TL;DR
The Model Context Protocol (MCP), introduced by Anthropic in 2024, is an open standard that connects AI applications to external tools and data sources — often described as "USB-C for AI." One MCP server can be reused by many model clients.
Key Points
- Three roles:
- Host: the main application (e.g., Claude Desktop, Codex, IDEs);
- Client: the in-host component that connects to servers;
- Server: a lightweight program exposing tools, resources, and prompts.
- Transports: stdio (local subprocess) and HTTP/SSE (remote); spec maintained at modelcontextprotocol.io.
- Capabilities: tools (actions), resources (readable data), prompts (reusable templates), exchanged over JSON-RPC.
- Why it matters: previously every agent needed bespoke integrations per tool; MCP standardizes the interface and cuts integration cost.
Status
- Widely supported by coding agents, IDEs, and frameworks.
- Remote servers can use OAuth, making them production-ready.
- The ecosystem is moving toward discoverability and remote services.
Sources
- Spec: modelcontextprotocol.io
- Wikipedia: Model Context Protocol
- Anthropic announcement (2024-11)