Model Context Protocol (MCP)

Core Concept

MCP is an open protocol that standardizes how applications provide context to LLMs. Think of MCP like a USB-C port for AI applications. Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools.

Purpose & Use Cases

MCP enables developers to connect AI models to custom tools and data sources in a standardized, interoperable manner—addressing a key need in extending Claude's capabilities beyond its base training.

MCP in Anthropic Products

The protocol integrates across multiple Anthropic offerings:

Messages API - Developers can use the MCP connector to link to MCP servers programmatically
Claude Code - Supports adding custom MCP servers or using Claude Code as a server itself
Claude.ai - Teams can enable MCP connectors through the platform
Claude Desktop - Allows local MCP server configuration

Build Your Own MCP Products

For protocol details, server/client development guidance, and community-created implementations, refer to the official MCP Documentation at modelcontextprotocol.io.

MCP in Claude Code

MCP Tool Integration

MCP tools follow the pattern: mcp__<server>__<tool>

Using MCP in Hooks

Configure MCP tools with regex matchers in hooks:

{
  "matcher": "mcp__memory__.*",
  "hooks": [{"type": "command", "command": "validate.py"}]
}

MCP Slash Commands

MCP servers expose prompts as commands with the pattern:

/mcp__<server-name>__<prompt-name> [arguments]

These are dynamically discovered from connected MCP servers and automatically available when the server is active.

Managing MCP Servers

Use the /mcp built-in command to manage MCP server connections.

Plugin Integration

Plugins can include MCP servers via .mcp.json configuration for external tool integration.

Resources

Official MCP Documentation: modelcontextprotocol.io
Anthropic's MCP implementation guides
Community-created MCP servers and implementations