Let’s build. In this lesson you’ll create a minimal MCP server that exposes one tool, run it over stdio, and connect it to a host. We’ll use the official TypeScript SDK, but the concepts transfer to the Python SDK too.

🎯 Learning Objectives

• Scaffold an MCP server with the official SDK
• Define a tool with an input schema and handler
• Run the server over stdio
• Connect and test it from a host

Step 1 — Set Up the Project

Install the official MCP SDK and a schema validator:

mkdir devops-mcp && cd devops-mcp
npm init -y
npm install @modelcontextprotocol/sdk zod

zod lets us declare and validate the tool’s input schema — critical, because the model will send arguments and we must not trust them blindly.

Step 2 — Create the Server & a Tool

Here’s a complete, minimal server exposing one tool, get_service_status, that reports whether a named service is healthy:

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { z } from "zod";

// 1. Create the server
const server = new McpServer({
name: "devops-tools",
version: "1.0.0",
});

// 2. Register a tool: name, description, input schema, handler
server.tool(
"get_service_status",
"Check whether a named service is currently healthy",
{ service: z.string().describe("The service name, e.g. 'payments'") },
async ({ service }) => {
  // (in real life: query your monitoring / health endpoint)
  const healthy = await checkHealth(service);
  return {
    content: [
      {
        type: "text",
        text: healthy
          ? service + " is healthy ✅"
          : service + " is DOWN ❌",
      },
    ],
  };
}
);

// 3. Connect over stdio
const transport = new StdioServerTransport();
await server.connect(transport);

Three things to notice:

• The description tells the model when to use the tool.
• The zod schema (service: z.string()) validates input automatically.
• The handler returns content the model reads back.

Step 3 — Run and Register It

An stdio server isn’t run directly by you — the host launches it as a subprocess. You tell the host how to start it via config. For a typical desktop host:

{
"mcpServers": {
  "devops-tools": {
    "command": "node",
    "args": ["/path/to/devops-mcp/server.js"]
  }
}
}

On start, the host runs that command, performs the initialize handshake, calls tools/list, and your get_service_status tool appears — ready for the model to call.

Step 4 — Test It

Before wiring it into a host, test with the MCP Inspector, an official dev tool that acts as a client:

npx @modelcontextprotocol/inspector node server.js

▶ tools/list
- get_service_status: Check whether a named service is currently healthy

▶ tools/call get_service_status { "service": "payments" }
← { "content": [{ "type": "text", "text": "payments is healthy ✅" }] }

Seeing the tool listed and returning a result means your server is correct — now any MCP host can use it.

🧪 Hands-on Lab

const ALLOWED = ["payments", "orders", "inventory"];

server.tool(
"restart_service",
"Restart one of the known services. Use only when explicitly asked.",
{ service: z.enum(["payments", "orders", "inventory"]) },
async ({ service }) => {
  if (!ALLOWED.includes(service)) {
    return { content: [{ type: "text", text: "Not allowed." }], isError: true };
  }
  await ops.restart(service);
  return { content: [{ type: "text", text: "Restarted " + service }] };
}
);

🧠 Knowledge Check

💼 Interview Preparation

Summary

You built a working MCP server, defined and validated a tool, ran it over stdio, and tested it with the Inspector. You now know how to give an AI safe, structured access to a system. Finally, let’s see how MCP is used in real DevOps workflows.