MCP 服务器

MCP 服务器是一种为 LLM 提供工具调用能力的一种方式，也是使用 OpenAI 兼容工具调用的替代方案。

通过将 MCP(Anthropic)工具定义转换为与 OpenAI 兼容的工具定义，您可以将 MCP 服务器与 OpenRouter 结合使用。

在本例中，我们将使用 Anthropic 的 MCP 客户端 SDK 与文件系统 MCP 进行交互，所有这些都在 OpenRouter 的后台进行。

注

请注意，与 MCP 服务器交互比调用 REST 端点更复杂。MCP 协议具有状态，需要会话管理。以下示例使用了 MCP 客户端 SDK，但仍然有些复杂。

First, some setup. In order to run this you will need to pip install the packages, and create a .env file with OPENAI_API_KEY set. This example also assumes the directory /Applications exists. 首先，进行一些设置。为了运行此程序，您需要使用 pip 安装软件包，并创建一个 .env 文件，同时设置 OPENAI_API_KEY。此示例还假设已有目录 /Applications。

import asyncio
from typing import Optional
from contextlib import AsyncExitStack

from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

from openai import OpenAI
from dotenv import load_dotenv
import json

load_dotenv()  # load environment variables from .env

MODEL = "anthropic/claude-3-7-sonnet"

SERVER_CONFIG = {
    "command": "npx",
    "args": ["-y",
              "@modelcontextprotocol/server-filesystem",
              f"/Applications/"],
    "env": None
}

接下来，我们的辅助函数将 MCP 工具定义转换为 OpenAI 工具定义：

def convert_tool_format(tool):
    converted_tool = {
        "type": "function",
        "function": {
            "name": tool.name,
            "description": tool.description,
            "parameters": {
                "type": "object",
                "properties": tool.inputSchema["properties"],
                "required": tool.inputSchema["required"]
            }
        }
    }
    return converted_tool

还有 MCP 客户端本身，令人遗憾的是，它竟然有大约 100 行代码。需要注意的是，SERVER_CONFIG 为硬编码到客户端，但当然可以针对其他 MCP 服务器进行参数化。

class MCPClient:
    def __init__(self):
        self.session: Optional[ClientSession] = None
        self.exit_stack = AsyncExitStack()
        self.openai = OpenAI(
            base_url="https://openrouter.co/v1"
        )

    async def connect_to_server(self, server_config):
        server_params = StdioServerParameters(**server_config)
        stdio_transport = await self.exit_stack.enter_async_context(stdio_client(server_params))
        self.stdio, self.write = stdio_transport
        self.session = await self.exit_stack.enter_async_context(ClientSession(self.stdio, self.write))

        await self.session.initialize()

        # List available tools from the MCP server
        response = await self.session.list_tools()
        print("\nConnected to server with tools:", [tool.name for tool in response.tools])

        self.messages = []

    async def process_query(self, query: str) -> str:

        self.messages.append({
            "role": "user",
            "content": query
        })

        response = await self.session.list_tools()
        available_tools = [convert_tool_format(tool) for tool in response.tools]

        response = self.openai.chat.completions.create(
            model=MODEL,
            tools=available_tools,
            messages=self.messages
        )
        self.messages.append(response.choices[0].message.model_dump())

        final_text = []
        content = response.choices[0].message
        if content.tool_calls is not None:
            tool_name = content.tool_calls[0].function.name
            tool_args = content.tool_calls[0].function.arguments
            tool_args = json.loads(tool_args) if tool_args else {}

            # Execute tool call
            try:
                result = await self.session.call_tool(tool_name, tool_args)
                final_text.append(f"[Calling tool {tool_name} with args {tool_args}]")
            except Exception as e:
                print(f"Error calling tool {tool_name}: {e}")
                result = None

            self.messages.append({
                "role": "tool",
                "tool_call_id": content.tool_calls[0].id,
                "name": tool_name,
                "content": result.content
            })

            response = self.openai.chat.completions.create(
                model=MODEL,
                max_tokens=1000,
                messages=self.messages,
            )

            final_text.append(response.choices[0].message.content)
        else:
            final_text.append(content.content)

        return "\n".join(final_text)

    async def chat_loop(self):
        """Run an interactive chat loop"""
        print("\nMCP Client Started!")
        print("Type your queries or 'quit' to exit.")

        while True:
            try:
                query = input("\nQuery: ").strip()
                result = await self.process_query(query)
                print("Result:")
                print(result)

            except Exception as e:
                print(f"Error: {str(e)}")

    async def cleanup(self):
        await self.exit_stack.aclose()

async def main():
    client = MCPClient()
    try:
        await client.connect_to_server(SERVER_CONFIG)
        await client.chat_loop()
    finally:
        await client.cleanup()

if __name__ == "__main__":
    import sys
    asyncio.run(main())

将以上所有代码组装到 mcp-client.py 中，您将获得一个行为如下的客户端(为简洁起见，某些输出已被截断)：

% python mcp-client.py

Secure MCP Filesystem Server running on stdio
Allowed directories: [ '/Applications' ]

Connected to server with tools: ['read_file', 'read_multiple_files', 'write_file'...]

MCP Client Started!
Type your queries or 'quit' to exit.

Query: Do I have microsoft office installed?

Result:
[Calling tool list_allowed_directories with args {}]
I can check if Microsoft Office is installed in the Applications folder:

Query: continue

Result:
[Calling tool search_files with args {'path': '/Applications', 'pattern': 'Microsoft'}]
Now let me check specifically for Microsoft Office applications:

Query: continue

Result:
I can see from the search results that Microsoft Office is indeed installed on your system.
The search found the following main Microsoft Office applications:

1. Microsoft Excel - /Applications/Microsoft Excel.app
2. Microsoft PowerPoint - /Applications/Microsoft PowerPoint.app
3. Microsoft Word - /Applications/Microsoft Word.app
4. OneDrive - /Applications/OneDrive.app (which includes Microsoft SharePoint integration)