. Props written as JSX, such as a Tooltip's tip, render as Markdown. See Components.
Go SDK references: Go packages that mix kinds group their pages under Interfaces, Structs, Types, and Functions. Pass -exclude-symbols to the extractor to leave symbols out. See Generated pages.
TypeDoc overview pages: TypeDoc artifacts with a README get an overview page that indexes the library's symbols. See Generated pages.

Improvements

Desktop app tabs: Tabs signed in to different organizations work side by side. In the editor, Cmd + click a link to one of your pages to open that page in a new tab, or right-click it to open it in a new tab or window.
Thumbs-only feedback: The feedback export API includes thumbs up and thumbs down votes submitted without a comment. See Get feedback.
Site settings: The project General settings page is now called Site.
Previews when you turn on authentication: Turning on authentication through the API or admin MCP deletes previews that don't have authentication, matching the dashboard. To keep them, pass previewDeployments: "keep".
Deleted subdomains: After you delete a project, Mintlify reserves its subdomain for your organization for 30 days. See Delete a project.

Bug fixes

Editor: Code blocks you type with three backticks in a list item stay in that list item. Cards render inside custom components in visual mode, table cell selections stay after you drag, and custom component previews load your fonts and repository CSS. Character references and hidden export blocks survive saves and deletions. Parent page titles show in the sidebar after you refresh. The editor keeps syncing when browser storage stops accepting writes, and Cmd + J toggles the agent while you type in it.
Published sites: Static files load on sites with authentication on custom domains. RSS feeds work on sites with authentication. Go references no longer repeat description sentences and render GitHub alerts as callouts. Sphinx references keep their module navigation and type links.
CLI: mint validate no longer crashes during connectivity checks. mint dev picks up snippet edits you made while it was stopped. MDX parse errors suggest escaping <.
Dashboard: The sign-in page explains when your organization requires SSO.
Desktop app: Session refreshes no longer clear stored data or reload your tabs.

v3.11.238

2026-10-30T22:57:34+00:00

v3.11.238

v3.11.235

2026-10-30T22:57:34+00:00

v3.11.235

v3.11.234

2026-10-30T01:23:23+00:00

v3.11.234

The Harness and Memory Architecture Behind Self-Updating Docs

2026-10-30T00:00:00+00:00

How using fixed schema memory can improve agent performance at complex work.

Modern docs playbook

2026-10-30T00:00:00+00:00

Five plays from Coinbase, Anthropic, Anaconda, RV Tech, and Decagon on making knowledge work for humans and agents.

Autonomous AI Agents That Verify Their Own Work

2026-10-30T00:00:00+00:00

How autonomous AI agents verify their own work: criteria set first, evidence with provenance, a separate judge, and a human sign-off.

LOL #19 - Projects, Channels, and a Goal That Reviews Itself

2026-10-30T00:00:00+00:00

This week at LobeHub: work moves inside Projects, agents collaborate in channels and decide for themselves when to speak, goals get a review loop before sign-off, Resources become Docs, and Canvas and Editor 7.0 land.

v3.11.231

2026-10-29T22:45:39+00:00

v3.11.231

Introducing Mintlify Base

2026-10-29T00:00:00+00:00

Mintlify Base is now in beta. It's a workspace where people and agents work from the same specs, runbooks, and onboarding pages, kept current as your work changes.

Your wiki is becoming a codebase

2026-10-29T00:00:00+00:00

Agents now write and read most of what companies know, but wikis are still updated by hand. Knowledge needs the same care engineers give code: tracked, reviewed, and kept current by agents.

Build an A2UI app with MCP

2026-10-29T00:00:00+00:00

Let the chat model generate interactive A2UI layouts and render them in an MCP App view.

Self-Hosted AI on a Raspberry Pi: A Complete 2026 Deployment Guide

2026-10-29T00:00:00+00:00

Step-by-step guide to running local AI on a Raspberry Pi 5 — hardware picks, memory math, Ollama setup, and a self-hosted LobeHub stack.

LM Studio 0.4.27

2026-10-29T00:00:00+00:00

Improved LM Link reliability during large transfers. Improved runtime discovery reliabilty. Added Radeon 8065S detection for ROCm runtime compatibility. Fixed chat Markdown formatting bugs. Fixed switching to bundled MTP retaining a previously selected draft model. [Mac] Added maximum-memory and Neural Engine Prefill controls for Splash models. [Mac] Fixed compatibility checks for M5-only Metal runtimes.

Your agent can now publish your docs

2026-10-28T00:00:00+00:00

Agents can publish a live Mintlify site from the terminal without an account or GitHub repository. You can claim the site within 24 hours.

How to host an MCP server: hosting options and setup

2026-10-28T00:00:00+00:00

Give your MCP server a remote endpoint, configure access, and test it in real clients. Follow the deployment workflow and monitor tool usage.

AI Agent vs Chatbot: The Architecture Difference That Actually Matters

2026-10-28T00:00:00+00:00

A chatbot returns text; an agent returns a finished task. The five parts of the loop that separate them — and a test for yours.

Step-level flow control and retries in Upstash Workflow

2026-10-28T00:00:00+00:00

Upstash Workflow now lets you set flow control and retries on a single step with the third argument of context.run, instead of applying them to the whole run.

October 28, 2026

2026-10-28T00:00:00+00:00

Summary

Inboxes can now carry a signature that AgentMail adds to every message they send. Your agents no longer have to remember to append contact details or branding, and every message from the inbox looks the same whether it was sent through the API, the MCP server, or SMTP.

What's new?

signature on create and update: an object with text and an optional html. Send null on update to remove it.
Every send path: sends, replies, reply-alls, forwards, and drafts, including scheduled ones, get the signature right after your content and above any quoted or forwarded message.
No duplicates: a message whose body already contains the signature is sent as written.
Returned on the inbox:signature appears on inbox responses when one is set.

Use cases

Build support agents whose replies always end with the team's name, hours, and help-center link.
Give sales agents a consistent signature with a logo linked by URL, without changing agent prompts.
Update one inbox setting to change the signature on every agent that sends from it.

UK AI Search Index 2026: Key Statistics Revealed

2026-10-27T16:19:49+00:00

Explore the UK AI Search Index, revealing British insights on local domains and citation trends across various industries in AI Visibility.

How a Local SEO Agency Sells AI Search with OtterlyAI

2026-10-27T07:09:08+00:00

How Thencan Designs turned AI visibility report cards into $1,000+ MRR and 100% client retention. Local SEO meets GEO, powered by OtterlyAI from $29/month.

Multi-Model AI Platform: How to Route 30+ Models Without Losing Context

2026-10-27T00:00:00+00:00

How multi-model AI platforms route work across 30+ models without losing context — topic pins, task tiers, provider fallback, and compaction.

Context7 vs Exa 3.0 on WebCode: The Same Accuracy in a Fifth of the Tokens

2026-10-27T00:00:00+00:00

On Exa's WebCode benchmark, Context7 Search and Exa 3.0 led to as many correct answers, with about a fifth of the tokens for Context7 and a third of the wrong answers.

October 27, 2026

2026-10-27T00:00:00+00:00

Summary

Calendar invitations your agent builds itself and sends from an inbox now go on that inbox's calendar. Attendees' replies are tracked on the event, the same as for invitations sent with the Calendar API.

What's new?

Your own .ics invitations land on your calendar: an invitation the inbox sends as an attachment, with METHOD:REQUEST and the inbox as ORGANIZER, becomes an event with source: api. The sent email gets calendar_event_id, and a calendar.event.created webhook is sent. AgentMail emails nobody, because your email already went out.
Replies apply: attendees' accept, decline and tentative replies update the event and send calendar.event.responded, where before they were ignored.
Your later emails sync it: a newer REQUEST updates the event, and a CANCEL cancels it or removes the attendees it names.
Move to the API whenever you like:PATCH the event with send_invites: true and attendees get the update under your original UID.
Limits: up to 100 new imported events per inbox and 200 per organization each day. PUBLISH files, replies, single-date changes and invitations organized by someone else are not imported.

v4.1.4

2026-10-26T06:25:36+00:00

v4.1.4: Token Gesture

Generated CLIs now preserve remote transports, empty arrays, and parameters whose names collide with their own helpers; mounted servers honor session visibility and mounted_components_raise_on_load_error, a stale client context no longer closes a newer session, and OAuth provider wrappers accept identity assertions. Tasks accept a redis-py credential provider that returns current credentials for each new connection (streaming providers such as redis-entraid's are rejected until Docket can refresh them safely), Redis 8.1.0 is excluded, and custom KeyringStore users must now install py-key-value-aio[keyring] explicitly. Thanks to our 16 contributors, including 9 first-time contributors.

Enhancements ✨

Add server authoring guidance through the docs site by @zzstoatzz in #5703
Support rotating Redis credentials for tasks by @zzstoatzz and @nazarii-plebanskyi-rf in #5970
fix(client): forward max_sse_event_size through StreamableHttpTransport by @PrathmeshPawar2110 in #6016
Add a downstream smoke for Redis credential providers by @zzstoatzz in #6044

Security 🔒

Bump certifi to 2026.10.19 by @jakekaplan in #5989
Let configured proxy headers replace forwarded ones regardless of case by @leopranzl in #5972

Fixes 🐞

Fix maintained skill discovery on the docs site by @zzstoatzz in #5954
Publish regular skill files for Mintlify discovery by @zzstoatzz in #5956
Serialize HTTP and SSE transports in generate-cli by @asasemahmed in #5891
fix(cli): preserve empty arrays in generated commands by @dailyun in #5961
Fix session visibility lookups through mounted servers by @HootRock and @leopranzl in #5966
Fix generated CLI JSON parameter name collisions by @Vittuu in #5959
Escape server text in the fastmcp call elicitation prompt by @asasemahmed in #5909
Map common audio extensions when Audio is given a path by @asasemahmed in #5919
Drop clamped log records the logger is not enabled for by @asasemahmed in #5915
Omit the output schema when an OpenAPI success response has no body by @asasemahmed in #5913
Create the MCP config file on install when it is missing or empty by @asasemahmed in #5887
Preserve credential providers when cloning task extensions by @zzstoatzz and @nazarii-plebanskyi-rf in #5984
Exclude Redis 8.1.0 to preserve task reconnection by @zzstoatzz and @nazarii-plebanskyi-rf in #5985
Remove unused keyring dependency by @zzstoatzz and @tom42-otera in #5987
Fix static checks broken by ty 0.0.89 and pydantic 2.14 by @zzstoatzz in #6001
Run dev inspector servers over stdio and stop mapping the deployment port to the proxy by @asasemahmed in #6029
fix(json-schema): apply a Draft-4 boolean exclusive bound to its sibling by @kbkb628 in #5234
Strip a URL scheme from a Clerk domain by @asasemahmed in #5921
Fix generate-cli commands for tools with a json parameter by @andreivince in #6032
Honor mounted_components_raise_on_load_error in AggregateProvider by @asasemahmed in #6026
Reject streaming credential providers in TasksExtension by @zzstoatzz in #6043
fix(client): keep stale contexts from closing newer sessions by @mikamikasuki in #6018
fix(auth): expose identity assertions on OAuth provider wrappers by @computeaholic and @jonaseck2 in #6006
Measure blocking duration with perf_counter in transform_fn generator test by @Lokeshrao69 in #6014

Docs 📚

Clarify request state across mounted servers by @HootRock and @leopranzl in #5967
Pass mcp_path through the remote OAuth custom routes example by @asasemahmed in #5925

New Contributors

@dailyun made their first contribution in #5961
@Vittuu made their first contribution in #5959
@leopranzl made their first contribution in #5966
@nazarii-plebanskyi-rf made their first contribution in #5970
@tom42-otera made their first contribution in #5987
@computeaholic made their first contribution in #6006
@jonaseck2 made their first contribution in #6006
@PrathmeshPawar2110 made their first contribution in #6016
@Lokeshrao69 made their first contribution in #6014

Full Changelog: v4.1.3...v4.1.4

Introducing Manufact MCP v2.0

2026-10-26T00:00:00+00:00

Manufact MCP v1 made deployment decisions the agent could not see. V2.0 splits them into small tools that report facts, and every error names the next step.

Agentic Payments Hit 68.3M Transactions. The Total Value Was $3.7 Million.

2026-10-26T00:00:00+00:00

AI agent payments doubled to 68.3M in Q3 2026, but total value was just $3.7M. What a five-cent average means for agent builders.

Better than Grok Bot: Running Customer Success with Hermes on Upstash Box

2026-10-26T00:00:00+00:00

How I run Customer Success at a lean startup with Hermes, the open-source agent from Nous Research, living on an always-on Upstash Box and connected to my work apps through Upsy.

How Upsy Uses MCP Tasks to Keep Agents in a Meeting

2026-10-24T00:00:00+00:00

Upsy Meet is a meeting room for AI agents. Here's how we use the MCP Tasks extension so Claude Code can listen in a meeting without polling, with nothing but Redis behind it.

October 24, 2026

2026-10-24T00:00:00+00:00

Summary

message.received for an email with a calendar invitation now arrives with calendar_event_id already set, so your agent can answer the invitation straight from the webhook instead of fetching the message again.

What's new?

Invitation emails wait for their event: for an email with a calendar attachment, message.received (webhook and WebSocket) waits until the invitation has been added to the inbox's calendar, at most 20 seconds, and the message carries calendar_event_id.
Other email is unchanged: only emails with a calendar attachment wait, typically a second or two.
No calendar_event_id means the invitation wasn't added: a reply or cancellation for an event the calendar doesn't have, a sender that couldn't be verified as the organizer, or a slow import. Get the message again later to check.

Autopilot, dashboard inbox, editor agent skills, and more

2026-10-23T23:13:30+00:00

Autopilot

Autopilot (beta): Autopilot is a proactive agent that opens pull requests to keep your documentation up-to-date, and performs ongoing maintenance such as a daily quality check. Turn it on from the Autopilot page in your dashboard. Available on Pro and Enterprise plans.
Review suggestions: Autopilot's pull requests appear under Suggestions, where you can preview, accept, close, or ask for changes in a chat. See Review suggestions.
Feature launches: For a large launch, Autopilot tracks the feature and keeps a draft pull request current until the code merges. See Feature launches.
Integration events: Ask Autopilot in a chat to run when an event happens in a connected shared integration, such as a new Linear issue. See Slack and other integrations.
Signed commits: GitHub signs Autopilot's commits and shows them as Verified, so they work with branch protection rules that require signed commits. See How Autopilot opens pull requests.

Editor and dashboard

Agent skills: Add reusable instructions that the agent loads when a request matches them. Create skills, upload SKILL.md files, .skill or .zip archives, or folders, or ask the agent to save a skill while you chat. Admins can move personal skills to your organization.
Inbox: The inbox collects replies, mentions in comments and page content, comments on pages you own, and shared private pages. Filter by section and date range, and mark everything as read at once.
Edits from comment threads: Ask @mintlify in a comment to change a page and the agent performs edits then updates the comment thread. See Ask the Mintlify agent.
Show all branches: Choose whether the branch menu lists every branch in your repository. See Show all branches.
Commit co-authors: Credit teammates as co-authors on editor commits. See Editor settings.
Delete agent chats: Delete chats from the agent's session history.
Source mode: Fold component blocks and code fences and keep your scroll position when you switch between visual and source mode. JSX attribute expressions highlight as JavaScript, and style objects autocomplete CSS properties and describe them on hover.
Mermaid diagrams: Visual mode hides Mermaid code by default and remembers whether you show it.
Inline SVG and custom pages: Visual mode draws inline <svg> blocks. On pages in custom or frame mode, components respond to clicks and you can select and copy text.
Page titles in URLs: Editor URLs and copied page links include the page title, such as /~/Getting-Started-<id>. Existing links still work.
Copy and move pages: The page menu's Copy to… and Move to… options are now Copy to a project and Move within project. After you copy a page to another project, the editor opens the copy. See Copy a page to another project.
Billing page: Set overages and spend limits for each project and compare plans on the redesigned Billing page. On Pro and Enterprise plans, reserve credits for a project. See Overages.
Automations in project settings: Automations moved to your project settings.

Desktop app

Copy tab links: Right-click a tab and select Copy Link to copy the link to its page.
What's new: See recent changes to the desktop app in its settings.

Published sites

Page styles: Add a <style> block to a page to style only that page's content. See Custom CSS.
Styling hooks: Style the search bar, search dialog, and theme layout with stable data-search-item and data-layout-item attributes. Log in and log out links in the sidebar have the login-link and logout-link classes, like the links in the navigation bar. See Class and element hooks.
JSON request bodies: Switch between Form and JSON to edit request bodies in the API playground. See Edit request bodies as JSON.
Optional API fields: Set api.optional to collapse, expand, or always show optional fields in the API playground. See API settings.
Markdown export: .md pages render components, Visibility content, and snippets as plain Markdown, and keep the text of components such as Frame, Tile, and Prompt. Links to pages in your navigation point to their .md versions, and on most sites internal links use your site's full URL. See Components and HTML.
Go SDK references: Generate SDK reference pages for Go packages with the go format and mintlify-go-extractor. See Go.
Javadoc navigation: SDK reference navigation for Javadoc artifacts groups symbols by kind within each package. See Generated pages.
Classic thumbnails: Generated OG images use the classic layout by default. Thumbnails with a custom background image and no preset keep their original layout. See thumbnails.
Google Fonts from your domain: On sites without authentication, Mintlify serves Google Fonts from your docs domain, so readers' browsers don't connect to Google.
Assistant widget text size: Set appearance.fontSize to a number from 12 to 20 to change the text size of the assistant widget in pixels.
Preview access through the API: Pass auth to the Trigger preview deployment endpoint to protect a preview with a password or dashboard login.
Old previews: If you change your repository, Mintlify deletes previews from the previous repository after seven days. See Preview lifetime.

CLI

Generate docs.json files: mint dev generates a docs.json file from your folder structure when your directory doesn't have one. See Local preview.
Preview on other devices: mint dev listens only on localhost. To open the preview on a phone or another computer, run mint dev --host. See Preview on other devices.

Breaking change

mint test removed: The CLI no longer includes the mint test command. Remove it from scripts and CI workflows that call it.

AI and MCP

Admin MCP editing: edit_page applies up to 100 replacements in one call, create_node writes into a page that already exists at the path, and write_page and edit_page report markup the editor drops. See Content.
Admin MCP navigation: list_nodes returns fewer fields for each node and leaves out uploaded files unless you filter for them. Update scripts that read other fields. list_deployments returns each project's projectType. See Navigation.
Admin MCP deploy status: deployment_status reports whether published changes are live, and suggest_edit accepts a comment that explains the change. See Branches and publishing.
Bring your own model: Starter questions and the commit messages and pull request details for editor publishes also run on your key. See Bring your own model. Available on Enterprise plans.
API request examples: When readers ask the assistant how to call an API, its answers include a minimal request example based on your docs.
Assistant API limits: The assistant API no longer has a per-IP daily request limit. See Create assistant message.
Conversation IDs: Assistant conversation exports include a conversationId that groups turns from the same conversation. See Assistant conversations.

Bug fixes

Editor: Imported .md files parse as Markdown, and Notion Markdown exports are cleaned up on import. HTML links to sections on the same page scroll in visual mode. Bold and italic formatting stays after you post a comment. New branches no longer reuse names of deleted branches. Interrupted agent responses reconnect. OpenAPI page descriptions render components. Typing --- in source mode starts frontmatter. Template literals inside HTML no longer gain indentation when you save, and inline HTML tables keep their layout when you click a cell.
Published sites: Switching versions keeps you on the same page when the other version files it elsewhere. Price-like dollar amounts render as text instead of math. Scrolling in Safari continues over horizontal scrollbars. MCP discovery documents use your custom domain. On sites with authentication, the contextual menu's MCP options point to the authenticated MCP server. RSS items keep their original dates. Tailwind classes passed through variables and expressions get their styles.
API playground: Request examples use server variable defaults when the playground is hidden. Requests without a body sent through the playground proxy no longer fail with 415 errors.
Previews: View Preview links open the preview's home page when the linked page isn't in the build. Leftover previews from a previous repository no longer block new previews.
Offline exports: API playground icons load from the export.
Desktop app: Switching organizations no longer reloads every tab. Tabs show the editor branch, and new tabs keep your theme.
Dashboard: Visit page stays disabled until your site's first build is live. Pending custom domains appear in your custom domain settings. Choosing no spend limit no longer turns off overages. Projects you delete leave the project list right away, and a new project that reuses a deleted project's subdomain starts with an empty editor.
Automations: Scheduled Update from code changes runs no longer skip commits that are pushed after their commit date.

MCP Tasks and Elicitations in Claude Code: Long-Running Tools That Can Ask You Things

2026-10-23T00:00:00+00:00

Articles and tutorials on serverless technologies from Upstash and community

Claude Code vs Codex (2026): Which Coding Agent to Use, and How to Run Them Together

2026-10-22T00:00:00+00:00

Claude Code or Codex? Compare five coding agents (Claude Code, Codex, Pi, Amp and OpenCode) and learn how to run them side by side in one LobeHub window with your own subscription.

Third-party libraries for your documentation

2026-10-21T21:23:29+00:00

Third-party libraries

Import React components from npm packages into your public documentation. Add interactive charts or reuse your design system in MDX pages.

Third-party libraries require an Enterprise plan. Contact your account manager to enable them for your site.

Follow the setup guide to add dependencies, commit a lockfile, and configure packages in docs.json. Use only trusted packages and review the limitations before getting started.

v2.11.1

2026-10-21T14:54:53+00:00

Inspector chat & hosted connections

Patch release updating the bundled Inspector and CLI, with no server API changes.

Update: Updated @mcp-use/inspector to v20.3.23, fixing thinking feedback between tool calls, chat backgrounds during history resizing, and hosted MCP proxy routing (#2891, #2899, #2900); see the Inspector changelog
Update: Updated @mcp-use/cli to v4.5.1 with synchronized Inspector dependency metadata

What Are MCP Apps? The Upstash Console Inside ChatGPT

2026-10-21T00:00:00+00:00

What an MCP App is, how hosts render one, and how the Upstash console runs as an MCP App in ChatGPT.

v4.1.3

2026-10-20T22:32:43+00:00

v4.1.3: Root Awakening

This release updates certifi-backed dependencies following the eMudhra CA compromise and fixes public-client OAuth authentication, empty visibility selections, generator transforms, and per-server timeouts for ordinary requests. FastMCP 4 uses system trust by default: follow the security notice to update operating-system, container, custom, and installer trust stores; upgrading FastMCP alone is insufficient. Thanks to our 11 contributors, including 6 first-time contributors.

Security 🔒

Require certifi 2026.10.19 or later for certifi-backed server dependencies and optional SDK integrations, and document system, custom, and installer trust-store updates by @zzstoatzz in #5945.

Enhancements ✨

Chart public backlog counts and closure outcomes by @zzstoatzz in #5942

Fixes 🐞

Default the upstream token auth method to none without a client secret by @asasemahmed in #5879
fix: preserve empty session visibility selections by @himanshuchawla0807-gif in #5944
test(mount): make unreachable proxy failures deterministic by @CharlesYXW and @ch-z-hc in #5584
fix(tools): materialize generator transform_fn results in TransformedTool.run by @BlueX888 and @nandanadileep in #5672
Honor per-server MCP config timeouts by @atty57, @phnmn, and @CRYPTONIKAV in #5758
fix: don't classify a tool body's own TimeoutError as the tool timeout by @Wilson0511 in #5940

New Contributors

@phnmn made their first contribution in #5758
@CRYPTONIKAV made their first contribution in #5758
@himanshuchawla0807-gif made their first contribution in #5944
@CharlesYXW made their first contribution in #5584
@atty57 made their first contribution in #5758
@Wilson0511 made their first contribution in #5940

Full Changelog: v4.1.2...v4.1.3

v4.1.2

2026-10-20T14:01:55+00:00

v4.1.2: License to Deny

Component authorization now applies after transforms, denied resources cannot fall through matching templates, and OAuth client credentials providers support issuer binding across fresh and cached-token flows. Existing configurations remain compatible, but users of ClientCredentialsOAuthProvider or PrivateKeyJWTOAuthProvider must set issuer to enable this protection; upgrading alone is insufficient. Thanks to our 18 contributors, including 9 first-time contributors.

Security 🔒

Authorize transformed components and prevent denied-resource template fallback by @zzstoatzz and @rishid in #5933
Bind OAuth client credentials to a configured issuer, including cached-token renewal, by @zzstoatzz and @LHMisme420 in #5933

Enhancements ✨

Add SEP-2640 Skills extension foundation by @strawgate in #5057
Write down triage and merge norms in docs/development by @zzstoatzz in #5872
Add experimental Skills discovery to SkillsDirectoryProvider by @zzstoatzz in #5704

Fixes 🐞

Do not require websockets for the OAuth callback server by @asasemahmed in #5693
Keep fastmcp.json resolvable when deployment.cwd changes the working directory by @asasemahmed in #5700
Fix ctx.log() failing when extra contains LogRecord attribute names by @asasemahmed in #5688
Use a monotonic clock in the rate limiters by @asasemahmed in #5689
Escape server metadata in the fastmcp inspect text summary by @asasemahmed in #5701
Respect an explicit is_binary=False on FileResource by @asasemahmed in #5696
Follow pagination in the MCP format of fastmcp inspect by @asasemahmed in #5691
Visit each reference once when finding access-restricted properties by @AarnavNoble in #5751
Run sync resource templates in the threadpool by @hardness1020 in #5742
Require openapi-pydantic 0.6.0 so OpenAPI 3.1.2 specs load by @jsphbtst in #5647
Scope FileUpload and session visibility to a stable client on MCP 2026-07-28 by @zzstoatzz in #5651
fix(auth): accept private_key_jwt without form client_id by @HootRock and @imranq2 in #5832
fix(cli): honor explicit URL transport by @mikamikasuki in #5529
Remove legacy logging state when HTTP sessions close by @mikamikasuki in #5539
Name data-backed File resources consistently with their MIME subtype by @nandanadileep in #5568
fix(client): run Python stdio children in UTF-8 mode by @ningmao-hlyz in #5642
Report an unbindable OAuth callback port instead of SystemExit by @asasemahmed in #5692
Accept ES256 and other RSA/EC algorithms for CIMD private_key_jwt assertions by @asasemahmed in #5695
Preserve icons on transformed tools by @asasemahmed in #5702
Hydrate valid schema defaults while preserving invalid defaults by @nandanadileep in #5738
fix(resources): preserve public metadata in built-in read contents by @HootRock in #5778
Preserve shared dependency scopes without optional task dependencies by @HootRock and @BlueX888 in #5817
fix: treat non-string skill frontmatter descriptions as absent by @BlueX888 in #5823
fix(resources): preserve component meta for list[ResourceContent] returns by @BlueX888 in #5843
Fall back to the client timeout when resolving claimed results by @marcinbelczewski in #5190
Discover Goose Streamable HTTP extensions by @jiuyueQ and @ch-z-hc in #5582
Raise ToolError when a tool with an output schema returns an unserializable value by @massimiliano1991 in #5063
Stop deploy tests from spawning PowerShell on every state access by @zzstoatzz in #5874
Give the timeout-override test headroom for the follow-up tools/list by @zzstoatzz in #5875
Keep disabled resources from being served by matching templates by @zzstoatzz in #5876

Docs 📚

Fix codespell typo in v4.1.1 changelog by @zzstoatzz in #5818
context: stop documenting session_id as a cross-request key by @zzstoatzz in #5862

Examples & Contrib 💡

fix(contrib): preserve BulkToolCaller tool prefixes by @mikamikasuki in #5555

Other Changes 🦾

chore(deps): Update github actions by @prefect-renovate[bot] in https://github.com/PrefectHQ/fastmcp/pull/5707

New Contributors

@AarnavNoble made their first contribution in #5751
@hardness1020 made their first contribution in #5742
@jsphbtst made their first contribution in #5647
@marcinbelczewski made their first contribution in #5190
@jiuyueQ made their first contribution in #5582
@massimiliano1991 made their first contribution in #5063
@imranq2 made their first contribution in #5832
@rishid made their first contribution in #5933
@LHMisme420 made their first contribution in #5933

Full Changelog: v4.1.1...v4.1.2

Bring Your Own Key: Accuracy, Cost & Latency by Model

2026-10-20T00:00:00+00:00

Accuracy, cost and latency for models you can bring to Mem0, with our picks for best value, best under $1 and best quality.

Managing Time in AI Memory

2026-10-20T00:00:00+00:00

Every memory has two clocks: when it happened and when it was stored. Eight places where it breaks in consumer agents, and the Mem0 settings that fix each one.

v2.11.0

2026-10-19T20:47:01+00:00

Custom landing pages

Minor release adding a customizable page at the MCP URL with landing.tsx, plus agent streaming, OAuth client, and CLI fixes.

New (server, @mcp-use/cli): a landing.tsx in the MCP root replaces the page a browser sees at the MCP URL. mcp-use dev and mcp-use build find it (also under --mcp-dir) and mcp-use start serves the built page. It renders on the server and hydrates in the browser; without it, the built-in page is served as before (#2648)
New (server): mcp-use/landing exports React-free hooks for the page: useServerInfo(), useToolList(), usePromptList(), useResourceList() and useInstallTargets(). Server info always ships to the page; each list ships only when the page's import graph uses its hook (#2648)
New (server): getInstallTargets({ url, name }) returns install links and commands for Claude Code, Cursor, VS Code, VS Code Insiders, ChatGPT and the Manufact Inspector, for use outside the server such as on a docs site (#2648)
New (server): landing.tsx can export const metadata (title, description, icon, image) to set the page title, meta description, and Open Graph and Twitter tags (#2648)
New (@mcp-use/cli): Tailwind classes work in landing.tsx without a CSS file and CSS imports are bundled. In dev, edits use Fast Refresh, adding or deleting landing.tsx switches pages live, and a broken edit keeps serving the last working page. Builds honor MCP_ASSETS_URL (#2648)
New (create-mcp-use-app): every template includes a Tailwind-styled landing.tsx; blank and mcp-server add React and its types, and every template's tsconfig.json typechecks landing.tsx (#2648)
Fix (server, @mcp-use/cli): landing pages set data-theme from the OS color scheme before first paint so dark: classes work; a page whose CSS imports Tailwind itself owns Tailwind; and in mcp-use dev the page's stylesheets load before hydration (#2880)
Update (server): ChatGPT install steps follow the Plugins flow (Plugins → Add → Add Custom MCP Server) instead of developer mode and the Connectors tab (#2648)
Fix (@mcp-use/cli): mcp-use start --with-inspector serves the Inspector at /inspector when basePath is "/", and mcp-use dev prints /inspector instead of //inspector (#2555)
Fix (@mcp-use/cli): mcp-use client and mcp-use screenshot install a matching @mcp-use/client when they auto-install it, including on Windows where the install always failed (#2853)
Fix (@mcp-use/agent): the OpenAI Responses provider emits tool-call argument events, so MCPAgent.stream() and streamEvents() report steps for tool calls on OpenAI models (#2551)
Fix (@mcp-use/agent): MCPAgent.stream() returns and saves to memory only the final turn's text, the same as run(), instead of gluing on text written before tool calls (#2723)
Fix (@mcp-use/agent): RemoteAgent.run no longer rejects successful responses whose body has no error field, and error messages show the error instead of [object Object] (#2651)
Fix (@mcp-use/client): useMcp().authTokens.expires_at no longer moves later on every reconnect for opaque (non-JWT) tokens; the expiry is saved when tokens are received (#2879)
Fix (@mcp-use/client): Node OAuth loopback connections close after the callback, cancellation, or timeout, so the process can exit (#2766)
Update: Updated @mcp-use/cli to v4.5.0, create-mcp-use-app to v2.1.0, @mcp-use/client to v2.4.5, and @mcp-use/agent to v2.0.25
Update: Updated @mcp-use/inspector to v20.3.22; see the Inspector changelog

October 19, 2026

2026-10-19T00:00:00+00:00

Summary

Calendar is now available to every organization in US production, with no opt-in. Every inbox already has a calendar: your agents can create events, send and answer invitations, and react to start and end webhooks without joining AgentMail Labs first.

What's new?

No opt-in: Calendar is no longer in AgentMail Labs. The calendar API works for every organization in US production (api.agentmail.to).
Invitations fill the calendar: calendar invitations emailed to any inbox are now added to its calendar, where before they were added only for organizations in Labs.
Console: every organization in US production sees the Calendar view and the calendar API key permissions.
Availability: Calendar is not yet available in EU production (api.agentmail.eu), which returns 403 with the message Calendar isn't available in this region yet.
Agent Armor stays in AgentMail Labs.

Devin updates — October 19, 2026

2026-10-19T00:00:00+00:00

Clearer Work Log While Devin Works The current step in the work log shows what Devin is doing as soon as that summary is ready instead of "Deciding action", and status lines no longer narrate internal housekeeping. Devin Oncall Is Now Generally Available Devin Oncall is now available to all organizations. Responders watch your Slack or Microsoft Teams channels and PagerDuty incidents, triage new alerts, bug reports and support requests, and investigate them using your telemetry tools, code, deployments and runbooks to post a likely root cause and recommended next actions. Devin can also join incident channels to keep investigating and maintain a self-updating incident report, and recurring O

v4.1.1

2026-10-18T02:14:29+00:00

v4.1.1: Regex Marks the Spot

Regex tool search now rejects patterns over 1,000 characters and stops after about a second, and SkillsDirectoryProvider discovers skill folders that are links again, including links to folders outside the skills root; a skill's own files must still resolve inside the folder it links to. All 4.x users are encouraged to upgrade. Thanks to our 4 contributors, including 1 first-time contributor.

Security 🔒

Keep linked skill folders discoverable by @zzstoatzz in #5766
Bound regex tool search by pattern length and time by @zzstoatzz and @SAYUTIM in #5771

Fixes 🐞

Keep an unparseable proxy setting from aborting the update check by @zzstoatzz and @ker2xu in #5639
Keep the JWKS refresh lock clear of subclass attributes by @zzstoatzz in #5770

Examples & Contrib 💡

fix: convert version query param to VersionSpec in component manager routes by @BlueX888 in #5575

New Contributors

@ker2xu made their first contribution in #5639

Full Changelog: v4.1.0...v4.1.1

v2.10.1

2026-10-17T19:41:55+00:00

OAuth discovery & browser tool calls

Patch fixes for public MCP connections, OAuth credential binding, browser tool calls, and ChatGPT widget descriptions.

Fix (@mcp-use/client): validate OAuth protected-resource metadata with the official SDK before classifying an anonymous connection as mixed auth. Root metadata for a different resource no longer makes a public MCP endpoint appear to require authentication (#2858)
Fix (@mcp-use/client): discard cached OAuth credentials with a missing, empty, or mismatched issuer stamp. Affected users may need to sign in again (#2858)
Fix (@mcp-use/client): browser tool calls send Mcp-Param-* headers for parameters annotated with x-mcp-header, so servers that require those headers accept calls from clients such as the Inspector (#2803)
Fix (server): copy a tool view's description to _meta["openai/widgetDescription"] on resources/list and resources/read, restoring the widget description for ChatGPT (#2845)
Update: Updated @mcp-use/client to v2.4.4, @mcp-use/agent to v2.0.24, and @mcp-use/cli to v4.4.1
Update: Updated @mcp-use/inspector to v20.3.21; see the Inspector changelog

API reference pages in the editor, Admin MCP improvements, and more

2026-10-16T22:16:57+00:00

Breaking changes

Mintlify Index discontinued: The Index API returns 410 Gone and the Index MCP server tells connected agents to remove it. The dashboard no longer shows Index API keys, and the CLI no longer includes the mint index command.
Admin MCP tool changes: The admin MCP renames save to publish. Tools that edit content require a branch argument, and create_branch, discard, and status replace checkout, discard_session, and get_session_state. Dedicated project management tools replace the search_code_operations and execute_code code mode tools. Update prompts or scripts that call the old tools.

OpenAPI and API playground

OpenAPI 3.2: Mintlify accepts OpenAPI 3.2 documents. Tag hierarchies become nested navigation groups, and the API playground supports 3.2 additions such as the QUERY method and whole query string parameters. To upgrade a 3.1 document, change its openapi version.
Streaming responses: The API playground shows streaming responses as they arrive. Click Stop to end the request.

Editor and dashboard

API reference pages in the editor: Pages with openapi or api frontmatter show their endpoint and generated fields in the editor, so you can edit the content around them. See Pages.
Section links: Links to /page#heading open the page at that heading. Copy link to section in the block actions menu copies a dashboard link to the heading, and pasted section links show the section name.
Thumbnail settings: Set thumbnail overrides in page settings and in the settings for groups, tabs, and other navigation sections. See Edit page settings.
Unpublished site configuration: The site configuration view lists the sections with unpublished changes in a banner and marks them in the sidebar.
Print a page: Press Cmd + P to print the page content without the editor interface. See Keyboard shortcuts.
Agent mentions: Ask the agent to add teammate and page mentions to your content.
Haiku 5.5: Choose Haiku 5.5 in the agent's model picker.
Publishes: The Activity section of the dashboard is now Publishes, with Overview and Previews pages. See Deployments.
Admin MCP in the sidebar: Admin MCP settings and credentials have their own MCP item in the dashboard sidebar. Search MCP client credentials moved to the Docs MCP section of your project settings.
Regional translation targets: The Translate content automation can translate into British, Australian, and Canadian English and into Mexican and Latin American Spanish. See Manage automations.

Published sites

OG image presets: Choose a classic, headline, or minimal layout for generated OG images, with custom fonts, a brand or neutral background, and an eyebrow label above the title. Override the settings for individual pages and sections. See thumbnails.
Preview OG images: Run mint og to render a page's OG image to a PNG file.
Visibility by sign-in state and group: Show content only to signed-in readers, signed-out readers, or specific groups, and hide it from groups with excludes. See Visibility.
Regional language variants: navigation.languages accepts regional codes such as en-GB, es-MX, and es-419, and the language switcher shows the region. Language routing sends visitors to the variant that matches their browser. See Regional variants.
hreflang tags: Sites with multiple languages link each page to its translations with hreflang tags, so search engines can show readers the right language. See Alternate language links.
TypeDoc categories: SDK reference navigation follows the @category and @group tags in your TypeDoc artifact. See Generated pages.
Preview deployments: The robots.txt file on preview deployments disallows crawling.
Navigation on authenticated sites: Signed-out visitors on sites with authentication see the public navigation on first load, without layout shifts.
Server-rendered custom components: Custom components that you define with export and that only place their props into HTML render on the server, so they no longer pop in after the page loads. See React components.

AI and MCP

Admin MCP comment tools: The admin MCP can read comment threads, start a comment, suggest a tracked change, reply, and resolve threads. See Comments.
Merge from the admin MCP: Pass merge: true to publish to merge the branch's pull request when your project allows edits to the deploy branch. See Branches and publishing.
User ID on MCP events: Search tool calls through an authenticated MCP server record the signed-in user's sub claim as user_id in analytics streaming. Available on Enterprise plans.
Redacted assistant conversations: Email addresses and phone numbers in assistant conversations appear redacted in the dashboard, CSV exports, and the analytics API. Available on Pro and Enterprise plans.

Bug fixes

Editor: Footnotes render instead of showing a warning. Keyboard text selection works across tables. Comment cards stay anchored when you switch tabs, and block menu actions follow their block when content shifts. Autolinked URLs include their paths, queries, and sections and leave out trailing punctuation. Bare domains link to https://. The docs.json tab is titled Site settings.
Comments: The comment popover hides reply and actions when you can't comment. If your organization uses roles, admins can delete any comment or thread.
Published sites: Tabs no longer clip ordered list numbers from 10 up. Related pages pick the right language for index pages.
CLI: mint a11y checks files that use custom heading IDs. mint broken-links reports every page it can't parse instead of stopping at the first, no longer flags valid code inside <MDX> fragments, and resolves relative links in snippets.
Desktop app: Switching organizations no longer switches you back to the previous organization.
Multi-repo: Each GitHub source fetches content through its own installation.
Analytics API: Export date parameters accept ISO 8601 time zone offsets and fractional seconds.

Devin updates — October 16, 2026

2026-10-16T00:00:00+00:00

Multiple Side Chats from One Message A message that already has a side chat can start another: its side-chat button opens a menu of the message's threads with a "New side chat" option. Command Palette Finds Archived Sessions Cmd+K session search now includes archived sessions, marked with an icon. Queued Messages Stay Editable While Devin Is Idle Queued messages can be sent, edited, removed or reordered after Devin stops or the session goes to sleep, instead of being locked until the next turn. Open Chat Images in a Tab Images in the chat can be opened as a tab from the image viewer, so you can view an image alongside the conversation. Numbers Follow Your Display Language Numbers across Devi

National Design Studio and Cognition Launch the American Forge Residency

2026-10-16T00:00:00+00:00

National Design Studio and Cognition are looking for engineers to help modernize America’s federal software infrastructure.

How Does MCP Work? The Protocol Step by Step (2026)

2026-10-15T00:00:00+00:00

How does MCP work? A step-by-step walkthrough of the 2026-07-28 spec: discovery, tools/list, tools/call, elicitation, notifications and auth, with real JSON.

MCP Client Guide: What It Is, Top Clients and How to Build One

2026-10-15T00:00:00+00:00

What an MCP client is, how it differs from the host, which MCP clients support which transports and features, and how to build one in TypeScript or Python.

A2A vs MCP: Differences, Use Cases and How They Fit

2026-10-15T00:00:00+00:00

A2A vs MCP explained: MCP connects an agent to tools and data, A2A connects agents to other agents. Side-by-side table, real payloads, and when to use each.

Blender MCP with Claude Code: Step-by-Step Setup Guide

2026-10-15T00:00:00+00:00

Set up Blender MCP with Claude Code: install uv, run one claude mcp add command, enable the Blender add-on, verify with /mcp, then prompt and fix common errors.

Figma MCP Claude Code Setup: Remote, Desktop or Framelink

2026-10-15T00:00:00+00:00

Figma MCP Claude Code setup: connect the official remote or desktop server, or community Framelink, sign in, verify with /mcp and pick the right option.

How to Install GitHub MCP in Claude Code (Remote or Local)

2026-10-15T00:00:00+00:00

How to install GitHub MCP in Claude Code: add the official remote server with a PAT or run it locally with Docker OAuth, then set toolsets and read-only.

How to Add Supabase MCP to Claude Code: Setup Guide

2026-10-15T00:00:00+00:00

How to add Supabase MCP to Claude Code: one claude mcp add command, OAuth login via /mcp, then scope to one project, go read-only and pick features.

Build MCP Server: Step-by-Step Guide with TypeScript SDK v2

2026-10-15T00:00:00+00:00

Build an MCP server from scratch: design good tools, write a TypeScript server for stdio and Streamable HTTP, test it, connect it to Claude Code, then publish.

v2.10.0

2026-10-14T23:07:52+00:00

OAuth Proxy for non-DCR providers

Canary release (mcp-use@2.10.0-canary.4) adding an OAuth authorization server in front of identity providers without dynamic client registration, secret-free production builds, and .gitignore-aware deployments. Stable v2.10.0 has not shipped.

New (server): oauthProxy() in mcp-use/oauth makes the MCP server its own authorization server in front of GitHub (github()), Google (google()), or any OAuth 2.0 provider (upstream()). Clients present a client ID metadata document or register (DCR), approve a server-owned consent page that names the server by its host, and receive the proxy's own opaque tokens (#2822)
New (server): the clients option limits which clients may connect, by metadata document URL, origin, or a client defined on the server (#2822)
New (server): the upstream client secret and user tokens stay on the server, encrypted in a shared store from mcp-use/oauth/storage: Upstash, Redis, Postgres, Firestore, Cloudflare D1 or Durable Object, or in-memory for local development. redisStorage takes a connected client or a lazy connect function (#2822)
New (server): tools read the upstream token with upstreamToken(ctx.auth) or ctx.auth.upstream (#2822)
Enhancement (server): failed sign-ins and refreshes log the cause without token or code values (provider error code and description, HTTP status and content type, mismatched iss, or network error); refused client metadata documents and records encrypted with a retired secret are also logged (#2822)
Enhancement (server): clients that signed in through the 1.x oauthProxy keep working while registration is on; users approve the consent page once and sign in again without removing the server (#2822)
New (server): OAuth providers can register public HTTP routes from setup() with host.route(), read host.mode and host.listenHost, and compute oauthMetadata from the resolved resource URL (#2822)
Fix (server): OAuth endpoints under the MCP path are never written to trace logs, and oauthEnvironmentValue no longer throws on runtimes without process (#2822)
Fix (server): trailing-slash trimming in upstashStorage and the GitHub and Google upstream providers no longer uses a backtracking regex
New (server, @mcp-use/cli): mcp-use build evaluates the entry with MCP_USE_PHASE=build and no longer needs runtime secrets; the new isBuilding() export lets module-scope code skip reading secrets or connecting to a database. oauthProxy and upstashStorage require no secrets or credentials while building (#2830)
Fix (@mcp-use/cli): mcp-use build exits when the build finishes, even if the entry opened a connection at module scope such as a Redis client (#2830)
Fix (@mcp-use/cli): mcp-use build builds for production when NODE_ENV is unset. Views no longer ship React's development build, and server-side JSX no longer compiles to jsxDEV, which failed under mcp-use start (#2830)
Fix (@mcp-use/cli): a failed build or dev start prints the cause of a refused connection instead of an empty line (#2830)
New (@mcp-use/cli): mcp-use deploy --no-github applies .gitignore rules when packaging, including wildcards, directory rules, and negation, while keeping the baseline cloud exclusions (#2485)
Fix (server): MCPServer.fromOpenAPI converts OpenAPI 3.0 boolean exclusiveMinimum and exclusiveMaximum into JSON Schema numeric bounds instead of throwing while registering tools (#2717)
Fix (@mcp-use/agent): the Anthropic provider sends "[no content]" for empty tool result content, avoiding HTTP 400 errors from the Messages API (#2693)
Documentation: add the OAuth Proxy guide, OAuth Proxy storage, v1 migration notes, and API reference; add GitHub and Google oauthProxy examples with Firestore (#2841, #2831)
Documentation: correct API reference comments: stdio env does not inherit the parent environment, defaultRequestOptions applies only to the initial tool listing, VMExecutorOptions.memoryLimitMb is not enforced, legacy stateless GET, DELETE, and HEAD return 204, and mcp-use/vite, mcp-use/tanstack-start, and mcp-use/tanstack-start/vite appear in the generated reference (#2838)
Update: patch vulnerable transitive dependencies (proxy-addr, shell-quote, source-map-js) and update Next.js examples to next ^16.3.8 (#2842)
Update: Updated @mcp-use/cli to v4.4.0 (canary.2)
Update: Updated @mcp-use/agent to v2.0.23 (canary.1) and @mcp-use/client to v2.4.3 (canary.0)
Update: Updated @mcp-use/inspector to v20.3.20 (canary.3); see the Inspector changelog

October 14, 2026

2026-10-14T00:00:00+00:00

Summary

Agent Armor can now hold mail. In Hold mode, your agents hear about a message only once its scan finishes, and a message marked review stays held from them until someone releases it.

What's new?

Hold mode: an organization admin switches Agent Armor between Monitor, the default, and Hold in the console under Settings, then Labs. Hold sends no message.security.completed: the result arrives on message.received, or on message.security.review for a held message.
agent-armor-held label: a held message carries it until it is released, and its thread while any of its messages is held. List them by it, with include_spam, include_blocked and include_unauthenticated, to see held mail.
message.security.review: a webhook and WebSocket event sent when a message is held.
Release:POST /v0/inboxes/{inbox_id}/messages/{message_id}/agent-armor/release releases a held message with a reason. Its message.received then goes to your agents, and message.security.override is sent too, in either order. Admins can also release from the console.
agent_armor_release permission: opt-in, so an agent's own key can't release mail held from it. See Permissions.
manual_override: a released message's agent_armor result records how, when and why it was released.

Use cases

Keep mail that may manipulate an agent away from it until a person has checked it.
Give a review tool, not your agents, the key that releases held mail.

Devin updates — October 14, 2026

2026-10-14T00:00:00+00:00

Chat Width Preference A new Chat width preference (Preferences → Appearance, or the command palette's Accessibility menu) widens the session chat and composer on large screens, so code blocks and tables need less horizontal scrolling. Interrupting Devin Keeps the Partial Reply When you send a message while Devin is still writing a reply, the part of the reply you already saw stays in the chat instead of disappearing. Virtual Environment Picker Uses Your Starred Default New sessions use your starred virtual environment default unless you pick another one in the composer; the picker resets your selection across reloads. Session Links in Devin's Messages Render as Chips Links to another Devin s

v2.9.0

2026-10-13T19:44:47+00:00

Typed composer mention search

Canary release (mcp-use@2.9.0-canary.0) adding typed search for app-owned items in supporting ChatGPT desktop composers. Stable v2.9.0 has not shipped.

New (server): server.mentions({ name, search }) registers a normal typed ToolRef for searching people, files, channels, and other app-owned items without a View or iframe handshake (#2792)
New (server): generate a required string query input schema and an { items: ResourceLink[] } output schema using the official SDK resource-link schema; export ResourceLink, MentionSearchParams, MentionSearchResult, and MentionsRegistration from mcp-use (#2792)
Enhancement (server): mention tools advertise mentions/search extension metadata and always include app visibility, preserving unrelated descriptor metadata; search is annotated read-only by default (#2792)
Enhancement (server): search callbacks receive ordinary request context, OAuth sign-in defaults, and ctx.signal; empty and whitespace queries pass through unchanged. Applications control authorization, ranking, and result limits (#2792)
Documentation: add composer mention registration, result contracts, visibility behavior, and authorization guidance to ChatGPT Plugin Extensions (#2792)
Update (@mcp-use/cli): rebuild bundled workspace code and synchronize published internal package metadata
Update: Updated @mcp-use/cli to v4.3.3 (canary.0)
Update: Updated @mcp-use/inspector to v20.3.19 (canary.0); see the Inspector changelog

Agent Zero v2.14: Faster Startup, Stable Prompt Caching, & Smoother Chats

2026-10-13T16:29:24+00:00

v2.14 speeds up startup, chat switching, and Browser tools, preserves long-conversation prompt caches, and improves presets, delegation, OAuth, and plugins.

October 12, 2026

2026-10-12T00:00:00+00:00

Summary

Agent Armor is now in AgentMail Labs. An organization admin can turn it on from the console, and every message your inboxes receive is scanned for attempts to manipulate an AI agent.

What's new?

Agent Armor in Labs: after joining AgentMail Labs, an organization owner or admin turns on Agent Armor under Settings, then Labs. It is free while it is in Labs.
Verdicts on messages: a scanned message carries an agent_armor result. A complete result has a verdict of clear or review, with a short reason.
message.security.completed: a webhook and WebSocket event sent when a scan reaches a verdict.
Feedback:POST /v0/inboxes/{inbox_id}/messages/{message_id}/agent-armor/feedback records whether a verdict was right.
Never in the way: Agent Armor never holds, delays or changes mail.
Availability: US production (api.agentmail.to), not yet EU production (api.agentmail.eu).

Use cases

Keep instructions hidden in inbound email from reaching your agent's tools unchecked.
Route suspicious messages to a person before your agent replies.

v2.8.2

2026-10-09T08:42:29+00:00

Tunnel cancellation & API reference navigation

Patch release keeping development tunnels connected when a cancellation arrives after its request has finished, with clearer API reference navigation.

Fix (@mcp-use/tunnel, @mcp-use/cli): ignore valid late or duplicate request cancellations so they no longer close the tunnel or interrupt unrelated requests; active requests still cancel normally (#2664)
Documentation: improve generated API reference navigation and clarify client and React hook documentation (#2791)
Update: Updated @mcp-use/tunnel to v0.2.2 and @mcp-use/cli to v4.3.2
Update: Updated @mcp-use/inspector to v20.3.18; see the Inspector changelog

Devin updates — October 9, 2026

2026-10-09T00:00:00+00:00

Memory and dreaming Devin now remembers what it learns about you and your work across sessions. Memory is personal to you and not shared with your organization; you can review its files and turn it on or off from Settings Customization Memory. Each night Devin reviews your recent sessions and refines what it remembers under Recent dreaming sessions in the Memory tab. Org admins can turn memory on or off for their organization, and enterprise admins can force it on or off for every organization. Session resume window is now 14 days Sessions can be resumed for 14 days after they go to sleep (previously 30 days). Session history remains available after that; starting new work from an older sess

v4.1.0

2026-10-08T22:57:20+00:00

v4.1.0: Fourth and One

FastMCP 4.1.0 is a maintenance release at heart: a broad sweep of fixes across OpenAPI, proxies, auth, tasks, and the CLI, plus full Python 3.15 support. It doesn't change how most servers behave. A few security fixes do tighten inputs that FastMCP previously accepted, though, and our versioning policy keeps breaking changes out of patch releases, so this ships as a point-one release.

Check these before upgrading:

MultiAuth now qualifies client IDs by source, so existing MultiAuth sessions and tasks start new ownership scopes. Finish pending tasks before upgrading.
Regex tool search uses Pydantic's regex engine, which doesn't support lookarounds or backreferences.
Skill files must resolve inside their configured directories.
OpenAPI path parameters reject . and .. segments.
HTTP sessions now expire after the MCP SDK's 30-minute idle default instead of never.
CodeMode requires Monty 1.1, which renames max_duration_secs to max_feed_duration_secs.

Thanks to our 48 contributors, including 21 first-time contributors!

Breaking Changes ⚠️

fix(http): preserve the SDK's default session idle timeout by @asts-top in #5229
Restore full Python 3.15 support by @Kludex in #5558

Enhancements ✨

enhancement: Configure token introspection audiences by @jlowin and @sajdakabir in #5461
fix: Make proxy HTTP header forwarding configurable by @jlowin in #5462
enhancement: Configure upload retention in memory by @jlowin and @warlocksmurf in #5463
Clarify maintainer review decisions by @jlowin in #5466
feat: persist encrypted CLI OAuth tokens by @andreivince in #5405
feat(auth): add TokenCache support to GoogleTokenVerifier and GoogleProvider by @jerive in #5366
Add gitleaks secret scanning to prek by @strawgate in #5355
Keep review design decisions with maintainers by @jlowin in #5468

Security 🔒

fix!: Qualify MultiAuth client identities by @jlowin in #5455
fix!: Validate OpenAPI path parameter segments by @jlowin in #5457
fix!: Resolve skill files within their configured directories by @jlowin and @GEONWOOHAN in #5459
fix!: Use Pydantic’s regex engine for tool search by @jlowin in #5467

Fixes 🐞

fix: Share JWKS refreshes across key lookups by @jlowin in #5456
fix: Reuse shared OpenAPI composition members by @jlowin in #5460
Resolve app tool names against the transformed catalog by @jlowin in #5464
fix: keep generator iteration inside dependency lifetime by @mikamikasuki in #5438
fix(tools): consume synchronous generators within dispatched invocation by @HootRock in #5376
fix(tools): reject results returned after timeout deadlines by @HootRock in #5379
Fix #5368: Tool.from_tool: a transform_fn that declares ctx: Context fails on eve by @nandanadileep and @chrikrah in #5371
fix(auth): preserve AuthKit issuer identity by @mikamikasuki and @paddynatte in #5452
fix(auth): advertise resource_metadata only when a route serves it by @alex-feel in #5363
fix: decode escaped JSON Pointer tokens during hydration by @HuaTNA in #5407
fix: read MCP config files as UTF-8 in the install commands by @ch-z-hc in #5359
Add regression coverage for root-context schema caching by @hulkbig in #5397
Keep the original cause when ErrorHandlingMiddleware re-raises an error unchanged by @asasemahmed in #5344
Resolve ASGIServer.http_client relative URLs against the server origin by @asasemahmed in #5341
Raise startup failures from run_server_async instead of hanging by @asasemahmed in #5340
Restrict Windows CLI state access without requiring SeSecurityPrivilege by @asasemahmed in #5312
Advertise the same list_changed capabilities over stdio as other transports by @asasemahmed in #5311
Validate responses to dict and nested-list elicitation choices by @asasemahmed in #5343
Inline enums shared by several fields in elicitation schemas by @asasemahmed in #5342
Honor the SDK ElicitResult returned by an elicitation handler by @asasemahmed in #5339
Look up call_tool output schemas across all tools/list pages by @asasemahmed in #5338
Deduplicate versions in PromptsAsTools and ResourcesAsTools list tools by @asasemahmed in #5331
Map File text formats to canonical text MIME types by @asasemahmed in #5334
Keep generated OpenAPI component names unique when a suffixed name is taken by @asasemahmed in #5337
Keep is_error when ResponseLimitingMiddleware truncates a result by @asasemahmed in #5309
Read +json OpenAPI resource responses as JSON text by @asasemahmed in #5310
Preserve provider challenge defaults through MultiAuth by @taylorwilsdon in #5454
Write generate-cli output as UTF-8 by @asasemahmed and @ch-z-hc in #5335
Strip internal visibility meta from prompt results and static resources by @asasemahmed in #5330
Register browser OAuth clients as public clients by @goran-revops in #5231
fix(client): preserve unique JSON arrays by @lzyruc in #5270
Keep filesystem reloads serialized during cancellation by @sherxlg-gif in #5184
Treat upstream HTTP errors from OpenAPI tools as HTTP status errors by @asasemahmed in #5308
chore(ty): pin the checked platform to the one CI uses (#5302) by @sclfcz in #5303
fix(tools): take partial tool names and docstrings from the wrapped function by @BlueX888 and @Ares16x16 in #5269
fix(tools): keep structured content for object schemas without explicit type by @BlueX888 in #5271
fix: return current tool schemas from cached BM25 searches by @baba9811 in #5153
fix(azure): send prompt=select_account to Entra instead of the consent URL by @mayankkumar-cmd in #5162
fix(auth): preserve upstream refresh error semantics by @77cjj in #5011
Clean up run_server_async when startup exits early by @syf2211 and @asasemahmed in #5469
Pass resource security policy through FileSystemProvider discovery by @asasemahmed in #5498
Forward annotations on proxied resources and resource templates by @asasemahmed in #5499
Forward the resources/read result meta through proxies by @asasemahmed in #5500
Fix proxy dropping log messages whose data is not a msg/extra dict by @asasemahmed in #5501
Forward stop_sequences in the proxy's sampling relay by @asasemahmed in #5502
Treat non-Bearer Authorization schemes as missing authentication by @asasemahmed in #5503
Reject positional-only parameters in prompts and resource templates by @asasemahmed in #5505
Carry domain and prefers_border onto the Prefab renderer resource by @asasemahmed in #5611
Keep GenerativeUI CSP on the renderer resource only by @asasemahmed in #5612
Keep OpenAPI path segments that contain a colon by @asasemahmed in #5613
Print install mcp-json and inspect JSON without Rich wrapping by @asasemahmed in #5615
Coerce optional parameters in fastmcp call by their non-null type by @asasemahmed in #5616
Return the existing component when an ignored duplicate is added by @asasemahmed in #5617
Keep a config client's servers connected when a new() clone connects by @asasemahmed in #5618
Strip discriminator under properties named default, const, enum or examples by @asasemahmed in #5619
Keep a finished task's status when it is cancelled by @asasemahmed in #5621
Forward --skip-source to the uv run subprocess in fastmcp run by @asasemahmed in #5623
Send numeric and boolean OpenAPI raw bodies as text by @asasemahmed in #5624
Map Anthropic stop_sequence to the MCP stopSequence stop reason by @asasemahmed in #5625
Escape the server spec and name in the generate-cli header docstring by @asasemahmed in #5626
Fix context docs: request_context.meta is a dict, not an attribute object by @asasemahmed in #5631
Validate strict tool input in JSON mode by @asasemahmed in #5504
Raise when a required OpenAPI path parameter is missing by @asasemahmed in #5614
Omit writeOnly properties from OpenAPI tool output schemas by @asasemahmed in #5336
fix(tasks): keep stored arguments alive with the other per-task keys by @sclfcz in #5144
fix(auth): omit scopes from generic authorization code exchange by @liang0417 and @kiranthakkar in #5183
Reject missing and unexpected arguments at task submission by @asasemahmed in #5622
fix(auth): preserve verifier-defined challenge scopes through wrappers by @mikamikasuki in #5478
fix(server): VersionFilter get_tool/get_resource/get_prompt respect include_unversioned=False by @Ashfaqbs and @Purin1410 in #4949
Stop asserting that the GitHub remote server lacks structured content by @jlowin in #5634
Skip skills whose SKILL.md is not valid UTF-8 during directory discovery by @asasemahmed in #5497
Fix tool body validation errors at the MCP boundary by @qingshungLI and @alastair in #5485
Detect client extensions without requiring clientInfo by @alex-feel in #5196
Propagate child cancellation from gather by @dennisimoo and @lokesh0186 in #5161
fix(resources): separate URI fragments from query values by @Harsh23Kashyap in #5132
fix: preserve input requests in resource and prompt tools by @HootRock and @andreivince in #5476
fix(docstrings): don't leak Returns/Raises into description for parameterless functions by @Ashfaqbs and @pvliesdonk in #5067
fix(client): hydrate formatted scalars and positional array items by @FanouZeng-TT, @MarcHuang168, @lzyruc, and @onk3sh in #5203

Docs 📚

docs: credit contributors in v4.0.11 changelog entries by @jlowin in #5449
Clarify CodeMode execution context by @jlowin in #5458
docs: explain Inspector entrypoint initialization by @andreivince and @nick-youngblut in #5404
Docs: show post-call inspection of tool text results by @Amazinghorseli and @roli-lpci in #5385
Document FastMCP Client use cases by @andreivince in #5482
docs: show tests for MCP mounted in FastAPI by @andreivince and @DeoLeung in #5471
Document OpenTelemetry log export by @andreivince in #5483
docs: explain OpenAPI client timeouts by @andreivince and @dimitribarbot in #5470
Clarify Python projects and Cursor workspaces in install docs by @andreivince and @nick-youngblut in #5475
Remove comments from fastmcp.json docs examples by @asasemahmed in #5627
Fix mcp-json Advanced Configuration example output by @asasemahmed in #5628
Remove FASTMCP_SERVER_DEPENDENCIES from the settings reference by @asasemahmed in #5629
Fix versioning docs: get_tool takes a VersionSpec and returns None for a missing version by @asasemahmed in #5630
Fix authorization docs: AccessToken field types and built-in check count by @asasemahmed in #5632

Examples & Contrib 💡

Don't evaluate properties when registering MCPMixin methods by @asasemahmed in #5620

Dependencies 📦

chore(deps): bump the uv group across 2 directories with 2 updates by @dependabot[bot] in https://github.com/PrefectHQ/fastmcp/pull/5393

New Contributors

@mikamikasuki made their first contribution in #5438
@HootRock made their first contribution in #5376
@alex-feel made their first contribution in #5363
@HuaTNA made their first contribution in #5407
@andreivince made their first contribution in #5404
@ch-z-hc made their first contribution in #5359
@Amazinghorseli made their first contribution in #5385
@hulkbig made their first contribution in #5397
@jerive made their first contribution in #5366
@goran-revops made their first contribution in #5231
@lzyruc made their first contribution in #5270
@sherxlg-gif made their first contribution in #5184
@baba9811 made their first contribution in #5153
@mayankkumar-cmd made their first contribution in #5162
@77cjj made their first contribution in #5011
@liang0417 made their first contribution in #5183
@Ashfaqbs made their first contribution in #4949
@qingshungLI made their first contribution in #5485
@dennisimoo made their first contribution in #5161
@FanouZeng-TT made their first contribution in #5203
@GEONWOOHAN made their first contribution in #5459

Full Changelog: v4.0.11...v4.1.0

v2.8.1

2026-10-08T04:44:53+00:00

Streaming, code mode & deployment fixes

Patch release fixing streamed tool notifications, code-mode discovery, CLI deployment links and remote validation, Windows scaffolding, and Inspector tool timeouts.

Fix (server): request logging no longer buffers SSE response bodies, so tool progress and log notifications reach clients before tool completion. This applies at every logging level, including trace; finite JSON response logging is preserved (#2789)
Fix (@mcp-use/client): CodeModeConnector.listTools() returns execute_code and search_tools, exposes tool capabilities and protocol metadata, and returns empty resource, resource-template, and prompt lists (#2488)
Fix (@mcp-use/client): CodeModeConnector honors request abort signals for listing and tool calls, including cancellation while waiting for code execution or tool search (#2488)
Fix (@mcp-use/cli): deployment dashboard links use the selected organization slug and server ID for managed uploads and GitHub deployments; organizations without a slug no longer receive an invalid link (#2744)
Fix (@mcp-use/cli): mcp-use deploy validates that Git remotes are hosted on github.com, rejecting other hosts that merely contain github.com in their hostname or path (#2687)
Fix (create-mcp-use-app): --install and the install prompt run npm and pnpm through a shell on Windows so their .cmd shims work instead of failing with spawn npm ENOENT (#2696)
Fix (@mcp-use/inspector): tool calls honor saved timeout settings and cleared timeouts stay cleared; see the Inspector changelog (#2656)
Update: Updated @mcp-use/agent to v2.0.22 and @mcp-use/client to v2.4.2
Update: Updated @mcp-use/inspector to v20.3.17; see the Inspector changelog
Update: Updated @mcp-use/cli to v4.3.1
Update: Updated create-mcp-use-app to v2.0.10

October 8, 2026

2026-10-08T00:00:00+00:00

Summary

Calendar is now in AgentMail Labs, so your organization can turn it on itself instead of asking us for beta access. An organization admin joins Labs once in the console, and every inbox's calendar becomes available to your agents right away.

What's new?

AgentMail Labs: a new Labs tab in the console's organization settings. An organization owner or admin joins (or leaves) Labs for the whole organization. Other members see the setting read-only.
Calendar in Labs: joining Labs turns on the calendar API for every inbox. API keys see the change within five minutes; the console shows it immediately.
Leaving Labs: calendar events are kept, but the calendar API returns 403, new invitations are not added, and event start and end webhooks stop until you rejoin.
Availability: Labs is available in US production (api.agentmail.to), not yet in EU production (api.agentmail.eu).
Organizations that were in the calendar beta were moved into Labs, so nothing changes for them.

Use cases

Give your agents a calendar the moment you need one, without waiting on a support request.
Try new AgentMail features early, for the whole organization, from one setting.

LM Studio 0.4.26

2026-10-08T00:00:00+00:00

Support for /v1/systemone Jev (TypeSafe AI) compatible endpoint Support for /v1/decisions OpenAI compatible endpoint Coexistence with Bionic — Run both apps together; Bionic uses LM Studio for local model inference while LM Studio is running. Requires Bionic 1.1.8 and newer. Added CUDA 13 support on Windows ARM. UI polish Updated the llama.cpp runtime to 2.50.0 (upstream b11337). Bug fixes and security hardening.

Devin updates — October 7, 2026

2026-10-07T00:00:00+00:00

Notification Inbox Devin now has a notification inbox. The bell in the sidebar collects notifications about your sessions, such as when Devin finishes working or needs your attention, and opening a session marks its notifications as read. A new Settings Notifications page lets you choose which notifications you receive in the inbox and which are also pushed to your phone. On mobile web, the bell opens your notifications full-screen. Command Palette Refinements The ⌘K palette lists the same navigation pages as the sidebar, with the same names and icons, lists settings pages directly (not individual settings), and when a pull request's review is open in a session's workspace pane its actions (

v2.8.0

2026-10-06T18:52:23+00:00

ChatGPT Plugin Extensions

Minor release adding typed ChatGPT Plugin Extensions across server and React APIs, with updated MCP SDKs and richer examples.

New (server): register typed global, thread, and file entrypoints on tool.view; launch contracts are validated and OpenAI tool metadata is preserved (#2737)
New (server): register native ChatGPT plugin settings with typed fields, optional layout, validated read/update tools, request context callbacks, and per-request capability advertisement. Applications keep control of persistence and authorization (#2739)
New (server): advertise top-level tool icons in tools/list, including proxied tools, and support public-relative icon assets (#2779)
New (@mcp-use/react): useDeepLink reads the initial and subsequent ChatGPT deep-link URLs through the existing MCP App connection (#2738)
New (@mcp-use/react): useModelContext manages keyed native text, image, resource-link, and embedded-resource attachments. It activates only for committed hook consumers and handles replacement, removal, cancellation, and delivery failures without restoring stale content (#2776)
Enhancement (server): define supported MCP App display modes and an optional initial ChatGPT display preference in the named viewConfig export; validate preferences while preserving broader host capabilities (#2740)
Enhancement (@mcp-use/cli): include display-mode configuration in development and production view manifests and pre-bundle the model-context schema for cold iframe startup (#2740, #2776)
New (examples): add a Bookshop Plugin Extensions example with entrypoint icons, plugin settings, and model-context attachments (#2773)
Fix (create-mcp-use-app): failed template selection or GitHub template cloning no longer leaves an empty project directory that blocks a retry (#2765)
Update (@mcp-use/client): update the official MCP client, core, and server SDKs to 2.3.1 and MCP Apps to 2.0.3. Node follows only same-origin redirects that preserve the request method; browsers reject redirects by default, so configure the final endpoint URL (#2784)
Documentation: add the ChatGPT Plugin Extensions guide for MCP Apps and clarify browser redirect restrictions (#2772, #2784)
Update: Updated @mcp-use/agent to v2.0.21 and @mcp-use/client to v2.4.1
Update: Updated @mcp-use/inspector to v20.3.16; see the Inspector changelog
Update: Updated @mcp-use/cli to v4.3.0
Update: Updated create-mcp-use-app to v2.0.9

Devin updates — October 5, 2026

2026-10-05T00:00:00+00:00

Terminal Tab in the Session Workspace The session workspace has a new Terminal tab for running commands on Devin's machine. The Files tab can now be opened while Devin is asleep, and opening Files or Terminal wakes Devin. Keyboard Shortcuts and Command Palette The Settings → Keyboard shortcuts page is now available to everyone. Mark the current session done (or not done) with ⌘⇧E / Ctrl+Shift+E, and actions from a session's … menu can be found by typing in the command palette (⌘K). Open HTML, PDF and SVG Attachments in a New Tab HTML, PDF and SVG attachments have a new "Open in new tab" action that opens the file on a page of its own, with the filename as the tab title, instead of downloadin

v4.0.11

2026-10-04T16:58:25+00:00

v4.0.11: Fort Knocks

This release contains important security and bug fixes. All users are encouraged to upgrade. Thanks to our 10 contributors, including 6 first-time contributors.

Enhancements ✨

agents: make AGENTS.md the entry point and tighten repo skills by @zzstoatzz in #5278
ci: publish maintenance status and add MAINTAINING.md by @zzstoatzz in #5284
rate_limiting: deprecate a burst_capacity below 1 by @zzstoatzz and @sclfcz in #5296
agents: consolidate PR review guidance by @zzstoatzz in #5297
status: unify dashboard and terminal issue ranking by @zzstoatzz in #5300
Enable ruff flake8-bandit (S) rules on the library by @strawgate in #5356
Bundle extensions with providers by @jlowin in #5360
Expand supported Python CI coverage by @jlowin in #5391
docs: credit issue authors in maintainer implementations by @jlowin in #5410
enhancement: scan release notes for shared contributor credit by @jlowin in #5433

Security 🔒

Pass client CLI arguments literally to Windows .cmd wrappers by @jlowin and @strawgate in #5419
Build client schema types with create_model and bounded caches by @jlowin and @strawgate in #5413
Bound schema nesting before client type conversion by @jlowin and @strawgate in #5434
OpenAPI components send only declared arguments and keep configured headers by @jlowin and @strawgate in #5423
Use the server's auth provider for component manager routes by @jlowin and @strawgate in #5425
Apply Host/Origin protection settings to the SSE transport by @jlowin in #5427
Require a startup-URL session for the fastmcp dev apps preview by @jlowin and @strawgate in #5417
Apply transforms, enabled state, and auth to hashed tool lookups by @jlowin and @strawgate in #5415
Key cached schema adapters by the normalized schema by @jlowin in #5436
Resolve injected tools through the server's tool lookup by @jlowin in #5431
Keep $ref pointers when inlining would produce an oversized schema by @jlowin and @strawgate in #5421

Fixes 🐞

rate_limiting: default burst capacity to at least 1 by @zzstoatzz and @Patrick-SCH03 in #5293
types: map Image jpg/svg/tif formats to canonical MIME types by @zzstoatzz and @Patrick-SCH03 in #5294
client: decode empty structured content into CallToolResult.data by @zzstoatzz and @Patrick-SCH03 in #5291
fix: use portable issue ranking cache filenames by @zzstoatzz in #5314
fix: validate cached version timestamp types by @Harsh23Kashyap in #5351
cli: read MCP and fastmcp.json config files as utf-8 by @zzstoatzz and @asasemahmed in #5345
Support Python 3.15 by @Kludex in #5374
Revert "Support Python 3.15" by @jlowin in #5389
examples: handle malformed media links and concurrent Fire TV setup by @jlowin in #5390
Fix mounted extension runtime tracking by @jlowin in #5392
Fall back to pure Python YAML loader in schema tests by @jlowin in #5395
tools: keep explicit Field(title=...) on parameters in the input schema by @zzstoatzz and @TJReinert in #5388
Run Windows tests on code pull requests by @jlowin in #5408
Avoid repeated traversal of shared tool schema definitions by @jlowin in #5412
Keep referenced body fields when a request body schema adds sibling properties by @jlowin in #5440
Require pydocket 0.26.0 so a cancelled task cannot stop the worker by @jlowin in #5444

Docs 📚

docs: consolidate developer and maintainer guidance by @zzstoatzz in #5298
examples: add interactive media picker MCP app by @zzstoatzz in #5281
docs: route SDK v1 users to the correct upgrade guide by @aritejhg in #5372
Clarify app visibility and configurable security guarantees by @jlowin in #5409
Clarify declared header parameters in the OpenAPI docs by @jlowin and @strawgate in #5437
docs: add v4.0.11 changelog entries by @jlowin in #5446

Dependencies 📦

chore(deps): bump the uv group across 2 directories with 2 updates by @dependabot[bot] in https://github.com/PrefectHQ/fastmcp/pull/5361

Other Changes 🦾

chore(deps): Update astral-sh/setup-uv action to v10.2.0 by @prefect-renovate[bot] in https://github.com/PrefectHQ/fastmcp/pull/5347

New Contributors

@Harsh23Kashyap made their first contribution in #5351
@aritejhg made their first contribution in #5372
@Patrick-SCH03 made their first contribution in #5291
@sclfcz made their first contribution in #5296
@asasemahmed made their first contribution in #5345
@TJReinert made their first contribution in #5388

Full Changelog: v4.0.10...v4.0.11

Desktop app improvements, agent comment replies, and more

2026-10-02T19:58:56+00:00

Desktop app

Deep links: mintlify:// links open in the desktop app as a new tab.
Zoom and new tabs: Zoom in, zoom out, or reset the zoom from the View menu. Cmd + N opens a new tab with a blank page.
Offline screen: Navigation changes you made while offline save when your connection returns. The app shows an offline screen if you lose your connection.
Find in page: Press Cmd + F to search the current page. See Keyboard shortcuts for more information.
Check for updates: Check for a new version from the app menu or settings. See Updates for more information.
Browser sign-in: Sign in to the desktop app through your default browser. If browser sign-in isn't available for your organization, sign in inside the app. See Sign in for more information.

Agent and AI

Agent replies to comments: The agent can read open comments and suggestions on a page, reply to threads, and resolve comment threads.
Google Docs integration: Connect Google Docs as a read-only personal integration so the agent can search and read your documents. Available on Enterprise plans.
Admin MCP: The admin MCP can edit your deploy branch directly when the branch doesn't require pull requests. It can also upload images and returns links to private pages. The diff shows who made each pending change.
Group-scoped MCP credentials: Assign user groups to MCP client credentials so they can access pages restricted to those groups. You can't change a credential's groups after you create it.
Custom skills on authenticated sites: Skills marked public: true appear for visitors who aren't signed in. See Share skills on authenticated sites.
Multiple pull requests per automation run: Custom automations can open several pull requests in one run, or add commits to a pull request that an earlier run opened. See Open multiple pull requests or append to open ones.

Editor and dashboard

Images in the editor: Click an image to select it. Drag it to move it, even inside frames, cards, tabs, and callouts, and resize it with handles that also work on touch screens. Dropped and pasted media lands between blocks.
Emoji reactions on text: Highlight text and react to it with an emoji. See Comments.
Publish warnings: The Publish menu warns you about open comments and labels unlisted pages as hidden.
Fewer confirmations: Source mode asks you to confirm a large change only when it exceeds 10,000 characters.
Keyboard shortcuts in the editor: Open Keyboard shortcuts from the dashboard help menu to see every shortcut. See View shortcuts in the dashboard for more information.
Starting page: Choose which dashboard page opens first. See Starting page for more information.
Preview access: Switch an existing preview between dashboard login and password protection. When you turn on authentication, choose whether to keep existing previews, delete public previews, or delete all previews. See Restrict access to preview deployments for more information.

Published sites and deployments

Page modes in every theme: Every theme supports center and frame page modes. In Maple and Almond, center mode keeps the sidebar and centers the content beside it. See Page mode for more information.
Endpoint frontmatter validation: Deployments fail if a page's api, openapi, or asyncapi frontmatter is a list instead of a single endpoint string, such as "GET /users".

Bug fixes

Editor: Ordered lists keep their numbering when you delete an empty middle item. Any block now renders between accordions in an AccordionGroup. Images pasted from other sites stay on the page. Source mode parses MDX import and export blocks correctly.
Private pages: Moving a private page to your site keeps its page ID and comments. Private pages load regardless of Git branch state.
Published sites: Images in table cells are zoomable. API playground requests without a body no longer send a Content-Type header. Search breadcrumbs match the page language. Willow renders custom page mode.
Assistant: The assistant widget on sites without authentication searches every page, not only pages marked public: true.
Authenticated sites: Generated OG images and PDF exports work on sites with authentication.
Previews: GitLab multi-repo preview deployments work, failed merge request comments no longer fail preview builds, and retargeting a pull request to your deploy branch creates a preview.
Agent: The editor agent uses your timezone.
Analytics: CSV exports escape spreadsheet formulas.
CLI: The CLI accepts AsyncAPI 3.0.0 specs on fresh installs.

Devin updates — October 2, 2026

2026-10-02T00:00:00+00:00

Sidebar Search and Filters Sidebar search now shows matching sub-sessions and workflow sessions even when their parent is collapsed; the Updated-time filter no longer defaults to the last month, and the Filters menu uses the available screen height. The session status no longer keeps asking you to review, approve or merge a PR once that step is done. Transcript Polish Edit cards show the changed lines even when the edit is far down the file and stream the newest lines while an edit is being written; chained shell commands keep their per-step rows while output streams and when a command fails; large scripted-tool and subagent runs expand instantly; and closing a screenshot viewer keeps your t

Metric charts for alert incidents

2026-10-01T00:00:00+00:00

When you review past alert incidents, you can expand each occurrence to see the measurements and thresholds recorded at the time, with charts covering the period before, during, and after the incident.

Manufact for Startups application program

2026-10-01T00:00:00+00:00

Eligible early-stage teams can apply for Startup plan access through Manufact for Startups on the marketing site.

User and latency insights in Analytics

2026-10-01T00:00:00+00:00

The Analytics tab now ranks active users, adds a dedicated tool-latency view with session drill-down, and shows period-over-period trends on the summary KPIs.

Public MCP quality reports

2026-10-01T00:00:00+00:00

The MCP catalog now publishes quality reports with an overall score, tool documentation feedback, basic publishing checks, and a score leaderboard.

Devin updates — September 30, 2026

2026-09-30T00:00:00+00:00

Devin Mono for Code and Terminals Code blocks and terminals in the app now render in Devin Mono, a new monospace typeface designed to pair with the app's UI font. Shell output in the Progress tab uses the same character spacing as the rest of the app. Read-Only Live Shell Preview The streaming terminal preview in the worklog is now a read-only text view: clicking it no longer places a cursor or captures keyboard input, text can be selected and copied normally, and multi-line progress bars redraw correctly. Messages Are No Longer Lost When the Connection Drops Messages sent to Devin from the browser, Devin Desktop or mobile right before the connection drops (closing the tab, backgrounding the

The Official Ruby SDK for MCP Reaches Tier 1

2026-09-29T09:00:00+00:00

The official Ruby SDK for the Model Context Protocol is now a Tier 1 SDK under the MCP SDK tiering system, with full support for the 2026-07-28 specification revision and a 100% conformance pass rate across both scored spec revisions.

Cursor Pricing 2026: Plans, Usage Models & Which to Choose

2026-09-29T00:00:00+00:00

Deep dive into Cursor pricing for AI agents, cost models by seat and usage, and how Mem0 reduces prompt tokens with persistent memory.

Testing memory placement against compaction cliff

2026-09-26T00:00:00+00:00

What happens to an agent's project rules when the context window fills, and whether keeping them somewhere else helps?

Desktop app, page narration, webhook-triggered automations, and more

2026-09-25T17:56:03+00:00

Desktop app

Desktop app (beta): Use your dashboard and editor in a native app for macOS, Windows, and Linux. Work in multiple tabs and windows, each with its own session. See Desktop app for more information and to download the app.

Editor

MDX checks in source mode: Source mode flags misspelled components, unknown or duplicate props, invalid prop values, missing required props, and unclosed tags. Hover over a flagged problem to apply a quick fix.
Autocomplete and hover definitions in source mode: Type < to get suggestions for components, props, and values, including page links, image paths, and icons. Hover over a component or prop to see its definition. See Visual and source mode.
Formatting in suggesting mode: In suggesting mode, bold, italic, underline, strikethrough, and inline code changes become suggestions that teammates can accept or reject. See Suggestions.
Granular publishing: The Selective publishing setting is now called Granular publishing. By default, the Publish menu shows and publishes only your own changes. Turn on granular publishing to also choose from changes your teammates made. See Granular publishing.
List and table shortcuts: Cmd + Shift + 7 creates a numbered list and Cmd + Shift + 9 creates a task list. Use Ctrl instead of Cmd on Windows and Linux. Type [] and a space to start a task list. Tab and Shift + Tab indent and outdent every selected list item. In tables, Tab moves to the next cell and adds a row from the last cell. See Keyboard shortcuts.
Copy and paste: Copying keeps the structure, numbering, and formatting of lists. Pasting from Google Docs keeps bold text and links in lists, and Google Docs checklists and GitHub task lists paste as task lists.
Comments: Unsent comment drafts survive page reloads, replies stay put when a thread collapses, and a comment can contain only a mention. See Comments.
Private page roles: The Publisher role for sharing private pages is now called Full Access. Permissions are unchanged. See Share a private page.
Files: The Hidden files section is now called Files and lists every file in your repository that isn't in your navigation. See Files.
Faster editor: Pages open faster, typing no longer reparses the whole page, and navigation loads without waiting on GitHub data. When you switch projects, the editor reopens the page you last had open.

Agent and assistant

Image reading: The editor agent can view JPEG, PNG, GIF, and WebP images on your pages to describe them, write alt text, or check screenshots against the surrounding content. The assistant can also inspect images on the pages it reads. See View images.
Agent panel: The editor agent panel has a new design, is faster in long chats, folds long messages, and lets you start typing from anywhere in the panel.
Links to new pages: When the agent creates pages, its final message links to each one.
Assistant exports: Assistant CSV exports list the sources each answer cited.

Automations

Webhook triggers: Start a custom automation with an authenticated POST request to its webhook URL, using an organization API key with write access. Each request runs the automation's saved prompt. See Trigger a webhook automation.
Multiple pull requests per run: A custom automation can open several pull requests in one run. When its prompt asks, it can also add changes to open pull requests from its earlier runs.
Managing automations: Editors can now delete automations and triggers link to their repositories and channels.

Published sites

Narration (beta): Readers can listen to a page from a Listen to this page player at the top of the page. Mintlify highlights each sentence as it plays and skips code blocks, tables, and other non-prose content. Turn on narration in your project settings. Each generation uses 50 credits. Available on Pro and Enterprise plans. See Narration.
WebMCP: Mintlify registers your site's MCP tools in the browser, so browsers and in-browser agents that support WebMCP can search and read your docs while someone visits your site. No setup required. Available on Pro and Enterprise plans. See WebMCP.
Folder-level metatags: Set metatags for every page in a folder with seo.paths in docs.json. Page frontmatter now overrides metatags set in docs.json. See SEO settings.
MCP on authenticated sites: Hosted MCP servers no longer return 403 errors on sites with authentication enabled, and MCP settings appear for every project that uses authentication. See Enable authenticated MCP.

Dashboard and deployments

Projects: The dashboard now calls your deployments Projects instead of Sites.
Workspace name: Change your workspace name in your organization settings. Your organization's slug stays the same.
Rollbacks for editors: Editors can now roll back a deployment, not just admins. Available on Enterprise plans.
Faster multi-repo updates: Deployments with multiple repositories rebuild only what changed instead of running a full update.
mint validate: The command no longer copies static files, so it uses less disk space in repositories with many images.

Bug fixes

Tables and lists: HTML tables no longer lose rows on save, ordered lists in table cells keep their starting number, and task lists written with [X] no longer fail MDX validation.
Editor: New page titles keep focus while you type, pasted dollar amounts no longer turn into math, image paste works on new private pages, and the editor shows clearer errors when a repository was renamed, can't be accessed, or is blocked by a GitHub IP allowlist.
Publishing: Publishing works around navigation errors instead of blocking, the editor rejects publish selections and docs.json edits that would drop navigation entries, and the publish button no longer gets stuck.
Published sites: Related pages render in preview deployments, locale redirects work behind proxies and custom base paths, breadcrumb groups without a root page render as text instead of a broken link, shell placeholders highlight correctly, and SDK type references link to their generated pages.
Search: Results from other versions show their version breadcrumb again.
Deployments: Mintlify removes stale OpenAPI pages when a spec changes.
Invites: Invite links work for users who are already logged in.

v4.0.10

2026-09-25T17:20:12+00:00

v4.0.10: Inside Job

Task-enabled tools now work behind search transforms and CodeMode: they're registered with the task backend even when hidden, and a tool, resource, or prompt that calls one through ctx.fastmcp.call_tool() (including the search call_tool proxy and CodeMode's execute) now gets its result instead of an empty task receipt.

Fixes 🐞

fix(stdio): tolerate a transport whose construction failed in __del__ by @kbkb628 in #5256
tasks: register task tools hidden by search and CodeMode by @zzstoatzz in #5262
tasks: run tools called from another tool in the foreground by @zzstoatzz in #5275

Docs 📚

Add Bluesky to docs footer by @zzstoatzz in #5274

New Contributors

@kbkb628 made their first contribution in #5256

Full Changelog: v4.0.9...v4.0.10

Cognition Crosses $1B in Annualized Revenue Run Rate

2026-09-25T00:00:00+00:00

A milestone that belongs to our customers.

v4.0.9

2026-09-24T14:35:40+00:00

v4.0.9: Cache and Release

ResourceTemplate now keeps its compiled URI pattern for its own lifetime while the shared cache is bounded again, preventing dynamic proxies from growing process memory without restoring the 4,096-template performance cliff.

Fixes 🐞

resources: keep each template's compiled pattern on the template by @zzstoatzz in #5253

Full Changelog: v4.0.8...v4.0.9

Agent Zero v2.13: File Browser, Project Folders, & Slash Commands

2026-09-23T18:47:44+00:00

v2.13 redesigns file navigation, adds project folders and messaging slash commands, streams native commentary, and speeds up settings, model, tool, and Codex paths.

AIHubMix updates — September 23, 2026

2026-09-23T00:00:00+00:00

New models Added claude-opus-5-5 : Anthropic’s flagship model for agentic coding, large codebases, computer use, and complex work. Added gpt-6-sol : OpenAI’s model for complex coding, agent workflows, and long-context reasoning. It supports text and image input, a 1.05M-token context window, and up to 128K output. Added gpt-6-luna : OpenAI’s efficient model for focused, high-volume tasks. It supports text and image input, a 1.05M-token context window, and up to 128K output.

Introducing DolphinBench: Mapping the Pareto Frontier of Agent Memory

2026-09-22T00:00:00+00:00

DolphinBench evaluates agent memory through task completion. Three personas, roughly 500k tokens of history each, 600 action tests, with accuracy, cost, and latency on one leaderboard.

Building the Future of Software Engineering in Latin America

2026-09-22T00:00:00+00:00

Cognition is expanding into Latin America, starting in São Paulo, building on years of work with the region’s largest banks, consumer leaders, and technology companies.

AIHubMix updates — September 22, 2026

2026-09-22T00:00:00+00:00

New models Added mimo-v2.6-flash : Xiaomi’s natively multimodal efficient model for frequent calls and large-scale professional tasks. Added mimo-v2.6-pro : Xiaomi’s natively multimodal flagship reasoning model for complex projects, coding, research, and long-running agent tasks. It supports a 1M-token context window. Added mimo-v2.6-pro-ultraspeed : Xiaomi’s speed-focused Pro variant, delivering up to 20 times faster output for latency-sensitive use cases. Added grok-4.7 : xAI’s flagship model for coding, agentic tasks, and professional knowledge work, with stronger self-verification and long-context handling.

AIHubMix updates — September 20, 2026

2026-09-20T00:00:00+00:00

New models Added step-5-preview : StepFun’s next-generation flagship preview for real-world tasks, especially programming and professional knowledge work.

LM Studio 0.4.25

2026-09-19T00:00:00+00:00

[Mac] Added support for the Splash engine on Apple Silicon M3 or newer running macOS 26.4 or newer.

Scoped automation triggers, editor agent integrations, and more

2026-09-18T21:10:59+00:00

Automations

Scoped GitHub triggers: Custom automations can now trigger on GitHub topics and filter to paths and specific authors.
GitHub topic search: When you configure a repository-topic trigger, Mintlify suggests topics from your organization's repositories so you can pick one without leaving the settings page.
Automatic merge conflict resolution: Automations now resolve merge conflicts on their own pull requests without needing you to open the PR first. See Resolve merge conflicts.
Automations menu: A new Automations menu in the dashboard topbar links to your automations, runs, and settings. Sidebar automation cards show a live count of runs that need review.
Runs tab improvements: The Runs tab now shows a preview deployment link and lets you open a run in the editor, view run details, or edit the automation's settings.

Editor agent

Connected integrations as tools: The agent can call your connected third-party apps while researching and editing. The agent can prompt you to connect a new integration when it needs one.
Roll back edits: Ask the agent to undo a change and it reverts the pages it edited in the current session.
Current page context: The agent now sees the page you have open and its unsaved edits, so requests like "fix this section" apply to what you are working on.

Editor collaboration

@mentions: Type @ in a page to mention a teammate. Mentioned teammates get a notification.
Profile picture uploads: Upload a profile picture from your account settings. It appears on comments, mentions, and collaborator avatars.
Workspace folder icons: Edit the icon on a workspace folder from its menu. Folder metadata is preserved when you move or rename it.
Emoji favicons: A new Emojis tab in the icon picker lets you set an emoji as a page icon, and emoji icons render as the editor tab favicon.
Reorder shared and starred pages: Drag pages in Shared with me and Starred to set the order you want. The order syncs across your devices.
Version history redesign: The version history panel has a refreshed layout that's easier to scan when comparing revisions.
Faster editor rendering: Visual mode renders, switches pages, and scrolls faster.

Dashboard

Rebuilt Activity page: The Activity page has a new layout and shows richer details.
SSO login route: A new login route detects SSO configuration for your email domain and sends you to the right identity provider. The SSO connection limit is raised to 25.
Assistant layout settings: Deployment settings include options for the assistant layout so you can control how the assistant appears on your published site.

Analytics

Sort and export improvements: Page ratings support sort order, feedback shows the page URL, and assistant analytics exports include feedback ratings.
Page filter polish: The page filter opens on the currently selected page, resets with the clear button, and shows All pages as the default label.
Page-scoped popular pages, referrals, and top agents: These analytics widgets accept a page filter so you can drill into a single page.

Published sites

Related pages on authenticated sites: Related pages now appear on sites protected by Mintlify authentication.
Copy assistant responses: Select text and copy full responses from the assistant.

Bug fixes

Editor publishing: Pending edits are saved before publishing, deleted page titles show in the publish panel, publish retries when the branch lock is held, GitHub same-path replaces succeed, GitLab create collisions no longer break publish, and rows the server keeps after publish are preserved.
Editor: Language switcher options are disabled for pages without a translation, new page paths update until the first commit, slash commands rank by relevance, frontmatter edits are preserved during Git sync, source-mode comments no longer show false conflicts, and the page title is preserved across navigation.
Automations: Folder selection and topic search work correctly in the create flow, automatic conflict resolution skips mandatory PR mode when needed, and the trigger options popover collapses cleanly.
OpenAPI and API playground: Reverse-proxy guides render for sites with a base path, schema validation errors stay on one line, and locally-ignored OpenAPI specs are marked stale on full updates.
Client and hosting: RSS URLs no longer duplicate base paths, custom and frame pages render without a layout shift on load, and color swatches pick their theme value in CSS.

Coding Agents Explained: What They Are and How They Differ from AI Assistants

2026-09-18T00:00:00+00:00

Learn what coding agents are, how they work, and how they differ from AI assistants, chatbots, and autocomplete tools.

AIHubMix updates — September 18, 2026

2026-09-18T00:00:00+00:00

New models Added jev-latest : TypeSafe’s structured decision model for fast, typed choices such as routing and classification. Added qwen3.8-omni-flash : Alibaba Qwen’s native multimodal model for audio/video analysis and agents. It accepts text, image, audio, and video input with a 1M-token context window. Added glm-5.3-flashx : Z.AI’s high-speed model for coding agents and real-time workflows, with a 1M-token context window and image and video understanding.

Splash Engine - the fastest local Qwen3.8 on Apple Silicon

2026-09-18T00:00:00+00:00

How to use Splash by Inco AI in LM Studio Bionic

September 17, 2026

2026-09-17T13:46:04+00:00

Tool Call Names are Stabilized

The Tool Call Name RFD has moved to Completed, and optional programmatic names for tool calls are now stabilized in ACP v1 and the v2 baseline.

Agents can identify the invoked tool separately from its human-readable title without introducing capability or authorization semantics.

Read the full announcement.

OpenAI Codex vs Claude Code: Which AI Coding Agent Wins in 2026?

2026-09-17T00:00:00+00:00

Codex leans toward parallel execution, isolated worktrees, cloud delegation, and reviewable task runs. Claude Code is terminal oriented and more engaging in long interactive sessions, and agent-to-agent coordination. OpenAIâs GPT-6 Astra and Anthropicâs Fable 5.1 have enhanced capabilities, but the harness determines how that intelligence is used. For serious repository work, the better choice is the one that performs more reliably on your codebase, constraints, and working style.

AIHubMix updates — September 17, 2026

2026-09-17T00:00:00+00:00

New models Added union-alpha-free : a free multimodal model for evaluation, research, coding, and agent workflows. It accepts text and image input with a 262K-token context window.

Session References and Introspection

2026-09-17T00:00:00+00:00

How Bionic can search persisted session transcripts to recover details across compactions and sessions.

Mem0 joins Vercel Marketplace

2026-09-16T00:00:00+00:00

Mem0 is now available to use in the Vercel Marketplace

Cognition and AWS team up to help ambitious teams ship more, faster

2026-09-15T00:00:00+00:00

Cognition and AWS have entered a multi-year Strategic Collaboration Agreement to help enterprises deploy autonomous engineers in production, move legacy workloads to AWS faster, and clear engineering backlogs.

AIHubMix updates — September 12, 2026

2026-09-12T00:00:00+00:00

New models Added qwen-3.8-27b : Alibaba’s open-source dense vision-language model with native understanding of text, images, and video. Built on the Qwen3.5 architecture, it improves coding, professional and research tasks, and long-running complex workflows. It supports a 131K-token context window and flexible cognitive control.

Introducing Fusion in Devin Desktop & CLI

2026-09-11T00:00:00+00:00

Fusion is the most efficient frontier harness for Fable and Astra, up to 39% more efficient compared to other model harnesses across major coding benchmarks. Today, we’re making it available in Devin Desktop and CLI.

Welcoming Dioxus to Cognition

2026-09-10T00:00:00+00:00

Jonathan Kelley and the Dioxus team are joining Cognition. We will continue supporting Dioxus, Blitz, Taffy, and Subsecond while increasing investment in Dioxus-Native and Blitz.

Introducing SWE-2: Pushing the Pareto Frontier

2026-09-10T00:00:00+00:00

Today we’re introducing SWE-2, our most advanced coding model yet. SWE-2 delivers highly competitive agentic coding performance across multiple effort levels, pushing the Pareto frontier of capability and inference cost.

Agent Zero v2.12: Remote Files, Native Tool Calls, & Security Hardening

2026-09-09T13:29:23+00:00

v2.12 hardens Connector and file transfers, unifies Files and Editor, preserves native Responses tool calls, improves prompts and providers, and fixes security issues across messaging plugins.

AIHubMix updates — September 9, 2026

2026-09-09T00:00:00+00:00

New models Added gpt-image-2.5-flare : OpenAI’s image model for fast, everyday high-quality image generation. It accepts text and image input, produces image output, and supports quality settings low, medium, high, xhigh, max, and auto. Added gpt-image-2.5-sunburst : OpenAI’s image generation and editing model for workflows requiring high editing precision. It accepts text and image input, produces image output, and supports quality settings low, medium, high, xhigh, max, and auto.

LM Studio 0.4.24

2026-09-09T00:00:00+00:00

[LM Studio Engine Protocol] Added advanced llama.cpp argument overrides for GGUF model loading Improve DSpark and DFlash assistant drafters compatibility heuristics Fixed the Loaded Instances UI showing an incorrect context length Fixed /api/v1/chat splitting text and image inputs into separate user messages Improved sign-in reliability

Factoring RSA-260

2026-09-09T00:00:00+00:00

How Devin and a Cognition researcher built the world’s highest-performance GPU lattice siever, to make factoring numbers 10x cheaper than the previous state of the art

Saying "Done" Isn’t Enough: How Lobe Eval Tests Agents on Real-World Tasks

2026-09-08T00:00:00+00:00

Lobe Eval runs agents through LobeHub’s actual product execution path in a stateful, resettable tool environment. Each run is judged using execution traces, tool results, and artifact evidence, while Pass³ measures how often a Case passes in all three runs.

AIHubMix updates — September 8, 2026

2026-09-08T00:00:00+00:00

New models Added deepseek-v4.1-flash : DeepSeek’s internal-beta model built on a new architecture with native multimodal support. It accepts text and image input, produces text output, and supports reasoning, tool calling, and structured outputs. Trial access is rate-limited and intended for testing only, and is not recommended for production.

Do it all with Devin: Announcing our Series E

2026-09-08T00:00:00+00:00

Cognition has raised over $2B at a $48B valuation, led by Andreessen Horowitz and Accel, to build the future of software engineering.

AIHubMix updates — September 7, 2026

2026-09-07T00:00:00+00:00

Model badges are more consistent The model catalog, model detail pages, and ?tag=new lists now derive the New badge from model launch time. New expires automatically after the configured window, and the old Hot badge is no longer shown. Recently launched parent models also keep their New badge when supplier self-binding data exists. GPT tool-call compatibility is improved gpt≥5.5 models called through /v1/chat/completions with tools now use the Responses compatibility path automatically, reducing upstream 400 errors. Claude Messages bridge calls for these GPT tool workflows follow the same compatibility path. OpenAI and Azure direct channels filter incompatible reasoning parameters before se

LM Studio 0.4.22

2026-08-28T00:00:00+00:00

Support for DFlash, DSpark, and MTP assistant drafters Requires llama.cpp engine version 2.29.1 or newer Requires llama.cpp engine version 2.29.1 or newer Added support for images returned by tools in OpenAI-compatible /v1/responses and /v1/chat/completions , and Anthropic-compatible /v1/messages [Linux] Fixed AppImage startup on systems without FUSE 2 [Linux] Fixed AppImage startup when Chromium sandboxing is unavailable [Linux] Fixed lms wake-up for AppImage installations [Linux] Fixed desktop icons for .deb installations [Linux] Enabled in-app updates for AppImage and .deb installations [Linux] Improved recovery after canceling .deb update authorization [Linux] Fixed keyboard shortcuts be

LM Studio 0.4.23

2026-08-28T00:00:00+00:00

Improve reliability of Qwen 3.8 Flash Next

Agent Zero v2.11: Context Doctor, Vision Sidecars, & Faster Streaming

2026-08-27T14:30:15+00:00

v2.11 adds Context Doctor, vision sidecars and a context usage indicator, makes streaming and WebUI synchronization faster, improves memory and embeddings, and defaults LiteLLM providers to Chat Completions.

How Auto Review works in Bionic

2026-08-27T00:00:00+00:00

AST parsing, capability extraction, and command matching for safe shell command execution in Bionic's Auto Review mode.

DeepSeek Harness: Mem0 as a Native Cordis Plugin

2026-08-24T00:00:00+00:00

DeepSeek Harness plugin

The New MCP Roadmap

2026-08-22T09:00:00+00:00

An update on the Model Context Protocol roadmap and focus areas for upcoming specification releases.

Agent Zero v2.10: Interactive Browser, Durable Subordinates, & Composer Context

2026-08-19T12:30:34+00:00

v2.10 introduces an interactive shared Browser runtime, durable subordinate lifecycles, scoped composer context and drafts, Responses streaming fixes, and a built-in ACP bridge.

Agent Zero v3: A Sneak Peek Inside the Sandbox

2026-08-17T11:57:41+00:00

Agent Zero's third major iteration is moving agent-run code and heavyweight tools into isolated, disposable gVisor sandboxes.

Bionic now supports skills

2026-08-17T00:00:00+00:00

Teach Bionic repeatable actions with Skills.

Kimi Code: Mem0 Joins the Editor Plugin Family

2026-08-13T00:00:00+00:00

Kimi Code plugin

Agent Zero v2.9: Agent Editor, Scoped Policies, & Security Hardening

2026-08-12T13:58:33+00:00

v2.9 introduces the Agent Editor, project-scoped profiles and tool policies, safer profile lifecycles, Time Travel retention, built-in plugin upgrades, and important security and reliability fixes.

LM Studio 0.4.21

2026-08-12T00:00:00+00:00

Improved llama.cpp model-load error messages Update advanced load settings for mmap, mlock, and direct io for llama.cpp engines 2.28.1 and newer. Support for local server API key in enterprise internal network endpoint mode

Run Muse Glimmer locally

2026-08-10T00:00:00+00:00

Meta's new 30B agentic open-source model is available LM Studio Bionic.

Work in Progress: Spynel — One Conversation, Many Agents at Work

2026-08-09T15:27:04+00:00

Jan Tomášek previews Spynel, a new orchestration layer that turns established AI coding harnesses into one durable, review-driven workflow from desk or phone.

Create slides and documents using Kimi K3 in Bionic

2026-08-04T00:00:00+00:00

See how Kimi K3 in LM Studio Bionic turns research into polished Word documents and PowerPoint decks.

Run DeepSeek V4 Flash locally or in the cloud (US-hosted)

2026-08-02T00:00:00+00:00

DeepSeek V4 Flash is now available locally and in LM Studio Bionic, hosted on US-based servers with Zero Data Retention by default.

Workflow Automation: Mem0 Memory in n8n and Zapier

2026-07-30T00:00:00+00:00

n8n and Zapier integrations

The 2026-07-28 Specification

2026-07-28T09:00:00+00:00

The 2026-07-28 Model Context Protocol specification is out, bringing a stateless protocol core, Multi Round-Trip Requests, header-based routing, cacheable list results, authorization hardening, a formal extensions framework, and updated Tier 1 SDKs.

LTM Partners with Cognition To Reduce Cyber Risk in Financial Services

2026-07-28T00:00:00+00:00

LTM has partnered with Cognition to deploy Devin, the AI software engineer, across its global client base and cybersecurity practice serving over 260 clients, including 26 of the Fortune 500 and the top 5 global banks.

The Official Ruby SDK for MCP Reaches 1.0

2026-07-27T09:00:00+00:00

The official Ruby SDK for the Model Context Protocol has reached version 1.0 with a stable public API and a 100% conformance pass rate. The SDK now meets every Tier 2 requirement of the MCP SDK tiering system, and development toward Tier 1 continues.

Run Kimi K3 in LM Studio Bionic

2026-07-27T00:00:00+00:00

MoonshotAI's Kimi K3 is now available in LM Studio Bionic.

Frontend Design vs. Impeccable vs. UI/UX Pro Max: Which Design Skill Fits the Job?

2026-07-26T00:00:00+00:00

A repository-level comparison of Anthropic's frontend-design, Impeccable, and UI/UX Pro Max, including where each wins, what their issue trackers reveal, and how GPT-5.6 changes the best default.

Grill Me vs. Superpowers: Use the Smallest Workflow That Controls the Risk

2026-07-26T00:00:00+00:00

A source-based comparison of Matt Pocock's Grill Me and obra's Superpowers: what each workflow actually controls, when a more capable model changes the trade-off, and when combining them creates more process than value.

Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)

2026-07-25T12:21:52+00:00

MCP solved client-to-server auth with OAuth 2.1, but the host itself has no identity of its own. WSO2's Ayesha Dissanayaka breaks down the Four A's framework and OAuth On-Behalf-Of tokens for securing AI agent hosts in production.

July 22, 2026

2026-07-24T14:10:48+00:00

Elicitation is Stabilized

The Elicitation RFD has moved to Completed, and structured form and URL-mode user interactions are now stabilized in ACP.

Agents can request structured non-sensitive input or direct users to secure out-of-band flows, while Clients explicitly advertise the modes they support.

Read the full announcement.

July 6, 2026

2026-07-20T20:49:45+00:00

Boolean Config Options are Stabilized

The Boolean Config Option Type RFD has moved to Completed and boolean session configuration options are stabilized.

Agents can now expose native on/off controls, while v1 clients explicitly opt in through the session.configOptions.boolean capability.

Read the full announcement.

June 29, 2026

2026-07-20T20:49:45+00:00

Request Cancellation is Stabilized

The Request Cancellation RFD has moved to Completed and the $/cancel_request notification is stabilized.

Either side can now request cancellation of an outstanding JSON-RPC request by ID without relying on method-specific cancellation.

Read the full announcement.

June 24, 2026

2026-07-20T20:49:45+00:00

Model Config Category is Stabilized

The Model Config Option Category RFD has moved to Completed and the model_config category is stabilized.

Agents can classify model-related parameters such as context size or speed and quality trade-offs so clients can present them alongside the main model selector.

Read the full announcement.

June 5, 2026

2026-07-20T20:49:45+00:00

Message IDs are Stabilized

The Message ID RFD has moved to Completed and optional Agent-generated messageId fields are stabilized in ACP v1.

Message IDs let clients group streamed chunks into individual messages and detect when a new message begins.

Read the full announcement.

June 1, 2026

2026-07-20T20:49:45+00:00

Additional Workspace Roots are Stabilized

The Additional Workspace Roots RFD has moved to Completed and additionalDirectories is stabilized for session lifecycle requests.

Clients can declare absolute additional workspace roots while cwd remains the primary working directory.

Read the full announcement.

Kimi K3 Hands-On Guide: New Parameters & API Support Matrix

2026-07-17T00:00:00+00:00

July 2026 Kimi K3 guide: reasoning_effort max, thinking history, dynamic tool loading, structured output, auto caching, partial prefix, and vision inputs.

TypeScript OSS SDK: 26 New Providers, Reranking, and Zero-Dependency Imports

2026-07-13T00:00:00+00:00

TypeScript provider expansion

GPT-5.6 Is Live: Prompt Caching Billing Changes Explained

2026-07-10T00:00:00+00:00

July 2026: GPT-5.6 gpt-5.6-sol / terra / luna on AIHubMix: 1.05M context, 1.25x cache writes, prompt_cache_key, explicit breakpoints, vs Claude caching.

SWE-1.7: Frontier Intelligence at a Fraction of the Cost

2026-07-08T00:00:00+00:00

Today, we’re launching SWE-1.7, the most capable model we’ve trained so far. It reaches frontier-level intelligence at a much lower cost, advancing the cost-performance Pareto curve.

Measuring the Trustworthiness of Open-Source-Derived Models

2026-07-08T00:00:00+00:00

We built an evaluation suite to assess model trustworthiness. Our results indicate that models developed from open-source models can be trusted, provided that sufficient thought and care is put into their development.

FrontierCode 1.1

2026-07-07T00:00:00+00:00

FrontierCode 1.1 refines our methodology to distinguish legitimate internet use from unfair use.

Build Your First Autonomous Loop with Claude Code Skills (Practical Guide) - Claude Skills Hub | Claude Skills Hub

2026-07-03T00:00:00+00:00

A practical loop engineering walkthrough: turn Claude Code from a chat tool into a self-verifying autonomous loop using skills, hooks, task lists, and checkers. Includes the loop contract, state files, verification layers, and human checkpoints.

Harness Engineering vs Prompt Engineering: Why the Industry Moved On - Claude Skills Hub | Claude Skills Hub

2026-07-03T00:00:00+00:00

Harness engineering vs prompt engineering explained — why Agent = Model + Harness replaced clever prompts, what a harness contains (skills, hooks, MCP, sub-agents, permissions), and how to migrate from prompting to harness building in 2026.

The Operator Loop Stack: Five Layers of Production Agent Loops - Claude Skills Hub | Claude Skills Hub

2026-07-03T00:00:00+00:00

A deep dive into the operator loop stack — harness, loop contract, state layer, checker, and human checkpoint. How each layer answers one failure mode of autonomous AI agents, with concrete Claude Code implementations for every layer.

Skills vs MCP vs Hooks: Choosing the Right Harness Component - Claude Skills Hub | Claude Skills Hub

2026-07-03T00:00:00+00:00

Skills vs MCP vs hooks compared — what each harness component does, when to use which, and how they compose. A decision guide for agent builders: knowledge goes in skills, external access goes through MCP, guarantees go in hooks.

Your AI Chatbot Just Exposed Your CEO's Salary to an Intern

2026-07-02T03:43:13+00:00

An HR chatbot leaks the CEO's salary to an intern, no jailbreak required. WSO2's Sahan Dilshan and Hasini Samarathunga break down why agent-only and user-only tokens both fail, and how on-behalf-of (OBO) delegation fixes the confused deputy problem in MCP.

Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)

2026-06-30T17:01:47+00:00

Three real MCP breaches: "GitHub" , "CVE-2025-6514" , "Microsoft Research" - share the same root cause: no hard execution boundary. This blog breaks down why IAM and Docker fall short for agentic workloads, and how WebAssembly + WASI's three core primitives (linear memory isolation, fuel counting, WASI pre-opens) solve the problem that your current security stack can't.

Beta SDKs for the 2026-07-28 MCP Spec Release Candidate Are Here

2026-06-29T09:00:00+00:00

Beta releases of the Python, TypeScript, Go, and C# SDKs are now available with support for the 2026-07-28 MCP specification release candidate. Here is what changes for your server, how to migrate, and how to test before the spec goes final on July 28.

SDK Memory Expiration: Expiring Memories Across Python and TypeScript

2026-06-27T00:00:00+00:00

SDK memory expiration

Mem0 Plugin: Shared Memory for Claude Code, Cursor, Codex, and Antigravity

2026-06-25T00:00:00+00:00

Mem0 Plugin v0.2.11

OpenCode Plugin: Native SDK Memory Tools for OpenCode

2026-06-22T00:00:00+00:00

OpenCode Plugin v0.2.0

Enterprise-Managed Authorization: Zero-touch OAuth for MCP

2026-06-18T17:00:00+00:00

The Enterprise-Managed Authorization extension to the Model Context Protocol is now stable, enabling organizations to centrally provision MCP server access through their identity provider so users get connected servers on first login without per-app OAuth.

Pi Agent Plugin: Persistent Memory for Pi Agent

2026-06-12T00:00:00+00:00

Pi Agent Plugin v0.1.2

ChatGPT Referral Traffic Increased ~60% Per Site: What I Found Across Three Analytics Sources

2026-06-08T14:59:21+00:00

ChatGPT referral traffic rose ~60% per site in May 2026. Here is what I found across three analytics sources, and why brand mentions becoming links caused it.

Do LLMs Read llms.txt? The Data from Adobe AEM Says Almost Never

2026-06-01T12:58:28+00:00

Verifiable LLMs are just 1.1% of llms.txt traffic, but Googlebot now tops the list. See the 4,685-domain data and decide if the file is worth it.

OpenAI Compatible Interface Upgrade: Deep Support for Claude

2026-06-01T00:00:00+00:00

AIHubMix has upgraded its OpenAI-compatible interface with deep Claude adaptation: multi-turn interleaved thinking with no extra parameters, automatic prompt cache hits, and support for enabling Anthropic beta features, so Claude delivers its full native capabilities under the OpenAI protocol.

Global Acceleration Network: 75% Lower Latency, 99.99% Availability

2026-06-01T00:00:00+00:00

AIHubMix built its own global edge nodes and intelligent traffic scheduling, cutting response latency by 75%, reducing connection stability fluctuations by 60%, and reaching 99.99% service availability. Multi-region probes run real-time health checks every minute with millisecond-level automatic failover, making AI application calls faster and more stable.

Claude Opus 4.7 New Parameters Guide

2026-06-01T00:00:00+00:00

AiHubMix is a comprehensive AI model API routing service, providing unified access to OpenAI, Claude, Gemini, Deepseek, Qwen, and more. Effortlessly integrate LLMs, multimodal AI, and developer tools with a single, standardized API.

Agentic Long-Term Memory and Meeting Prep

2026-05-01T00:00:00+00:00

The agentic era begins — multiturn agentic LongTerm Memory, Meeting Prep, Reflection Mode, and Google Calendar.

Lightport: Open-Sourcing Glama's AI Gateway

2026-04-27T19:47:28+00:00

A lightweight, open-source AI gateway that makes 80+ LLM providers OpenAI-compatible.

The Zettelkasten Method in Obsidian: A Practical Setup Guide

2026-04-09T07:42:48+00:00

Learn how to set up a working Zettelkasten in Obsidian: note types, vault structure, MOCs, and how AI handles the maintenance that kills most systems.

Tool Definition Quality Score (TDQS)

2026-04-03T19:30:54+00:00

97% of #MCP tool descriptions have quality defects. We are developing Tool Description Quality Score (TDQS) framework to fix it.

gateway.glama.ai

2026-04-02T00:00:00+00:00

[Improvement] Glama Gateway has been moved from https://glama.ai/api/gateway/openai/v1 to just https://gateway.glama.ai/v1 This gives us: Drop-in SDK compatibility — Works out of the box with OpenAI, LangChain, LiteLLM, etc. One line: base_url="https://gateway.glama.ai/v1" . Independent scaling — Gateway traffic

How to Set Up OpenClaw With Your Pieces Long-Term Memory

2026-03-27T00:00:00+00:00

A stepbystep guide to setting up OpenClaw with Pieces LongTerm Memory so your AI agent has full context of your work from day one.

The Hackers Who Tracked My Sleep Cycle

2026-03-26T15:58:43+00:00

How a series of overnight attacks revealed a card testing vulnerability – and the countermeasures that actually worked.

How to Bulk Rename Obsidian Files Without Breaking Every Link

2026-03-23T14:02:39+00:00

Learn how to bulk rename Obsidian files without breaking wikilinks. Compare the Bulk Rename plugin, VS Code workarounds, and AI-powered renaming.

Open Source Has a Bot Problem

2026-03-19T15:46:00+00:00

How a hidden prompt injection in CONTRIBUTING.md revealed that 40% of pull requests to a popular GitHub repository were generated by AI bots

Obsidian Find and Replace Across All Files: How to Make it Work

2026-03-19T07:01:13+00:00

Learn how to find and replace text across all files in your Obsidian vault using built-in tools, plugins, VS Code, and Desktop Commander AI.

Improved text streaming performance

2026-03-17T00:00:00+00:00

[Improvement] Responses now stream at a more consistent pace with smoother text rendering for better readability

New web search tools

2026-03-17T00:00:00+00:00

[Improvement] We just just refactored how the integrated web search tools work. You will see significantly faster response times and more accurate results

How to make a release?

2026-03-15T01:29:29+00:00

A Glama release is not a GitHub release. It is a containerized build of your MCP server that Glama uses to run security checks and to let users deploy the server

Reasoning Capabilities

2026-03-15T00:00:00+00:00

[Improvement] Glama now supports model reasoning capabilities. Additionally, we have standardized every model reasoning capabilities into 3 levels effort: low, medium, and high

Projects & Memory

2026-03-14T00:00:00+00:00

[Improvement] We just added projects with per project documents, memories, and MCP configuration

Spec-Driven Development with Claude Skills: Ship Features Faster

2026-03-10T00:00:00+00:00

Learn how to implement spec-driven development using Claude Skills. A practical guide to the SDD workflow: brainstorm, plan, implement, and ship — with less friction and higher quality.

Top Claude Skills for Security Professionals in 2026

2026-03-10T00:00:00+00:00

Discover the best Claude Skills for security professionals in 2026. From digital forensics and web fuzzing to code auditing and secure data handling, these skills turn Claude into a powerful security assistant.

Notion MCP Server: How to Give Your AI Full Access to Your Workspace

2026-03-04T07:12:23+00:00

Set up Notion MCP to give Claude, ChatGPT, or Cursor direct access to your Notion workspace. Covers hosted and open-source options, 17 tools, and troubleshoo...

The Personal Management System Guide: What Works, What Fails, and How to Build Your Own

2026-03-02T08:49:26+00:00

Build a personal management system that works across all your files and tools using AI-powered local file access — no app switching required.

Scheduled Summaries and a rebuilt local LLM engine

2026-03-01T00:00:00+00:00

The frictionremoval release — Scheduled Summaries, Modality Focus, a rebuilt local LLM engine, and expanded BYOK.

14 Best Obsidian Plugins in 2026: Must-Haves, Nice-to-Haves, and Power Tools

2026-02-23T16:02:19+00:00

Discover the 14 best Obsidian plugins in 2026 for beginners to power users. Clear verdicts, use-cases, and tips for each plugin.

Knowledge Management System: Why Most Fail and What Actually Works

2026-02-18T12:15:50+00:00

Most knowledge management systems fail because they demand constant upkeep. Learn how AI-native local tools like Desktop Commander turn your existing files i...

How to Write GEO Prompts for Reliable LLM Insights

2026-02-17T14:15:17+00:00

Learn about LLM prompts and how they can transform your approach to SEO, aligning with real user inquiries and behaviors.

How Do LLMs Choose Citations? The Selection Process

2026-02-16T07:16:44+00:00

Understand how llm citations help bridge the knowledge gap in large language models and enhance generated content.

I’m biased, but here’s my comparison of Desktop Commander vs Cowork

2026-02-13T08:28:26+00:00

An honest comparison of Desktop Commander and Claude Cowork from someone on the DC team. See where each tool excels — terminal access, file editing, deployme...

Bing AI Performance Report: GEO Impact Analysis

2026-02-11T19:00:26+00:00

Bing AI Performance Report: GEO Impact Analysis

New Sidebar

2026-02-11T00:00:00+00:00

[Improvement] The number of features available has grown substantially over the last few months. As a result, we've evolved the sidebar to only show the most used features. To see the rest, click 'More'

What Is LLM Crawling and Why Does It Matter?

2026-02-05T12:00:00+00:00

What Is LLM Crawling and Why Does It Matter?

Audio capture for Long-Term Memory

2026-02-01T00:00:00+00:00

The biggest capture and integration release yet — LTM Audio, custom templates, and a 39tool MCP server.

Time Breakdown for billable hours

2026-01-27T00:00:00+00:00

The professional productivity release — Time Breakdown, Timeline Search, and model management.

Code Execution

2026-01-24T00:00:00+00:00

[Improvement] Our built-in MCP server has gained the ability to execute code

What Is the Difference Between AI Mentions and Citations?

2026-01-23T20:00:00+00:00

What Is the Difference Between AI Mentions and Citations?

What Is AI Visibility Score and How Do You Measure It

2026-01-23T19:00:00+00:00

What Is AI Visibility Score and How Do You Measure It

New model: glm-4.7-flash

2026-01-22T00:00:00+00:00

[Improvement] glm-4.7-flash

A new Home Base and single-click summaries

2025-12-23T00:00:00+00:00

The foundation release — a new Home Base, personalization, and singleclick summaries.

A comprehensive review of the best AI Memory systems

2025-12-15T00:00:00+00:00

Indepth 2025 review of the best AI memory systems for developers. Compare architectures, features, use cases, and performance

Not a standup tool: the AI memory that makes standups great

2025-12-12T00:00:00+00:00

Ditch standup bots. Use Pieces, an ondevice artificial memory that captures your work to generate accurate, detailed standup reports with a simple prompt.

Building a daily productivity app with Pieces — Part 4: polish & production ready

2025-12-09T00:00:00+00:00

Build a daily productivity app with Pieces in Part 4 of the series: polish the UX, tighten performance, fix edge cases, and make your app productionready with deployment best practices.

Building a daily productivity app with Pieces — Part 2: Adding AI Intelligence with Gemini

2025-12-04T00:00:00+00:00

Build a daily productivity app with Pieces (Part 2) by adding AI intelligence with Google Gemini, covering architecture, prompts, integrations, and practical tips to ship smarter workflows.

Building a daily productivity app with Pieces - Part 3: bringing it all together with Flutter UI

2025-12-04T00:00:00+00:00

Part 3: Flutter UI to bring your Pieces daily productivity app together.

Building daily stand-up generator using Pieces API — Part 1: The SDK overview

2025-12-01T00:00:00+00:00

Learn how to build a daily standup generator using the Pieces API. This first part of the series covers the SDK overview, key capabilities, and how developers can streamline workflow automation with Pieces.

How we stopped watching our engineers struggle through stand-ups

2025-11-27T00:00:00+00:00

Tired of awkward standup meetings where great engineers sound like they did nothing? I automated our team's standups with AI and got 3 people promoted. Here's exactly how we changed standups and made real work visible to managers.

Version 1.143.0-next.2

2025-10-30T00:00:00+00:00

2.0 next init.

Version 1.143.0-next.1

2025-10-30T00:00:00+00:00

Try 2.0 next. starting V2

Version 1.142.8

2025-10-30T00:00:00+00:00

Version 1.142.8

Version 1.142.7

2025-10-28T00:00:00+00:00

Update i18n.

Version 1.142.6

2025-10-28T00:00:00+00:00

Version 1.142.6

Version 1.142.5

2025-10-27T00:00:00+00:00

Add MiniMax-M2 model.

Version 1.142.4

2025-10-27T00:00:00+00:00

Pre render ModelSwitchPanel, The error details of the connectivity check lead to a layout problem.

Version 1.142.3

2025-10-27T00:00:00+00:00

Adjust modal setting form styles for improved layout and responsiveness, Unzip file when uploading in knowledge base \LOB-500].

LM Studio 0.3.29 — OpenAI /v1/responses and variant listing

2025-10-06T00:00:00+00:00

New OpenAI‑compatible endpoint: POST /v1/responses . Stateful interactions via previous_response_id . Custom tool calling and Remote MCP support (opt‑in). Reasoning support with reasoning.effort for openai/gpt‑oss‑20b . Streaming via SSE when stream: true . Stateful interactions via previous_response_id . Custom tool calling and Remote MCP support (opt‑in). Reasoning support with reasoning.effort for openai/gpt‑oss‑20b . Streaming via SSE when stream: true . CLI: lms ls --variants lists all variants for multi‑variant models. Docs: /docs/developer/openai-compat . Full release notes: /blog/lmstudio-v0.3.29 .

LM Studio 0.3.27 — CLI: model resource estimates, status, and interrupts

2025-09-24T00:00:00+00:00

New: lms load --estimate-only prints estimated GPU and total memory before loading. Honors --context-length and --gpu , and uses an improved estimator that now accounts for flash attention and vision models. lms chat : press Ctrl+C to interrupt an ongoing prediction. lms ps --json now reports each model's generation status and the number of queued prediction requests. CLI color contrast improved for light mode. See docs: /docs/cli/local-models/load . Full release notes: /blog/lmstudio-v0.3.27 .

LM Studio 0.3.26 — CLI log streaming: server + model

2025-09-15T00:00:00+00:00

lms log stream now supports multiple sources and filters. --source server streams HTTP server logs (startup, endpoints, status) --source model --filter input,output streams formatted user input and model output Append --json for machine‑readable logs; --stats adds tokens/sec and related metrics (model source) --source server streams HTTP server logs (startup, endpoints, status) --source model --filter input,output streams formatted user input and model output Append --json for machine‑readable logs; --stats adds tokens/sec and related metrics (model source) See usage and examples: /docs/cli/serve/log-stream . Full release notes: /blog/lmstudio-v0.3.26 .

LM Studio 0.3.25 — New model support (API)

2025-09-04T00:00:00+00:00

Added support for NVIDIA Nemotron‑Nano‑v2 with tool‑calling via the OpenAI‑compatible endpoints ‡ . Added support for Google EmbeddingGemma for the /v1/embeddings endpoint ‡ .

LM Studio 0.3.24 — Seed‑OSS tool‑calling and template fixes

2025-08-28T00:00:00+00:00

Added support for ByteDance/Seed‑OSS including tool‑calling and prompt‑template compatibility fixes in the OpenAI‑compatible API ‡ . Fixed cases where tool calls were not parsed for certain prompt templates ‡ .

Flat Capital invests in Pieces

2025-08-12T17:20:55+00:00

Flat Capital invests in Pieces

LM Studio 0.3.23 — Reasoning content and tool‑calling reliability

2025-08-12T00:00:00+00:00

For gpt‑oss on POST /v1/chat/completions , reasoning content moves out of message.content and into choices.message.reasoning (non‑streaming) and choices.delta.reasoning (streaming), aligning with o3‑mini ‡ . Tool names are normalized (e.g., snake_case) before being provided to the model to improve tool‑calling reliability ‡ . Fixed errors for certain tools‑containing requests to POST /v1/chat/completions (e.g., "reading 'properties'") and non‑streaming tool‑call failures ‡ .

LM Studio 0.3.19 — Bug fixes for streaming and tool calls

2025-07-21T00:00:00+00:00

Corrected usage statistics returned by OpenAI‑compatible streaming responses ‡ . Improved handling of parallel tool calls via the streaming API ‡ . Fixed parsing of correct tool calls for certain Mistral models ‡ .

LM Studio 0.3.18 — Streaming options and tool‑calling improvements

2025-07-10T00:00:00+00:00

Added support for the stream_options object on OpenAI‑compatible endpoints. Setting stream_options.include_usage to true returns prompt and completion token usage during streaming ‡ . Errors returned from streaming endpoints now follow the correct format expected by OpenAI clients ‡ . Tool‑calling support added for Mistral v13 tokenizer models, using proper chat templates ‡ . The response_format.type field now accepts "text" in chat‑completion requests ‡ . Fixed bugs where parallel tool calls split across multiple chunks were dropped and where root‑level $defs in tool definitions were stripped ‡ .

Nano-Models power LTM-2.5

2025-04-17T00:00:00+00:00

Nano-Models power LTM-2.5

Connect Long-Term Memory to your AI tools

2025-04-15T00:00:00+00:00

You can now connect your Pieces LongTerm Memory to your favorite AI tools—like GitHub Copilot, Cursor, Block (codename Goose), and more.

SSO for AI Agents with OpenID Connect

2024-12-29T00:00:00+00:00

SSO for AI agents is plain OpenID Connect. The AgentID discovery values, open vs registered clients, and setup on Clerk, Supabase, Auth0, Descope and Better Auth.

Rate Limit AI Agents per Human, Not per Account

2024-12-07T00:00:00+00:00

Rate-limit AI agents per human, not per account: key your limiter on the AgentID owner_sub claim so ten agents from one person share one quota.

Adding a Sign in with AgentID Button

2024-12-07T00:00:00+00:00

Add a Sign in with AgentID button with two configuration values. Setup for Clerk, Supabase, Auth0, Descope, Better Auth, and Auth.js, plus the button guidelines.

Know Your Agent (KYA): Verify an Agent's Owner

2024-12-01T00:00:00+00:00

Know Your Agent (KYA) at sign-in: verify the human behind an AI agent with AgentID. Request owner_email and read it from the id_token or /userinfo.

The Best MCP Servers in 2026: A Curated List for Real Work

2024-11-21T00:00:00+00:00

A vetted, opinionated list of the best MCP servers you can add to Claude Code, Claude Desktop, Cursor, and other MCP clients today, grouped by what they're actually good for.

AgentID vs Auth0 for AI Agents

2024-11-12T00:00:00+00:00

Auth0 for AI Agents secures agents acting for your users. AgentID gives agents their own cross-app identity with owner accountability, and plugs into Auth0.

Is an AI Agent Acting for Itself or a User?

2024-10-14T00:00:00+00:00

Tell in code whether an AI agent acts for itself or for a user: principal tokens vs delegated access vs workload identity, and what revocation means in each.

AI Agent Identity vs Delegated Access vs Workload Identity

2024-09-07T00:00:00+00:00

A field guide to agent identity terms: agent ID, agentic identity, delegated access, workload identity, token vaults, SPIFFE and auth.md, with named examples.

What Is an MCP Gateway? Architecture, Benefits, and When You Need One

2024-07-31T00:00:00+00:00

An MCP gateway sits between your AI clients and every MCP server you run, turning a sprawl of point-to-point connections into one governed, auditable endpoint.

AgentID with Supabase: What Breaks and Why

2024-06-30T00:00:00+00:00

AgentID in Supabase Auth after setup: comma-separated scopes, the claims allowlist for owner claims, the free-plan provider limit, and what symptoms mean.

A2A vs MCP: How the Two Agent Protocols Fit Together

2024-04-16T00:00:00+00:00

A2A vs MCP is often framed as a rivalry, but A2A handles agent-to-agent communication while MCP connects a single agent to its tools and data — most production systems end up using both.

Claude Marketing Skills: How to Put Claude to Work on Real Marketing Tasks

2024-04-02T00:00:00+00:00

Claude marketing skills turn a general-purpose chatbot into a repeatable operator for brand voice, campaign briefs, SEO checklists, and social repurposing — here's what's official and how to build the rest.

Claude Skills 101: From Zero to Your First Custom Skill

2024-03-20T00:00:00+00:00

A comprehensive guide to understanding and creating your first custom Claude Skill.

Getting Started with MCP: A Developer's First Steps

2024-02-24T00:00:00+00:00

A comprehensive guide for developers to get started with the Model Context Protocol (MCP), including setting up the environment and building a simple server.

June 2026 Release Spotlight: ~20 New Models

2023-11-18T00:00:00+00:00

In June 2026 AIHubMix added ~20 new models (claude-fable-5 [Retired], glm-5.2, minimax-m3, qwen3.7-plus, kimi-k2.7-code, Kling video and more) and shipped the LLM Router, any-model Responses protocol, Mapping & Fallback, the AIHubMix CLI and Skill, and a backup domain.

The 2026-07-28 MCP Specification: A Stateless, Extensible Future

2023-11-09T00:00:00+00:00

A deep dive into the largest MCP protocol revision since launch: a stateless core, multi round-trip requests, an official extensions framework, and hardened authorization.

MCP Security Best Practices: A Practical Guide for 2026

2023-08-26T00:00:00+00:00

A senior engineer's guide to MCP security best practices, covering prompt injection, tool poisoning, confused deputy attacks, and OAuth 2.1 hardening for anyone building or deploying Model Context Protocol servers.

Mem0 Research: LoCoMo, LongMemEval & BEAM Benchmarks

2023-07-25T00:00:00+00:00

Mem0 scores 92.5 on LoCoMo and 94.4 on LongMemEval, achieving high accuracy at 3-4x lower token cost than full-context memory systems.

Free AI Models on AIHubMix

2023-04-19T00:00:00+00:00

Free AI Models: The Ultimate Guide to Building with Zero-Cost AI in 2026 on AIHubMix