Editing Model Context Protocol (section)

== Architecture ==
MCP follows a '''client-server architecture''' with three main components:

* '''MCP Host''': The AI application or environment (e.g., Claude Desktop, ChatGPT, Cursor, Visual Studio Code with Copilot, or a custom agent) that contains the LLM.
* '''MCP Client''': Embedded within the host, it manages connections to one or more MCP servers and handles communication using the protocol (based on [[JSON-RPC]] 2.0).
* '''MCP Server''': A lightweight program that exposes capabilities from external systems. Servers can provide:
  ** '''Resources''' (e.g., files, database records)
  ** '''Tools''' (e.g., functions for calculations, API calls, code execution)
  ** '''Prompts''' (specialized workflows or contextual instructions)

The protocol supports two-way communication, allowing AI models to discover available capabilities, request data or actions, and receive formatted responses. It includes features for security (e.g., permissions and human-in-the-loop approvals) and supports both local and remote servers.<ref name="mcp-site" /><ref name="a16z">[https://a16z.com/a-deep-dive-into-mcp-and-the-future-of-ai-tooling/ A Deep Dive Into MCP and the Future of AI Tooling], a16z, March 20, 2025.</ref>