MCP 詳解

MCP 详解

AI 導讀

MCP 是連接 AI 應用程式與其使用的工具和資料的開放標準。伺服器列出可提供的功能，應用程式交由模型選擇；模型提出需求後，應用程式呼叫伺服器並將結果交回模型。MCP 不只是 JSON-RPC，亦定義其中的方法、能力協商、傳輸方式及驗證。

正文

Model Context Protocol

What is MCP?

An open standard for connecting AI apps to the tools and data they use. A server says what it can do, the app shows that to the model, and when the model wants something the app calls the server and hands the result back.

Host: the AI app you use: Claude, Cursor, ChatGPT, an agent
Server: a small program exposing tools, data or prompts
Model: decides what to use; never talks to a server directly
Not REST: JSON-RPC over stdin/stdout or one HTTP endpoint, and servers can talk back
Not just JSON-RPC: JSON-RPC is the envelope. MCP defines what goes in it: the methods, capability negotiation, transports and auth

or jump to Tools · the spec time machine · Break it

Spec

#	From	To	Method	Info