What is an MCP server? A plain definition with examples

An MCP server is a program that gives AI applications extra capabilities through the Model Context Protocol (MCP), an open-source standard for connecting AI apps to outside systems. A server can offer tools the model calls, resources the app reads as context, and prompt templates the user picks. An AI app such as Claude, ChatGPT, VS Code, or Cursor connects to the server, asks what it offers, and uses it when a task needs it.

The definitions come from the Model Context Protocol's own docs: What is MCP?, Architecture overview, Understanding MCP servers, and Example servers, read on 2026-09-28. The worked example is Sume's hosted MCP server, described in Sume's MCP overview.

How does an MCP server work?

MCP has three roles. The host is the AI application, such as Claude Code or Claude Desktop. For each server it connects to, the host creates one MCP client, which keeps a dedicated connection to that server. The server is the program on the other end, and the two sides exchange JSON-RPC 2.0 messages. The server only offers capabilities; the model, the app, or the user decides when to use them, which is how an MCP server differs from an AI agent.

A session with tools runs in three steps, and the example after them is a tool definition from MCP's docs:

• Discover: the client asks what the server offers. For tools, a tools/list request returns each tool's name, description, and inputSchema, a JSON Schema for its arguments.
• Collect: the app merges the tools from every connected server into one registry the model can use.
• Call: when the model decides to use a tool, the app routes the call to the right server with tools/call and hands the result back to the model.

{
  "name": "searchFlights",
  "description": "Search for available flights",
  "inputSchema": {
    "type": "object",
    "properties": {
      "origin": { "type": "string", "description": "Departure city" },
      "destination": { "type": "string", "description": "Arrival city" },
      "date": { "type": "string", "format": "date", "description": "Travel date" }
    },
    "required": ["origin", "destination", "date"]
  }
}

What can an MCP server offer?

Three kinds of building blocks, each controlled by someone different. A server can offer any mix of them; MCP tools vs resources vs prompts covers how to choose among them.

What is an MCP server used for?

Connecting an AI app to a system it can't reach on its own. MCP's docs name file system servers for document access, database servers for data queries, GitHub servers for code management, Slack servers for team communication, and calendar servers for scheduling. The MCP project also maintains reference servers that show the protocol at work:

• Filesystem: secure file operations with configurable access controls.
• Git: tools to read, search, and manipulate Git repositories.
• Fetch: web content fetching and conversion for LLM use.
• Memory: a knowledge graph-based persistent memory system.
• Time: time and timezone conversion.
• Everything: a reference and test server with prompts, resources, and tools.

Where does an MCP server run?

On your machine or somewhere else: MCP's docs call the program a server regardless of where it runs. A server that the app launches on the same machine and talks to over standard input and output (the stdio transport) is commonly called a "local" MCP server, and it typically serves one client. A server hosted elsewhere and reached over the Streamable HTTP transport is called a "remote" MCP server, and it typically serves many clients. Local vs remote MCP server compares the two.

What does a real MCP server look like?

Sume's hosted MCP server is a remote one. Its production URL is https://mcp.sume.com/mcp, which you add to Cursor, Claude Code, Codex, or another remote MCP client. It lets agents call Sume account, catalog, job, asset, generation, and Avatar tools without wrapping Sume's HTTP API; its paid generation tools include generate_image, generate_video, music_create, and tts_create.

In current code, the server declares only the tools capability, so it offers tools but no resources or prompts. Two of the read-only first calls in Sume's quickstart are mcp_health, which confirms the endpoint, auth source, and safety posture, and tools_list, which lists every tool visible to your session. You sign in with OAuth, read-only unless you turn Write on at consent, or send a Sume API key in a header, as MCP OAuth and API keys explains.

Sume's basics page says hosted MCP still works but is not part of the primary path today, and it names Formats, called over HTTP, as the surface most partners should integrate.