Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 327 Agent status: 1 with issues Updated Oct 7, 2026
AgentGid

LangChain changelog: what's new each month

Every stable LangChain release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

September 2026

6 releases: 1.4.0a3 → 1.4.3

September brought a major new feature for integrating Model Context Protocol (MCP) servers with LangChain, along with several refinements to tool handling and model support. The month saw rapid iteration with six releases shipping improvements to MCP integration and fixes for edge cases in tool management.

Highlights

  • MCPAdapter lets you connect MCP servers to LangChain—whether via URL, local script, in-process server, or pre-built client—with optional response caching.
  • Tool metadata from MCP servers is now organized under an mcp namespace for easier inspection and management.
  • Mid-call server questions can surface as LangGraph interrupts for human-in-the-loop workflows.
  • New aslangchaintool() function converts individual MCP tools for applications managing their own clients.
  • Added support for Bedrock Mantle chat models in initchatmodel.

New

  • MCPAdapter with support for multiple server sources and caching modes.
  • MCPAdapter.listtools() with optional client-side response caching.
  • aslangchaintool() for single MCP tool conversion.
  • elicitation="interrupt" mode to surface server questions as LangGraph interrupts.
  • langchain.mcp namespace for MCP-related functionality.
  • Bedrock Mantle chat model support in initchatmodel.

Improved

  • Tool metadata grouped under mcp namespace with annotations and server identity.
  • MCP elicitation now driven via group member configuration.
  • Omit middleware trace inputs for anthropic and langchain (performance).
  • Mixed-era ClientGroup handling in MCP adapter.

Fixed

  • Preserve open MCP object arguments during tool call handling.
  • Correct InterruptOnConfig documentation.
  • Preserve model-generated tool calls in human-in-the-loop tool call edits.
  • Sanitize cache settings for fallback models.

August 2026

5 releases: 1.3.15 → 1.4.0a1

August brought incremental refinements across middleware, error handling, and content processing, culminating in an early alpha release. The month focused on stability and extending integration capabilities.

Highlights

  • Exposed tracepolicy configuration on AgentMiddleware for better observability control.
  • Added standard model exception types for more consistent error handling across the platform.
  • Introduced support for custom token counters in ContextEditingMiddleware.
  • Improved content preservation in PIIMiddleware redaction and fixed streaming content indexing issues.

New

  • Standard model exception types for consistent error handling.
  • Custom token counter support in ContextEditingMiddleware.
  • Tracepolicy exposure on AgentMiddleware.

Improved

  • Content block shape is now preserved during PIIMiddleware redaction.
  • Streaming content indexing in genai v1 was enhanced.
  • Custom HITL rejection reasons are now properly framed.

Fixed

  • Content block shape preservation in PIIMiddleware redaction.
  • Streaming content indexing in genai v1.
  • Custom HITL rejection reason handling.
  • MCP content type conversion error naming.

July 2026

3 releases: 1.3.12 → 1.3.14

July brought bug fixes to middleware components and new support for OpenAI prompt caching and Meta models. Three releases focused on improving tool handling reliability and expanding model provider coverage.

Highlights

  • Added support for OpenAI's explicit prompt caching to reduce API costs and latency.
  • Introduced new ToolErrorMiddleware for better error handling in tool calls.
  • Fixed interrupt propagation through ToolRetryMiddleware for more predictable behavior.
  • Added Meta model support with new langchain-meta integration option.

New

  • Meta extra and langchain-meta support in InitChatModel
  • OpenAI explicit prompt caching capability
  • ToolErrorMiddleware for tool call error handling

Improved

  • ToolRetryMiddleware now retries only retryable exceptions instead of all failures
  • Shell middleware process group handling to avoid unintended process kills

Fixed

  • Interrupt propagation through ToolRetryMiddleware
  • Shell middleware process-group kill behavior affecting background processes

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. LangChain: pricing, features and alternatives · All changelogs