Model Context Protocol (MCP)

Model Context Protocol (MCP) is an open protocol that standardizes how applications provide tools and context to language models. LangChain agents call tools defined on MCP servers through MCPAdapter, which discovers a server’s tools and adapts them into LangChain tools you can pass straight to create_agent.

MCPAdapter is built on FastMCP, which handles transport inference, protocol negotiation, connection management, and authentication. This section covers the LangChain-specific layer and links out to the FastMCP client documentation for the connection details underneath.

Note

The langchain.mcp namespace requires langchain[mcp]>=1.4.0 and is in beta. Importing from it raises a LangChainBetaWarning once per process. The API may change.

If you used MCP before v1.4.0, see Migrate from langchain-mcp-adapters.

Install

Install LangChain with the mcp extra, which pulls in FastMCP:

pip

pip install "langchain[mcp]"

uv

uv add "langchain[mcp]"

Quickstart

Open an MCPAdapter, discover the server’s tools with list_tools(), and build the agent inside the context. The tools hold the client, so the agent stays usable after the context exits:

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter


async def main():
    async with MCPAdapter("https://example.com/mcp") as adapter:
        tools = await adapter.list_tools()
        agent = create_agent("claude-sonnet-5", tools)
        return await agent.ainvoke({"messages": [{"role": "user", "content": "..."}]})

• View example trace: Open a public LangSmith run for this example.

LangChain docs MCP server

The LangChain docs MCP server is a public HTTP endpoint at https://docs.langchain.com/mcp.

Connect an agent to it to search and read documentation without writing custom tools:

from langchain.agents import create_agent
from langchain.mcp import MCPAdapter


async def main():
    async with MCPAdapter("https://docs.langchain.com/mcp") as adapter:  # [!code highlight]
        tools = await adapter.list_tools()
        agent = create_agent("claude-sonnet-5", tools)
        return await agent.ainvoke(
            {
                "messages": [
                    {
                        "role": "user",
                        "content": "How do I add short-term memory to a LangChain agent?",
                    }
                ]
            }
        )

Note

The docs MCP server is public and does not require an API key. For IDE and coding-agent setup (Claude Code, Cursor, and others), see Use docs programmatically.

The server exposes these tools:

Transports

MCPAdapter infers the transport from the target you provide. The target is the only difference between an in-process server, a local script over stdio, and a remote URL:

from pathlib import Path

from langchain.mcp import MCPAdapter

# An in-process FastMCP server: no subprocess, no socket. Ideal for tests.
in_memory = MCPAdapter(server)  # a FastMCP instance

# A script path is launched over stdio, one subprocess per adapter.
stdio = MCPAdapter(Path("weather_server.py"))

# A string must be an http(s) URL, reached over streamable HTTP.
http = MCPAdapter("https://example.com/mcp")

A target can be any of the following:

• An http/https URL (str): reached over Streamable HTTP.
• A script path (Path): launched as a subprocess over stdio.
• A transport object (StreamableTransport): a preconfigured transport object. See Client Transports.
• An in-process FastMCP server: connected in memory, without a subprocess or socket.
• An MCPConfig dict ({"mcpServers": {...}}): several servers behind one adapter. See Connections.
• A prebuilt fastmcp.Client: for full control over transport, caching, and protocol negotiation.

Warning

A str target must be an http or https URL. FastMCP resolves a string by testing it as a filesystem path before testing it as a URL, so a string naming an existing .py or .js file would launch that file as a subprocess. Because strings are the form a target most often arrives in from configuration or from a model, MCPAdapter rejects strings that do not match the shape of a URL.

Next steps

• Tools: Load MCP tools into agents, control their execution, and handle their outputs.
• Connections: Connection lifecycle, multiple servers, protocol eras, and caching.
• Authentication: Bearer tokens, OAuth 2.1, and per-user server auth.