LangChain ships a first-party MCP adapter

LangChain shipped an alpha of langchain.mcp, a first-party adapter that turns any Model Context Protocol server into LangChain-compatible tools, in releases 1.4.0a1 and 1.4.0a2 on August 27 and 28. The adapter supports multiple MCP servers at once with automatic tool namespacing to avoid collisions, works with both legacy and current MCP protocol versions, and handles OAuth and bearer-token authentication out of the box. It also caches responses and, notably, supports elicitation, MCP’s mechanism for a tool to pause mid-execution and ask the user a question, by routing that pause through LangGraph interrupts rather than requiring a custom workaround.

The companion langchain-anthropic package was updated the same week to add Anthropic SDK 1.0 support, container-based skills configuration, and updated thinking-display modes.

Why the elicitation piece matters most

Most MCP adapters handle the easy part, calling a tool and returning its result, but elicitation is the harder case: a tool that needs to ask a follow-up question mid-run without breaking the surrounding agent loop. Wiring that natively into LangGraph’s interrupt mechanism means a LangChain-based agent can use an MCP server that needs clarification, credentials, or a confirmation step without the developer writing bespoke pause-and-resume logic for each server. It follows the same MCP-adoption pattern seen across the ecosystem this year, including Vercel’s AI SDK stabilizing its own tool-repair handling and the broader Agent Plugins standard several rivals agreed to in June, one more sign that MCP support is becoming table stakes rather than a differentiator for agent frameworks.