LangChain Changelog

Engine usage, spend limits, and pricing now display in LangChain Standard Units at $1 each, and OAuth applications can use a single access token for the LangSmith API and Remote MCP server. Other changes include fetch_runs deprecated in favor of query_runs, cost resolution for OpenAI and Vertex AI-hosted open models, and Claude Sonnet 5.5 and Gemini 3.7/3.8 Flash in the Playground.

Evaluators using organization-scoped model configurations no longer attempt per-model OAuth token exchange, and Amazon Bedrock models are no longer rejected; annotation queues can now add selected thread items to a dataset. LangSmith Go and Java SDKs can save Insights report configurations, and the thread traces API accepts trace_filter and tree_filter parameters.

Starting September 14, 2026, traces ingested into LangSmith SaaS with extended retention are kept at most 180 days, existing settings above that are applied as 180 days for new traces, and new settings above 180 days are rejected, while self-hosted and BYOC deployments are unchanged. Organization Operators can no longer grant the Organization Admin role to a service key, and multipart feedback ingest now records a rejection error on the run instead of returning a 400 that dropped the run and all accompanying feedback.

Adds /v2/threads/stats for thread/trace metrics, preserves OAuth/session auth across Gateway routing, and requires download permission for dataset exports alongside stricter validation for malformed IDs, shares, and null artifacts. Trajectory batch reads move to a dedicated gRPC service.

LLM gateway traces no longer capture request and response content by default, and Engine usage now costs 40% fewer Usage Credits. Also added ABAC project trace access control, OpenPull request creation for Engine issues, and Bedrock routing for newer OpenAI models.

New annotation queue management endpoints (add to queue, list items, bulk-add runs) with thread support, plus numerous observability fixes: captured examples now copy attachments, annotation queue items include tenant IDs, and evaluation rule runs with num_epochs drift beyond boundaries.

LLM Gateway now reaches Amazon Bedrock with an assumed IAM role or AWS access keys instead of a static API key, and routes requests to each model family's native API format. The Turns view in thread detail is deprecated in favor of Messages or Details, with LangSmith Cloud removing it on October 31, 2026.

Fleet gains Google connection grants for Docs, Sheets, and Slides, with native .docx/.xlsx/.pptx attachment handling and PPT output. Added Google Drive search/organize/share, a built-in spreadsheet skill via openpyxl, user-identity GitHub requests, custom Slack app channel subscriptions, artifact feedback chips, and file conversion for legacy Office, OpenDocument, RTF, EPUB, and text-based PDFs, plus graceful fallback when a model can't accept attachments. Skills can be created from local folders preserving SKILL.md, and folder uploads accept up to 500 files.

Access-policy endpoints now report errors as RFC 7807 problem details, moving error messages to the detail field and returning 422 for invalid request bodies. LangSmith MCP's fetch_runs tool can now include invocation parameters, and manual automation triggers now work for all rule types. Numerous fixes across datasets, tracing, prompts, and deployments.

New annotation queue item endpoints cover CRUD, counting, positioning, and reviewing run/thread items. Sandboxes offer streaming execute requests as Server-Sent Events, and legacy endpoints are deprecated with a January 31, 2027 sunset date.

Fleet now includes a built-in configuration-hardening skill, an edge-to-edge file workspace with collapsible panel, and self-hosted sandbox-backed computer access without cloud billing. Slack file attachments are available under /workspace/uploads, with a fix for agent chat loading and sandbox routing for split Go/Python self-hosted deployments.

Annotation queue item endpoints are now available through the public API and SDK generation flow, and the sandbox supports a new streaming execute request that returns stdout and stderr as Server-Sent Events. All legacy endpoints replaced by the SmithDB SDK migration guide are marked deprecated with a January 31, 2027 sunset date.

Fleet adds a Slack file sending tool (slack_send_file) and a monthly LCU spend meter on the usage dashboard. Fixes a chat view crash when reopening active runs and requires OAuth (not API-key auth) for Arcade MCP gateways, plus expands thread API details and Google error explanations.

LLM Gateway gains full rate-limit policy management (create, edit, delete, enable/disable) and now lives in a dedicated sidebar section. Custom dashboard charts can query P90/P95 latency, first-token time, token, and cost metrics. Annotation queue size limits are now enforced consistently for runs and threads.

Annotation queue item endpoints are now available through the public API for adding, listing, updating, deleting, counting, positioning, and reviewing run or thread queue items. All legacy endpoints replaced by the SmithDB SDK migration guide are marked deprecated with a January 31, 2027 Sunset date and a link to the migration guide. Also fixes CSV uploads on Windows browsers and adds a streaming sandbox execute request.

Legacy feedback formula endpoints are deprecated in favor of composite evaluators and scheduled for removal on 2026-08-20. The experiments table and comparison views received major updates: chips now lay out from real measurements with clickable overflow menus, splits display as interactive chips, multi-turn evaluators gained a Test action, and run stats now populate reliably from SmithDB on self-hosted deployments using SSO/OAuth.

Fleet agents can now delete files from memory and skills using a new delete tool, while MCP servers requiring client-secret authentication now complete OAuth successfully. Fleet agents can read files shared in Slack, and the Agent Builder welcome flow is streamlined with two clear paths instead of a multi-step wizard.

Fleet agents can now delete files from memory and linked workspace skills using a new delete tool, receive files shared in Slack, and send the negotiated MCP protocol version during handshakes so servers requiring newer versions no longer fail. The agent builder now sources workspace and tenant lists from the Fleet API for easier switching, and OAuth now completes for MCP servers requiring client-secret authentication.

Select multiple rows in an experiment and add, replace, or remove their dataset splits in one action, or copy selected examples to another dataset. Fixed experiment comparison views to load run rows on self-hosted SSO/OAuth deployments, expand run trees reliably for repetition runs with project IDs, and correctly fetch run inputs and outputs from SmithDB when creating dataset examples from source runs. Plus dozens of observability, tracing, and evaluation improvements across datasets, experiments, evaluators, and run management.

LangSmith now enforces user-defined monthly trace limits scoped to individual projects and users, rejecting new traces that exceed a configured limit while allowing patches and feedback to continue. The experiment comparison view now shows each example's current dataset split rather than the split it had when the experiment ran, enabling users to see whether failures already belong to a split without re-running. Plus dozens of improvements across evaluators, tracing, playground model support, and dashboard rendering.