## What is MCP [What is the Model Context Protocol (MCP)?](https://modelcontextprotocol.io/docs/getting-started/intro) [Model Context Protocol](https://developers.openai.com/codex/mcp/) [Architecture overview](https://modelcontextprotocol.io/docs/learn/architecture) [Build your MCP server](https://developers.openai.com/apps-sdk/build/mcp-server/) []() ## 基本概念 ### MCP Host (运行 LLM 并管理 MCP Server 的程序) - 职责 1. 启动 MCP Server 2. 发现可用 tools 3. 把 tools 提供给 LLM 4. 当 LLM 需要工具时调用 MCP Server 5. 把结果返回给 LLM ``` User │ ▼ Cursor (MCP Host + LLM) │ ├─ 管理 MCP Servers ├─ 提供 tools 给 LLM └─ 调用 tools ``` ### MCP Server (提供工具能力的进程) - 职责 向 MCP Host 提供工具(tools / resources / prompts) ``` Cursor (MCP Host) │ │ JSON-RPC ▼ mcp-atlassian (MCP Server) │ ▼ Jira / Confluence ``` #### MCP Server 的运行模式 1. 本地进程 ``` Cursor │ ▼ local MCP server mode: stdin/stdout ``` 2. HTTP MCP server ``` Cursor │ ▼ remote MCP server mode: HTTP ``` 3. container MCP server ``` Cursor │ ▼ Docker MCP server ``` ### MCP Host 和 MCP Server 的关系 ``` Host = orchestrator Server = plugin MCP Host (e.g Cursor / IDE) │ ┌──────────┼──────────┐ ▼ ▼ ▼ MCP Server MCP Server MCP Server GitHub Atlassian Slack 每个 server 提供不同能力 ``` ### MCP 的真实运行流程 ``` User │ ▼ Cursor (Host) │ │ 1 发现 tools │ ▼ LLM │ │ 2 选择 tool │ ▼ Cursor │ │ 3 tools/call ▼ MCP Server │ │ 4 调用外部API ▼ External System │ │ 5 返回结果 ▼ Cursor │ │ 6 LLM生成最终回答 ▼ User ``` ## 整体机制 当 Cursor 启动 MCP Server 后,它会: 1. 向 MCP Server 请求工具列表 2. 将工具信息转换为 LLM可理解的函数定义 3. 将这些定义 附加到系统 prompt 中 4. LLM 推理是否调用工具 ``` User Prompt │ ▼ Cursor (MCP Host) │ ├─ tools/list → MCP Server │ ▼ Tool Definitions │ ▼ LLM Prompt Injection │ ▼ LLM decides: call tool? ``` - Cursor Prompt 的真实结构 ``` SYSTEM PROMPT -------------- You are an AI coding assistant. You have access to the following tools: Tool: get_jira_issue Description: Retrieve a Jira issue by key Parameters: issue_key (string) Tool: search_confluence Description: Search Confluence pages. Rules: - Use tools when needed - Do not fabricate data -------------------------------- CHAT HISTORY -------------------------------- User: Summarize Jira issue VEC-123 -------------------------------- TOOLS -------------------------------- [get_jira_issue definition] [search_confluence definition] ``` > tools 被 作为 metadata 提供给模型 - Cursor Tool Calling 时序 ``` User │ ▼ Cursor │ │ inject tools ▼ LLM │ │ choose tool ▼ Cursor │ │ tools/call ▼ MCP Server │ │ fetch data ▼ Cursor │ │ add result to context ▼ LLM │ │ final answer ▼ User ``` ai/mcp/ ├── mcp_llm_tool │ ├── [README.md](./ai/mcp/mcp_llm_tool/README.md) │ ├── [api_server.py](./ai/mcp/mcp_llm_tool/api_server.py) │ ├── [config.py](./ai/mcp/mcp_llm_tool/config.py) │ ├── [llm_client.py](./ai/mcp/mcp_llm_tool/llm_client.py) │ ├── [main.py](./ai/mcp/mcp_llm_tool/main.py) │ ├── [mcp_base.py](./ai/mcp/mcp_llm_tool/mcp_base.py) │ ├── [requirements.txt](./ai/mcp/mcp_llm_tool/requirements.txt) │ └── [text_analyzer.py](./ai/mcp/mcp_llm_tool/text_analyzer.py) ├── [mcp_llm_tool_part1_concepts.md](./ai/mcp/mcp_llm_tool_part1_concepts.md) ├── [mcp_llm_tool_part2_implementation.md](./ai/mcp/mcp_llm_tool_part2_implementation.md) └── [mcp_llm_tool_part3_deployment_testing.md](./ai/mcp/mcp_llm_tool_part3_deployment_testing.md) ## MCP Market [1](https://mcp.so/) [2](https://mcpmarket.com/) [3](https://smithery.ai/) []() ## Related Tools - uvx - npx ## Common MCP Servers > Web content fetching and conversion for efficient LLM usage - fetch (https://mcp.so/server/fetch/test) - spec-workflow-mcp (https://github.com/Pimzino/spec-workflow-mcp) - -