The spec defines three server features. Tools are functions the model can choose to call, each with a name, a description and a JSON Schema for its inputs. A tool can also declare an output schema and return structured results. Resources are data identified by a URI, such as files or database schemas, and the client application decides how to use them. Prompts are templates that the user picks, often as slash commands.
There are two standard transports. With stdio, the client launches the server as a subprocess and they exchange newline-delimited JSON-RPC messages; the server may log to stderr but must write only MCP messages to stdout. With Streamable HTTP, the server exposes one HTTP endpoint and the client sends each message as its own POST, answered with a single JSON object or a stream scoped to that request. The current spec describes MCP as stateless: each request carries what the server needs, and state that spans requests is passed as an explicit identifier. Earlier revisions allowed server-assigned sessions, so a server may need to handle older clients too.
Our own production MCP server, for a database-performance SaaS, runs on its own hostname as a stateless HTTP endpoint with no session store. It uses the same analysis pipeline, rate limits and billing as the web app, and it handles long-running analyses and repeated requests so an assistant does not cause duplicate work or charges.