What is an MCP server?
An MCP server is a program that gives AI applications access to one system through the Model Context Protocol. It exposes tools the model can call, resources the app can read and prompts a person can pick, and answers JSON-RPC requests from an MCP client. It can run as a local process or as a service at a URL.
Part of Glossary: the terms behind MCP, coding agents and shared pages, defined plainly.
A client reaches the server: it launches it as a subprocess and talks over stdin and stdout, or sends HTTP POSTs to its one MCP endpoint.
The client may call server/discover first. The server replies with the protocol versions it supports, its capabilities (tools, resources, prompts) and its name.
The client lists what the server offers, such as tools/list. Each tool comes with a description and a JSON Schema for its arguments.
On tools/call the server does the work against the real system and returns content, or a result marked isError so the model can retry.
If the server needs something from the person or the model mid-call, it returns an input_required result and the client retries with the answer.
Source: MCP specification: architecture
MCP server versus the system behind it
An MCP server is rarely the system itself. It is an adapter that sits in front of a database, a filesystem, a SaaS API or a shared document and translates MCP messages into whatever that system understands. That is why the specification's first design principle is that servers should be extremely easy to build: the host app carries the orchestration, consent and model handling, and the server focuses on a small, well-defined set of capabilities.
The split has a consequence people underestimate. A server never sees the conversation. It receives only the request in front of it, so it cannot read what the user said three turns ago or what another server returned. Since the 2026-07-28 revision it also cannot assume anything from the previous request on the same connection: the protocol is stateless, and a server that needs to carry state between calls returns an explicit handle, such as a basket id, that the model passes back on the next call.
When to write one
Write an MCP server when you want many AI apps to reach one system without a plugin for each. Expose a few focused tools rather than mirroring a whole API: the model has to read every description, and a long list costs context on every turn. Give tools clear names, a one-paragraph description and strict input schemas. If the system needs sign-in, serve it over Streamable HTTP and use the protocol's OAuth-based authorization; a stdio server should take credentials from its environment instead.
Common mistakes
- Writing to stdout. A stdio server must write nothing to stdout except MCP messages. Logs go to stderr, or the client will fail to parse the stream.
- Relying on connection state. A server that stores "the current user" from an earlier request breaks under load balancing and breaks the spec. Read everything from the request, including the authorization it carries.
- Skipping input validation. The specification requires servers to validate all tool inputs, apply access controls, rate limit calls and sanitize outputs. A model will send arguments a human never would.
FAQ
- Does an MCP server need to run on my computer?
- No. A local server runs as a subprocess your client starts and talks to over stdio. A remote server runs elsewhere and you reach it at a URL over Streamable HTTP. Both speak the same protocol.
- Can one MCP server serve many clients at once?
- A remote server can: the specification says a Streamable HTTP server operates as an independent process that can handle multiple client connections. A stdio server is started per client, as a child process.
- What does an MCP server have to implement at minimum?
- The base protocol, versioning, the message patterns, and server/discover. Everything else (tools, resources, prompts, authorization) is optional and advertised through capabilities.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.