模型上下文协议 (MCP) 是一个开放协议，用于标准化应用程序如何向语言模型提供工具和上下文。LangGraph 代理可以通过 langchain-mcp-adapters 库使用 MCP 服务器上定义的工具。

MCP

安装 langchain-mcp-adapters 库以在 LangGraph 中使用 MCP 工具

<span id="__span-0-1">pip install langchain-mcp-adapters

langchain-mcp-adapters 包使代理能够使用一个或多个 MCP 服务器上定义的工具。

使用 MCP 服务器上定义的工具的代理

<span id="__span-1-1">from langchain_mcp_adapters.client import MultiServerMCPClient
<span id="__span-1-2">from langgraph.prebuilt import create_react_agent
<span id="__span-1-3">
<span id="__span-1-4">client = MultiServerMCPClient(
<span id="__span-1-5">    {
<span id="__span-1-6">        "math": {
<span id="__span-1-7">            "command": "python",
<span id="__span-1-8">            # Replace with absolute path to your math_server.py file
<span id="__span-1-9">            "args": ["/path/to/math_server.py"],
<span id="__span-1-10">            "transport": "stdio",
<span id="__span-1-11">        },
<span id="__span-1-12">        "weather": {
<span id="__span-1-13">            # Ensure your start your weather server on port 8000
<span id="__span-1-14">            "url": "https://:8000/mcp",
<span id="__span-1-15">            "transport": "streamable_http",
<span id="__span-1-16">        }
<span id="__span-1-17">    }
<span id="__span-1-18">)
<span id="__span-1-19">tools = await client.get_tools()
<span id="__span-1-20">agent = create_react_agent(
<span id="__span-1-21">    "anthropic:claude-3-7-sonnet-latest",
<span id="__span-1-22">    tools
<span id="__span-1-23">)
<span id="__span-1-24">math_response = await agent.ainvoke(
<span id="__span-1-25">    {"messages": [{"role": "user", "content": "what's (3 + 5) x 12?"}]}
<span id="__span-1-26">)
<span id="__span-1-27">weather_response = await agent.ainvoke(
<span id="__span-1-28">    {"messages": [{"role": "user", "content": "what is the weather in nyc?"}]}
<span id="__span-1-29">)

自定义 MCP 服务器

要创建您自己的 MCP 服务器，您可以使用 mcp 库。该库提供了一种定义工具并将其作为服务器运行的简单方法。

安装 MCP 库

使用以下参考实现来测试您的代理与 MCP 工具服务器的交互。

数学服务器示例（stdio 传输）

<span id="__span-3-1">from mcp.server.fastmcp import FastMCP
<span id="__span-3-2">
<span id="__span-3-3">mcp = FastMCP("Math")
<span id="__span-3-4">
<span id="__span-3-5">@mcp.tool()
<span id="__span-3-6">def add(a: int, b: int) -&gt; int:
<span id="__span-3-7">    """Add two numbers"""
<span id="__span-3-8">    return a + b
<span id="__span-3-9">
<span id="__span-3-10">@mcp.tool()
<span id="__span-3-11">def multiply(a: int, b: int) -&gt; int:
<span id="__span-3-12">    """Multiply two numbers"""
<span id="__span-3-13">    return a * b
<span id="__span-3-14">
<span id="__span-3-15">if __name__ == "__main__":
<span id="__span-3-16">    mcp.run(transport="stdio")

天气服务器示例（可流式 HTTP 传输）

<span id="__span-4-1">from mcp.server.fastmcp import FastMCP
<span id="__span-4-2">
<span id="__span-4-3">mcp = FastMCP("Weather")
<span id="__span-4-4">
<span id="__span-4-5">@mcp.tool()
<span id="__span-4-6">async def get_weather(location: str) -&gt; str:
<span id="__span-4-7">    """Get weather for location."""
<span id="__span-4-8">    return "It's always sunny in New York"
<span id="__span-4-9">
<span id="__span-4-10">if __name__ == "__main__":
<span id="__span-4-11">    mcp.run(transport="streamable-http")

附加资源