Anatomy of an MCP Server: What's Inside and How It's Built

An MCP server is typically a Node.js script (TypeScript compiled to JavaScript) or a Python script. When your AI client starts it, the server process launches and communicates over stdio. The first thing it does is declare its capabilities in response to the client's initialization request.

Each tool is defined with a name, description, and inputSchema (JSON Schema). The description is what the AI reads to decide when to use the tool — so well-written descriptions lead to better AI behavior. When the AI calls the tool, the server receives the arguments, runs the logic, and returns a result object.

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js"; import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js"; import { z } from "zod"; const server = new McpServer({ name: "my-server", version: "1.0.0" }); server.tool("greet", { name: z.string() }, async ({ name }) => ({ content: [{ type: "text", text: `Hello, ${name}!` }], })); const transport = new StdioServerTransport(); await server.connect(transport);

Resources are optional. They're declared with a URI template and a handler that returns content when the AI requests that URI. Resources are read-only — they don't execute actions, they return data.

If a tool throws, the error is returned as a structured error response. Good servers validate inputs (using Zod in TypeScript or Pydantic in Python) and return clear error messages so the AI can handle failures gracefully.

Most servers are published to npm (for Node.js) or PyPI (for Python). Installation is just `npm install -g package-name` or `pip install package-name`. The install config in your AI client's JSON file points to the installed binary.