What is MCP server?

In the MCP architecture, a server is the party that supplies context and capabilities. It exposes three kinds of primitives: resources, tools and prompts. A server might wrap a file system and Git, a database, or an external API, and each one is meant to have a focused job. The specification describes servers as running either as local processes or as remote services.

Each server is connected to exactly one client, and the client sits inside a host application. The host coordinates the connection to the language model and holds the full conversation, so a server only receives the context it needs and cannot see into other servers. The architecture page lists this isolation as a design principle, along with the goals that servers should be easy to build and easy to combine.

A server advertises what it supports as capabilities, for example tool support, resource subscriptions and prompt templates. Clients can ask for these through the server/discover request, and a server must declare a feature before clients can rely on it.

When a server needs something from the client side mid-request, such as an LLM completion, a list of roots or an answer from the user, it returns an InputRequiredResult in its reply. The client then retries the request with the answer attached.