Phase 11 · MCP Tool Interface: let Agents call standard-protocol tools

Maps to report P0 "MCP tool interface". Source: examples/common/mcp_server.py, examples/p11/hello_chain.py. Requires pip install mcp langchain-mcp-adapters.

Why MCP

Earlier tools were hardcoded @tool inside the Agent. In production, tools often come from external systems / another Agent and should not be hardcoded. MCP (Model Context Protocol) is the hot protocol for Agent interconnection: it decouples tool provider from tool caller — swap a server via config, no Agent change.

Tool server (standard MCP)

from mcp.server.fastmcp import FastMCP
mcp = FastMCP("langgraph-course-tools")

@mcp.tool()
def get_weather(city: str) -> str:
    """Query current weather for a city."""
    ...

if __name__ == "__main__":
    mcp.run(transport="stdio")

Agent side (langchain-mcp-adapters)

from langchain_mcp_adapters.client import MultiServerMCPClient
from langgraph.prebuilt import create_react_agent

client = MultiServerMCPClient({
    "course-tools": {"command": sys.executable,
                     "args": ["-m", "examples.common.mcp_server"],
                     "transport": "stdio"},
})
tools = await client.get_tools()
agent = create_react_agent(get_chat_model(), tools)

You can later replace mcp_server.py with any third-party MCP server (weather, DB, internal API) with zero Agent changes.

Run

make run p=11

Summary

• MCP decouples tool provider from tool caller: swap a tool / server via config, no Agent code change.
• This tutorial uses langchain-mcp-adapters to launch the built-in mcp_server.py over stdio; in production you can point it at any third-party MCP server (weather, DB, internal API).
• Note: create_react_agent is itself a state graph under the hood — exactly the "state graph + node + conditional edge" paradigm applied to tool calling.