【大模型实战篇】基于Claude MCP协议的智能体落地示例

1. 背景

        之前我们在《MCP(Model Context Protocol) 大模型智能体第一个开源标准协议》一文中，介绍了MCP的概念，虽然了解了其概念、架构、解决的问题，但还缺少具体的示例，来帮助进一步理解整套MCP框架如何落地。

        今天我们基于claude的官方例子--获取天气预报【1】，来理解MCP落地的整条链路。

2. MCP示例

        该案例是构建一个简单的MCP天气预报服务器，并将其连接到主机，即Claude for Desktop。从基本设置开始，然后逐步发展到更复杂的使用场景。

        大模型虽然能力非常强，但其弊端就是内容是过时的，这里的过时不是说内容很旧，只是表达内容具有非实时性。比如没有获取天气预报和严重天气警报的能力。因此我们将使用MCP来解决这一问题。

        构建一个服务器，该服务器提供两个工具：获取警报（get-alerts）和获取预报（get-forecast）。然后，将该服务器连接到MCP主机（在本例中为Claude for Desktop）。

        首先我们配置下环境：

        （1）安装uv

curl -LsSf https://astral.sh/uv/install.sh | sh 

        安装完成后，会提示：

downloading uv 0.6.9 aarch64-apple-darwin
no checksums to verify
installing to /Users/nicolas/.local/bin
  uv
  uvx
everything's installed!       

      （2）安装所需的依赖包

        （3）在server.py中构建相应的get-alerts和 get-forecast工具：

from typing import Any import asyncio import httpx from mcp.server.models import InitializationOptions import mcp.types as types from mcp.server import NotificationOptions, Server import mcp.server.stdio NWS_API_BASE = "https://api.weather.gov" USER_AGENT = "weather-app/1.0" #@server.list_tools() - 注册用于列出可用工具的处理器 #@server.call_tool() - 注册用于执行工具调用的处理器 server = Server("weather") @server.list_tools() async def handle_list_tools() -> list[types.Tool]: """ List available tools. Each tool specifies its arguments using JSON Schema validation. """ return [ types.Tool( name="get-alerts", description="Get weather alerts for a state", inputSchema={ "type": "object", "properties": { "state": { "type": "string", "description": "Two-letter state code (e.g. CA, NY)", }, }, "required": ["state"], }, ), types.Tool( name="get-forecast", description="Get weather forecast for a location", inputSchema={ "type": "object", "properties": { "latitude": { "type": "number", "description": "Latitude of the location", }, "longitude": { "type": "number", "description": "Longitude of the location", }, }, "required": ["latitude", "longitude"], }, ), ] async def make_nws_request(client: httpx.AsyncClient, url: str) -> dict[str, Any] | None: """Make a request to the NWS API with proper error handling.""" headers = { "User-Agent": USER_AGENT, "Accept": "application/geo+json" } try: response = await client.get(url, headers=headers, timeout=30.0) response.raise_for_status() return response.json() except Exception: return None def format_alert(feature: dict) -> str: """Format an alert feature into a concise string.""" props = feature["properties"] return ( f"Event: {props.get('event', 'Unknown')}\n" f"Area: {props.get('areaDesc', 'Unknown')}\n" f"Severity: {props.get('severity', 'Unknown')}\n" f"Status: {props.get('status', 'Unknown')}\n" f"Headline: {props.get('headline', 'No headline')}\n" "---" ) @server.call_tool() async def handle_call_tool( name: str, arguments: dict | None ) -> list[types.TextContent | types.ImageContent | types.EmbeddedResource]: """ Handle tool execution requests. Tools can fetch weather data and notify clients of changes. """ if not arguments: raise ValueError("Missing arguments") if name == "get-alerts": state = arguments.get("state") if not state: raise ValueError("Missing state parameter") # Convert state to uppercase to ensure consistent format state = state.upper() if len(state) != 2: raise ValueError("State must be a two-letter code (e.g. CA, NY)") async with httpx.AsyncClient() as client: alerts_url = f"{NWS_API_BASE}/alerts?area={state}" alerts_data = await make_nws_request(client, alerts_url) if not alerts_data: return [types.TextContent(type="text", text="Failed to retrieve alerts data")] features = alerts_data.get("features", []) if not features: return [types.TextContent(type="text", text=f"No active alerts for {state}")] # Format each alert into a concise string formatted_alerts = [format_alert(feature) for feature in features[:20]] # only take the first 20 alerts alerts_text = f"Active alerts for {state}:\n\n" + "\n".join(formatted_alerts) return [ types.TextContent( type="text", text=alerts_text ) ] elif name == "get-forecast": try: latitude = float(arguments.get("latitude")) longitude = float(arguments.get("longitude")) except (TypeError, ValueError): return [types.TextContent( type="text", text="Invalid coordinates. Please provide valid numbers for latitude and longitude." )] # Basic coordinate validation if not (-90 <= latitude <= 90) or not (-180 <= longitude <= 180): return [types.TextContent( type="text", text="Invalid coordinates. Latitude must be between -90 and 90, longitude between -180 and 180." )] async with httpx.AsyncClient() as client: # First get the grid point lat_str = f"{latitude}" lon_str = f"{longitude}" points_url = f"{NWS_API_BASE}/points/{lat_str},{lon_str}" points_data = await make_nws_request(client, points_url) if not points_data: return [types.TextContent(type="text", text=f"Failed to retrieve grid point data for coordinates: {latitude}, {longitude}. This location may not be supported by the NWS API (only US locations are supported).")] # Extract forecast URL from the response properties = points_data.get("properties", {}) forecast_url = properties.get("forecast") if not forecast_url: return [types.TextContent(type="text", text="Failed to get forecast URL from grid point data")] # Get the forecast forecast_data = await make_nws_request(client, forecast_url) if not forecast_data: return [types.TextContent(type="text", text="Failed to retrieve forecast data")] # Format the forecast periods periods = forecast_data.get("properties", {}).get("periods", []) if not periods: return [types.TextContent(type="text", text="No forecast periods available")] # Format each period into a concise string formatted_forecast = [] for period in periods: forecast_text = ( f"{period.get('name', 'Unknown')}:\n" f"Temperature: {period.get('temperature', 'Unknown')}°{period.get('temperatureUnit', 'F')}\n" f"Wind: {period.get('windSpeed', 'Unknown')} {period.get('windDirection', '')}\n" f"{period.get('shortForecast', 'No forecast available')}\n" "---" ) formatted_forecast.append(forecast_text) forecast_text = f"Forecast for {latitude}, {longitude}:\n\n" + "\n".join(formatted_forecast) return [types.TextContent( type="text", text=forecast_text )] else: raise ValueError(f"Unknown tool: {name}") async def main(): # Run the server using stdin/stdout streams async with mcp.server.stdio.stdio_server() as (read_stream, write_stream): await server.run( read_stream, write_stream, InitializationOptions( server_name="weather", server_version="0.1.0", capabilities=server.get_capabilities( notification_options=NotificationOptions(), experimental_capabilities={}, ), ), ) # This is needed if you'd like to connect to a custom client if __name__ == "__main__": asyncio.run(main()) 

        这段代码中，最核心的其实就是@server.list_tools() 以及 @server.call_tool() 这两个注解。

@server.list_tools() - 注册用于列出可用工具的处理器
@server.call_tool() - 注册用于执行工具调用的处理器

        调用函数的逻辑也比较简单，匹配到对应的工具名称，然后抽取对应的输入参数，然后发起api的请求，对获得的结果进行处理：

async def main(): # Run the server using stdin/stdout streams async with mcp.server.stdio.stdio_server() as (read_stream, write_stream): await server.run( read_stream, write_stream, InitializationOptions( server_name="weather", server_version="0.1.0", capabilities=server.get_capabilities( notification_options=NotificationOptions(), experimental_capabilities={}, ), ), ) # This is needed if you'd like to connect to a custom client if __name__ == "__main__": asyncio.run(main())

      （4）服务端与客户端交互

        测试服务器与 Claude for Desktop。【2】也给出了构建MCP 客户端的教程。其中核心的逻辑如下：

async def process_query(self, query: str) -> str: """Process a query using Claude and available tools""" messages = [ { "role": "user", "content": query } ] response = await self.session.list_tools() available_tools = [{ "name": tool.name, "description": tool.description, "input_schema": tool.inputSchema } for tool in response.tools] # Initial Claude API call response = self.anthropic.messages.create( model="claude-3-5-sonnet-20241022", max_tokens=1000, messages=messages, tools=available_tools ) # Process response and handle tool calls final_text = [] assistant_message_content = [] for content in response.content: if content.type == 'text': final_text.append(content.text) assistant_message_content.append(content) elif content.type == 'tool_use': tool_name = content.name tool_args = content.input # Execute tool call result = await self.session.call_tool(tool_name, tool_args) final_text.append(f"[Calling tool {tool_name} with args {tool_args}]") assistant_message_content.append(content) messages.append({ "role": "assistant", "content": assistant_message_content }) messages.append({ "role": "user", "content": [ { "type": "tool_result", "tool_use_id": content.id, "content": result.content } ] }) # Get next response from Claude response = self.anthropic.messages.create( model="claude-3-5-sonnet-20241022", max_tokens=1000, messages=messages, tools=available_tools ) final_text.append(response.content[0].text) return "\n".join(final_text)

        启动客户端，需要打开 Claude for Desktop 应用配置文件：

~/Library/Application Support/Claude/claude_desktop_config.json

        如果该文件不存在，确保先创建出来，然后配置以下信息，以示例说明，我们uv init的是weather，所以这里mcpServers配置weather的服务，args中的路径设置为你weather的绝对路径。

{ "mcpServers": { "weather": { "command": "uv", "args": [ "--directory", "/ABSOLUTE/PATH/TO/PARENT/FOLDER/weather", "run", "server.py" ] } } }

        保存文件，并重新启动 Claude for Desktop。可以看到Claude for Desktop 能够识别在天气服务器中暴露的两个工具。

        然后在客户端询问天气，会提示调用get-forecast的tool：

3. MCP到底解决了什么问题

        工具是智能体框架的重要组成部分，允许大模型与外界互动并扩展其能力。即使没有MCP协议，也是可以实现LLM智能体，只不过存在几个弊端，当有许多不同的 API 时，启用工具使用变得很麻烦，因为任何工具都需要：手动构建prompt，每当其 API 发生变化时手动更新【3，4】。

        如下图所示：

        MCP其实解决了当存在大量工具时，能够自动发现，并自动构建prompt。

        整体流程示例：

      （1）以总结git项目最近5次提交为例，MCP 主机（与客户端一起）将首先调用 MCP 服务器，询问有哪些工具可用。

MCP 主机：像 Claude Desktop、IDE 或其他 AI 工具等程序，希望通过 MCP 访问数据。MCP 客户端：与服务器保持 1:1 连接 的协议客户端。

       （2）MPC 客户端接收到所列出的可用工具后，发给LLM，LLM 收到信息后，可能会选择使用某个工具。它通过主机向 MCP 服务器发送请求，然后接收结果，包括所使用的工具。

（3）LLM 收到工具处理结果（包括原始的query等信息），之后就可以向用户输出最终的答案。

总结起来，就一句话，MCP协议其实是让智能体更容易管理、发现、使用工具。

4. 参考材料

【1】For Server Developers - Model Context Protocol

【2】For Client Developers - Model Context Protocol

【3】AI Agent框架综述

【4】MCP工作原理