What is an MCP server?

What is MCP in AI?

MCP stands for Model Context Protocol. In AI products it is the open way for a host application to give a language model safe, structured access to capabilities you control—lookup, search, run a script, call an internal API—without baking every integration into the model weights or pasting secrets into the chat.

Three roles matter:

• Host — the app the user talks to (Claude Desktop, Cursor, Claude Code, VS Code with MCP, Windsurf, Codex, and other MCP clients).
• Model — the LLM that decides when a tool might help and which arguments to pass.
• MCP server — the small program you write (or generate) that implements the tools. The host launches it and speaks MCP to it; the model never talks to your server directly.

That split is why MCP shows up in “what is MCP in AI” and “what’s an MCP server” searches: people want the protocol idea first, then a concrete server they can run next to Cursor or Claude Desktop.

How an MCP server works

At runtime the loop is short and deliberate. Understanding tools/list and tools/call is enough to read most MCP docs and generator output.

1. Launch. The host starts your server as a child process (stdio) or connects to a remote endpoint (Streamable HTTP). Local generators on this site write stdio servers.
2. Initialize. Host and server exchange protocol version and capabilities.
3. tools/list. The host asks for every tool: name, description, and a JSON Schema (inputSchema) for arguments. Those descriptions are written for the model, not for a human API consumer.
4. Model chooses. When the user asks something that needs your capability, the model picks a tool name and fills arguments that match the schema.
5. tools/call. The host sends the call to your server. Your handler runs and returns text content (and optionally structured content). The host feeds that result back into the conversation.

stdio vs Streamable HTTP

A tested walkthrough of the same tools over both transports is in the TypeScript and Python MCP server template. For connecting a finished server to five hosts, see connect an MCP server to Claude Desktop, Cursor and Claude Code.

MCP server vs API

An API is a contract for clients you write. An MCP server is a contract for the model: tool names and descriptions are written so the LLM can choose when to call them. You can wrap an existing REST API as MCP tools, or expose a script or database query that never had an OpenAPI spec.

Side-by-side on one example API (form-built tools vs OpenAPI-converted tools): MCP Server Generator vs OpenAPI generators.

What an MCP server can expose

The protocol defines more than tools. Most first servers ship tools only; resources and prompts are optional but useful once you know the basics.

Tools

Tools are callable actions. Each has a name, a description the model reads, and an inputSchema (JSON Schema). Optional annotations (for example read-only or destructive hints) and an outputSchema help hosts and models use the tool safely. Design and validate schemas with the free MCP tool schema generator.

Example shape the host sees from tools/list:

{
  "name": "get_weather",
  "description": "Get current weather for a city",
  "inputSchema": {
    "type": "object",
    "properties": {
      "city": { "type": "string", "description": "City name, e.g. San Francisco" }
    },
    "required": ["city"]
  }
}

Resources

Resources are readable data the host can fetch by URI (for example a file, a config blob, or a curated document). They are not “called” with arguments the way tools are; they are listed and read. Use them when the model needs reference material more than an action.

Prompts

Prompts are reusable prompt templates the host can offer to the user or inject into a workflow. They package instructions you want applied consistently, separate from tool execution.

Working TypeScript and Python examples of a resource and a prompt (alongside tools) are in the MCP server template walkthrough.

How to get an MCP server without OpenAPI

You do not need an OpenAPI or Swagger file to create an MCP server. Many useful first servers are three or four task tools around a script, a database query, or a private endpoint that was never published as a spec.

• Form → ZIP. On the MCP Server Generator homepage, name the server, add tools with typed parameters, choose TypeScript or Python, and download a stdio project with README host config. Handlers start as stubs—replace the TODOs.
• Step-by-step tutorial. How to create an MCP server without OpenAPI walks through SDK, tools, handlers, stdio, test, and host config, then the form path.
• Template walkthrough. MCP server example and template for TypeScript and Python shows every generated file and how to add tools, resources and prompts.

Prefer someone else to build it? See the MCP server development service.

Already have an OpenAPI spec?

If you maintain OpenAPI 3 or Swagger 2.0, converting selected operations into MCP tools that send real HTTP requests is often faster than rewriting every handler by hand.

• OpenAPI to MCP converter — paste, upload or fetch a spec; download a TypeScript or Python server with one tool per selected operation.
• Form generator vs OpenAPI — same example API, compared on tool count, naming, auth and real HTTP handlers.

Convert, then prune: long lists of endpoint-shaped tools make it harder for the model to choose correctly. Expose only what the agent needs.

Connect, configure, schema and test

Once you have a server project:

1. Connect & config — Connect an MCP server to Claude Desktop, Cursor and Claude Code covers paths for five hosts. The MCP config generator and validator writes or checks claude_desktop_config.json, Cursor / VS Code mcp.json, Claude Code .mcp.json and Codex config.toml.
2. Schema — Use the MCP tool schema generator for richer inputs (enums, arrays, annotations, optional outputSchema) and to validate pasted tools/list JSON.
3. Test — Before touching a host, run how to test an MCP server with MCP Inspector (web UI and CLI), failure paths and a CI smoke pattern.

When to use a generator vs write by hand

FAQ