MCP (Model Context Protocol)
An open protocol for exposing tools, data and prompts to an AI client through one standard interface.
MCP lets a server publish capabilities — tools to call, resources to read, prompts to reuse — that any compatible client can consume. Its purpose is to stop every integration being bespoke: instead of writing one adapter per assistant, a system exposes an MCP server once and every client can use it.
Key facts
- MCP separates the client from the integration, so one server can serve any compatible AI client without per-client code.
- A server exposes three kinds of thing: tools the model can call, resources it can read, and prompt templates it can reuse.
- MCP servers run as separate processes and communicate over stdio or HTTP, so a crash in a server does not take down the client.
- Tools from an MCP server occupy the same context budget as native tools; connecting many servers shrinks the room left for conversation.
Frequently asked questions
How is MCP different from just writing a tool?
A native tool is bound to the application you wrote it in. An MCP server is reusable: the same server works with any compatible client, and the client does not need to know anything about your system beyond the protocol.
Does connecting more MCP servers make an agent more capable?
Only up to a point. Every connected tool is described on every request, so a large catalogue costs context and makes tool selection harder. Connecting what the task needs beats connecting everything available.
Machine-readable copy of this page:
/glossary/mcp.md