What is MCP Server

MCP(Model Context Protocol)是 Anthropic 提出的开放Protocol,用于标准化 AI 模型与外部工具、数据源之间的通信方式。MCP Server 是实现该Protocol的服务端程序,为 AI Clients提供工具调用能力。

Core Concepts

MCP defines three roles:

A Host can connect to multiple MCP Servers simultaneously, each providing different tool sets.

Differences Between MCP Server and Regular Plugins

ComparisonMCP ServerRegular Plugins
ProtocolStandardized JSON-RPC protocolPlatform-specific private APIs
ReusabilityOne Server can be used by multiple clientsUsually only works with specific platforms
RuntimeIndependent process, communicating via stdio or HTTPUsually runs within the host process
LanguageAny language (Node.js, Python, etc.)Limited to platform SDK
Security BoundaryProcess isolation, controllable permissionsUsually shares host permissions

How It Works

The MCP communication flow is as follows:

  1. When the Host starts, it launches MCP Server child processes based on the configuration file.
  2. The Client establishes a JSON-RPC connection with the Server via stdio (standard input/output).
  3. The Server declares its available tools and resources to the Client.
  4. During user-AI conversations, the AI calls Server-provided tools as needed.
  5. The Server executes operations and returns results, which the AI integrates into its response.
用户 → Host(Claude Desktop) → Client → MCP Server → 外部服务/文件系统
                                                    ↓
                                              Returns execution results

What Can MCP Server Do

Important Notes

FAQ

Is MCP proprietary to Anthropic?
MCP is an open protocol with open-source specifications and SDKs. Any AI client can implement MCP support. Currently Claude Desktop, Cursor, Windsurf, and others support it.
Do I need programming skills to use MCP?
Basic usage only requires editing JSON configuration files, no coding needed. However, understanding command-line operations and environment variables is helpful.
Does MCP Server auto-update?
No. MCP Servers run via npx or uvx will pull the latest version, but locally installed ones need manual updates. We recommend regularly checking official repository changelogs.

Related Links