What is an MCP server? A plain-English guide for 2026

MCP server, defined

An MCP server is a small program that exposes a set of tools(callable functions) and resources (read-only context) over the Model Context Protocol — an open specification, originally published by Anthropic, that any AI client can speak. Claude, Cursor, ChatGPT, and custom widgets all use the same wire format, so one MCP server works with every client without bespoke glue.

The mental model: REST is for code that already knows your schema. MCP is for an LLM that will discover your tools at runtime, read their descriptions, and pick the right one for the user's request.

Anatomy of an MCP server

• Tools — named actions with JSON-schema inputs (create_issue, send_email).
• Resources — read-only blobs the LLM can fetch by URI (a doc page, a row, a file).
• Prompts — optional reusable prompt templates the host can offer the user.
• Transport — stdio for local clients, HTTP+SSE or Streamable HTTP for remote.

Why MCP matters

Before MCP, every AI integration was bespoke: OpenAI function calling looked one way, Anthropic tool use another, every IDE had its own plugin shape. MCP collapses that to one protocol. Write the server once; every model and every client gets it for free. For the full breakdown of where the two layers actually differ, see MCP vs function calling.

For the strategic version of that argument — why this is worth doing before a competitor ships it, not after — see Why MCP matters.

A minimal example

A simplest MCP server exposes one tool. In TypeScript with the official SDK:

server.tool("get_weather", {
  description: "Get current weather for a city",
  inputSchema: z.object({ city: z.string() }),
  handler: async ({ city }) => fetchWeather(city),
});

That tool is now callable by Claude, Cursor, ChatGPT, or any custom widget — no per-client adapter code.

Three on-ramps to your first server

1. Convert an OpenAPI spec — fastest if you already have a REST API.
2. Wrap existing endpoints — best when you want control over which tools the LLM sees.
3. Use a hosted server — Linear, GitHub, Postgres, Slack already have production MCP servers you can connect to.

See Convert your REST API for the OpenAPI flow.

Production checklist

• Auth: short-lived JWTs or OAuth per end-user. Never share a long-lived key across users.
• Scopes: each tool checks the caller's permissions, same as your REST API would.
• Audit: log every tool call with request ID, user ID, inputs (redacted), outcome.
• Quotas: a gateway in front for per-user rate limits and cost caps.
• Observability: latency and error rate per tool, not just per server.

When you don't need an MCP server

If your AI feature is a pure summarization or classification job with no external action, you don't need MCP — a direct model call is fine. The moment the LLM needs to do something on behalf of a user (create a record, search a database, send a message), MCP is the right shape.