LangChain Revamps MCP Support with Stateless Core

LangChain has overhauled its support for the Model Context Protocol (MCP) by moving it directly into the main library under the langchain.mcp namespace. This update aligns with the protocol's major July rewrite and addresses massive growth in adoption, with official Tier 1 SDKs nearing half a billion monthly downloads and ChatGPT users increasing their MCP tool calls by 98x across 2026. Developers can access these features in beta by installing langchain[mcp] version 1.4.0 or higher, with TypeScript support expected to follow the initial Python release.

The integration is built on FastMCP 4, which handles transport negotiation, authentication, and connection management. The biggest architectural shift is the transition to a stateless core. Previously, every MCP call required establishing a session, forcing developers to manage sticky routing and shared session stores at scale. The new stateless specification eliminates session pinning entirely, meaning server redeployments no longer disrupt active workflows. FastMCP automatically negotiates between the new stateless protocol and legacy handshake-based servers, allowing models like google_genai:gemini-3.8-flash to interact with diverse tools seamlessly.

Practitioners can now leverage client-side caching and elicitation. By setting the cache to true, the client respects the server's time-to-live parameters, preventing redundant round-trip requests to fetch tool catalogs on every run. Elicitation allows tools to pause execution to ask users for clarification or confirmation—such as verifying a deletion—without keeping a connection open. LangChain surfaces these pauses as standard LangGraph interrupts, resuming once the user provides an answer.

For developers migrating from the older langchain-mcp-adapters package, the MultiServerMCPClient has been consolidated into a single MCPAdapter class. When managing multiple servers, tool names are automatically prefixed with their source server to prevent naming conflicts, turning generic search tools into distinct options like billing_search and docs_search. The client also supports advanced configurations, including OAuth 2.1, bearer tokens, and shared httpx2 connection pools.