Minimal Client Weather Server Sample

MCP（模型上下文协议）客户端-服务器交互示例

一个简单的天气工具，演示使用模型上下文协议（MCP）的服务器-客户端交互。仅用于演示目的。

用户：“东京的天气怎么样？”\ 回复：“东京的天气晴朗。”

图1：初始化和工具发现

sequenceDiagram
    autonumber

    participant User
    participant ClientApp as ClientApp (Host)
on Local PC
    participant LLM as LLM (e.g., Claude)
Remote Service
    participant MCPClient as MCPClient (Internal Component)
in ClientApp on Local PC
    participant MCPServer as MCPServer (e.g., Tool Server)
on Local PC

    Note over User, MCPServer: User starts ClientApp, initiating connection to MCPServer

    ClientApp->>+MCPClient: Instruct preparation to connect to a specific MCPServer
    Note right of ClientApp: Host manages MCPClient instances per server

    MCPClient->>+MCPServer: 1. initialize (Request)
[protocol_version, client_capabilities]
    Note over MCPClient, MCPServer: Start connection establishment and capability exchange (JSON-RPC)
    MCPServer-->>-MCPClient: 2. initialize (Response)
[selected_protocol_version, server_capabilities (tools, resources, etc.)]
    Note over MCPClient, MCPServer: Server notifies its available capabilities

    MCPClient->>MCPServer: 3. notifications/initialized (Initialization Complete Notification)
    Note over MCPClient, MCPServer: Handshake complete, normal communication possible

    MCPClient-->>-ClientApp: Initialization Success & Server Capabilities notified (implicitly via successful initialize await)
    Note right of ClientApp: Host (ClientApp) now knows the server is ready

    ClientApp->>MCPClient: 4. Instruct to get the list of tools provided by the server (session.list_tools())
    MCPClient->>+MCPServer: 5. tools/list (Request)
    Note over MCPClient, MCPServer: Request the list of tools defined on the server
    MCPServer-->>-MCPClient: 6. tools/list (Response)
[{"name": "get_weather", "description": "...", "inputSchema": {...}}]
    Note over MCPClient, MCPServer: Example: Returns the definition of the 'get_weather' tool

    MCPClient-->>ClientApp: 7. Notify Tool List (containing 'get_weather', etc.) via tools_response object
    Note right of ClientApp: Host stores/processes the retrieved tool information (format_tools_for_llm)

    ClientApp->>LLM: 8. Send available tool information to LLM
(e.g., via system prompt using formatted tools_prompt)
    Note over ClientApp, LLM: Host informs LLM that 'get_weather' is available
LLM can now decide to use the tool based on this info

图2：工具执行流程

sequenceDiagram
    autonumber

    participant User
    participant ClientApp as ClientApp (Host)
on Local PC
    participant LLM as LLM (e.g., Claude)
Remote Service
    participant MCPClient as MCPClient (Internal Component)
in ClientApp on Local PC
    participant MCPServer as MCPServer (e.g., Tool Server)
on Local PC

    User->>ClientApp: 1. "What's the weather in Tokyo?"
    Note right of User: User asks a question that might require a tool

    ClientApp->>LLM: 2. Forward user's question with context
[System Prompt (with tool info) + User Question]
    Note over ClientApp, LLM: Host sends the question and formatted tool info ('get_weather') to the LLM

    LLM->>ClientApp: 3. LLM responds, requesting tool usage
Response: `{"tool_name": "get_weather", "arguments": {"location": "Tokyo"}}`
    Note over ClientApp, LLM: Based on provided tool info, LLM decides to use the weather tool
and generates the required JSON structure with arguments

    ClientApp->>User: (Optional) 4. Confirm tool execution
"Execute 'get_weather' on Tool Server?"
    User->>ClientApp: (Optional) 5. "Yes"
    Note over ClientApp, User: For security/transparency, host might seek user permission (Not implemented in the sample code)

    ClientApp->>+MCPClient: 6. Instruct execution of 'get_weather' tool (session.call_tool())
Arguments: {"location": "Tokyo"}
    Note right of ClientApp: Host instructs the MCP server via MCPClient

    MCPClient->>+MCPServer: 7. tools/call (Request)
[name: "get_weather", arguments: {"location": "Tokyo"}]
    Note over MCPClient, MCPServer: Invoke server function with specified tool name and arguments (JSON-RPC)

    MCPServer->>MCPServer: Internal processing (executes get_weather function)
    Note over MCPServer: Logs execution, returns result (e.g., "Sunny")
    MCPServer-->>-MCPClient: 8. tools/call (Response)
[result: {content: [{type: "text", text: "Sunny"}]}]
    Note over MCPClient, MCPServer: Server returns the processing result (weather info) in structured format

    MCPClient-->>-ClientApp: 9. Notify Tool Execution Result
Result object contains: "Sunny" (extracted from tool_result_obj.content[0].text)

    ClientApp->>LLM: 10. Send tool execution result back to LLM
New User Message: "Tool 'get_weather' returned: 'Sunny'. Answer the original question based on this."
    Note over ClientApp, LLM: Host feeds back the information obtained from the server to the LLM

    LLM->>ClientApp: 11. Generate final response (as text)
"The weather in Tokyo is Sunny."

    ClientApp->>User: 12. Display the final response generated by LLM
    Note right of ClientApp: Present the final answer to the user