跳转到主要内容

跨模型、工具、服务、公司、提供商、API、套餐、文章和新闻进行搜索。

MCP 协议入门 10 分钟

什么是 MCP(模型上下文协议)?2026 指南

简明而前沿的讲解:主机/客户端/服务器架构、三种能力类型,以及 MCP 为何被称为「AI 界的 USB-C」。

什么是 MCP(模型上下文协议)?2026 指南

The problem MCP solves

Before MCP, every AI application had to build a custom connection to every data source — files, databases, GitHub, Slack. It was the pre-USB-C world where every device shipped its own cable. The Model Context Protocol is an open standard: a tool advertises itself via MCP once, and every compatible application can use it.

Three roles in the architecture

  • Host: the application the user works in — say, a code editor with an AI assistant.
  • Client: a connector inside the host that keeps a one-to-one session with each server.
  • Server: a small program that exposes one source's capabilities in the standard format.

One application can talk to many servers; servers stay independent of each other.

Three kinds of capability

  • Tools: actions the model can execute (create an issue, run a query, send a message).
  • Resources: data the model can read (files, records, content).
  • Prompts: ready-made templates the server author defines for common tasks.

Why it matters

For users: your AI assistant connects to everyday tools with no per-app setup. For developers: instead of ten custom integrations, you write one standard implementation that works everywhere. For security: the trust boundary is explicit — the server decides what is visible and executable, and the client asks the user for consent.