MCP (Model Context Protocol)

核心 · Key Idea

In one line: MCP = Model Context Protocol — Anthropic-led open protocol between "model ↔ outside world." It exposes three resource types (tools, data, prompts) through a unified interface, so any MCP server can plug into any MCP client (Claude Desktop / IDE / your own agent).

What it is#

Before MCP every product invented its own tool protocol (OpenAI's, Anthropic's, your platform's). MCP standardises:

Tools — callable functions
Resources — readable data (files, DB rows)
Prompts — reusable prompt templates

A server exposes these; a client connects via stdio or WebSocket and the model immediately gains the new capability.

Analogy#

打个比方 · Analogy

Before: every laptop had a proprietary charging port — a new adapter for every brand.
MCP: Type-C — one cable, plug it anywhere.
For models this means: write a tool once, use it everywhere.

Key concepts#

MCP ServerMCP server

A process that exposes Tools / Resources / Prompts. Can be local or remote-hosted.

MCP ClientMCP client

An MCP-aware host — Claude Desktop / Cursor / Cline / your own agent.

TransportTransport

stdio (local subprocess) or SSE / WebSocket (remote). Messages are JSON-RPC 2.0.

Tool / Resource / PromptThree resource kinds

Tools are executable; Resources are readable (files, DB); Prompts are pre-defined templates.

How it works#

The model reaches any server via Host → Client → Server. Add one server and the whole ecosystem inherits the capability.

Practical notes#

Spinning up a local server is trivial. npm i -g mcp-server-filesystem and one line of claude_desktop_config.json lets Claude read your machine.
Official / community servers cover a lot. GitHub, Slack, Postgres, Notion, Brave Search, Puppeteer — try off-the-shelf before writing your own.
Use the official SDK to write servers. mcp (Python) / @modelcontextprotocol/sdk (TS) — a few dozen lines and you're up.
Scope permissions tightly. MCP servers run locally and can read files / call APIs — principle of least privilege, otherwise a malicious prompt-injection can pivot through them.
Not all-or-nothing. Stay on Function Calling internally, and adopt MCP for the surfaces you want to open up.

Easy confusions#

MCP

A **vendor-neutral protocol**.
One server, every host.

Function Calling

A **per-model schema**.
Tools and model are tightly coupled.

MCP

**Designed for LLMs**: tool descriptions in natural language.
Native semantics for Resources and Prompts beyond pure functions.

OpenAPI / REST

APIs designed for human engineers / programs.
No LLM-friendly capability-discovery mechanism.