MCP (Model Context Protocol)
Also: Model Context Protocol · 模型上下文協議 · MCP server · MCP 協議
An open standard for exposing tools, resources and prompt templates to AI clients — write the server once, and every MCP-compatible client can use it.
When you will meet it
You meet it when you connect the second and third external system — a database, GitHub, a browser — to your agent. Without a standard the world is N×M: N clients each write bespoke glue for M services. MCP compresses that to N+M: each service writes one server, each client implements the protocol once. Missing this, you write every integration as one-off glue and drown by the fifth.
An analogy
Like USB-C. Before it, every brand had its own charging port and travel meant a bag of adapters; one standard connector replaced the bag. MCP aims to be the USB-C of AI tooling — the protocol itself provides no capability, it only defines the shape of the plug.
Minimal example
// 在客戶端設定裡註冊一個 MCP server（示意）
{
"mcpServers": {
"postgres": {
"command": "npx",
"args": ["@anthropic-ai/mcp-server-postgres",
"postgresql://localhost/mydb"]
}
}
}
// 重啟客戶端後，模型的工具清單裡就多了「查詢這個資料庫」的能力Note the server is a separate process running with its own permissions and credentials — the database connection string sits right in the config. Installing an MCP server is therefore not a light config line: you are plugging your agent into something that touches real systems (see supply-chain-risk).
What people get wrong
- Thinking MCP is a model or a framework. It is a protocol defining how client and server talk (JSON-RPC). Capability comes from what the server is wired to, not from the protocol.
- Assuming MCP handles permissions. The protocol defines how to connect, not whether you should, or how much may be touched. Least privilege, sandboxing and auditing all still apply — a filesystem server should never point at /.