Skip to content

Building Your Own MCP Server

  • Intermediate · Module 05
  • Lesson 4 of 5 · then sources and quiz

Step 1 of 4Knowledge

In this section, you'll build your own MCP server step by step. By the end, you'll have a working server that you can integrate into Claude Desktop or Cursor.

We'll build a Notes MCP Server that can:

  • Create notes (Tool)
  • Read notes (Resource)
  • Provide a summarization template (Prompt)

Step 2 of 4Understand

Anatomy of an MCP Server

An MCP server consists of four parts:

  1. Server instance -- Name and version
  2. Tool definitions -- Functions with Zod validation
  3. Resource definitions -- Data endpoints with URI
  4. Transport binding -- stdio for local servers

With stdio transport, an important detail to note: stdout is used for JSON-RPC communication. Log messages must therefore go through stderr (console.error).

Step 3 of 4Apply

Step 1: Set Up the Project

Create a new directory and initialize the project:

mkdir mcp-notes-server
cd mcp-notes-server
npm init -y
npm install @modelcontextprotocol/sdk zod
npm install -D typescript @types/node

Create a tsconfig.json:

{
  "compilerOptions": {
    "target": "ES2022",
    "module": "Node16",
    "moduleResolution": "Node16",
    "outDir": "./dist",
    "rootDir": "./src",
    "strict": true,
    "esModuleInterop": true
  },
  "include": ["src/**/*"]
}

Step 2: Server Skeleton

Create src/index.ts -- the heart of your server:

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

// In-memory storage for notes
const notes: Map<string, { title: string; content: string; createdAt: string }> = new Map();

// Create server
const server = new McpServer({
  name: "notes-server",
  version: "1.0.0",
});

This is the skeleton: you import the SDK, create a data store, and initialize the server with a name and version.

Step 3: Define a Tool -- Create Note

Now let's add the first tool:

// Tool: Create note
server.tool(
  "create_note",
  "Creates a new note with a title and content",
  {
    title: z.string().describe("Title of the note"),
    content: z.string().describe("Content of the note"),
  },
  async ({ title, content }) => {
    const id = `note_${Date.now()}`;
    notes.set(id, {
      title,
      content,
      createdAt: new Date().toISOString(),
    });

    return {
      content: [
        {
          type: "text",
          text: `Note "${title}" created (ID: ${id})`,
        },
      ],
    };
  }
);

Note the structure: server.tool() takes four parameters:

  1. Name -- Unique identifier
  2. Description -- What the tool does (for the LLM)
  3. Schema -- Parameters with Zod validation
  4. Handler -- The actual logic

Step 4: Define a Resource -- Read Notes

// Resource: List all notes
server.resource(
  "notes-list",
  "notes://list",
  async (uri) => {
    const allNotes = Array.from(notes.entries()).map(
      ([id, note]) => `- [${id}] ${note.title} (${note.createdAt})`
    );

    return {
      contents: [
        {
          uri: uri.href,
          mimeType: "text/plain",
          text: allNotes.length > 0
            ? allNotes.join("\n")
            : "No notes available.",
        },
      ],
    };
  }
);
Tip
Resources vs. Tools

The Resource only reads data -- it doesn't change anything. The LLM can use this Resource to enrich the context of its responses with current notes.

Step 5: Define a Prompt -- Summarization Template

// Prompt: Summarize notes
server.prompt(
  "summarize_notes",
  "Creates a summary of all notes",
  async () => ({
    messages: [
      {
        role: "user",
        content: {
          type: "text",
          text: `Here are my current notes:\n\n${
            Array.from(notes.entries())
              .map(([id, n]) => `## ${n.title}\n${n.content}`)
              .join("\n\n")
          }\n\nPlease create a structured summary with the key points.`,
        },
      },
    ],
  })
);

Step 6: Start the Server

Finally, connect the server to the stdio transport:

// Start server with stdio transport
async function main() {
  const transport = new StdioServerTransport();
  await server.connect(transport);
  console.error("Notes MCP server running...");
}

main().catch(console.error);
Note
console.error Instead of console.log

With stdio transport, stdout is used for JSON-RPC communication. That's why log messages must be sent via stderr (console.error) so they don't interfere with the protocol.

Step 7: Build and Test

# Compile TypeScript
npx tsc

# Test the server with the MCP Inspector
npx @modelcontextprotocol/inspector dist/index.js

The MCP Inspector is an interactive debugging tool that lets you test your server without needing to integrate it into an AI application.

Step 8: Integrate with Claude Desktop

Add your server to the Claude Desktop configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "notes": {
      "command": "node",
      "args": ["/absolute/path/to/dist/index.js"]
    }
  }
}

After restarting Claude Desktop, your Notes server will appear in the tool list. You can now say:

"Create a note titled 'Meeting Outcome' with the content 'Budget was approved'."

And Claude will call your create_note tool.

The Complete Code

Here's the full src/index.ts again:

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

const notes: Map<string, { title: string; content: string; createdAt: string }> = new Map();

const server = new McpServer({
  name: "notes-server",
  version: "1.0.0",
});

server.tool(
  "create_note",
  "Creates a new note with a title and content",
  {
    title: z.string().describe("Title of the note"),
    content: z.string().describe("Content of the note"),
  },
  async ({ title, content }) => {
    const id = `note_${Date.now()}`;
    notes.set(id, {
      title,
      content,
      createdAt: new Date().toISOString(),
    });
    return {
      content: [{ type: "text", text: `Note "${title}" created (ID: ${id})` }],
    };
  }
);

server.resource(
  "notes-list",
  "notes://list",
  async (uri) => {
    const allNotes = Array.from(notes.entries()).map(
      ([id, note]) => `- [${id}] ${note.title} (${note.createdAt})`
    );
    return {
      contents: [{
        uri: uri.href,
        mimeType: "text/plain",
        text: allNotes.length > 0 ? allNotes.join("\n") : "No notes available.",
      }],
    };
  }
);

server.prompt(
  "summarize_notes",
  "Creates a summary of all notes",
  async () => ({
    messages: [{
      role: "user",
      content: {
        type: "text",
        text: `Here are my current notes:\n\n${
          Array.from(notes.entries())
            .map(([id, n]) => `## ${n.title}\n${n.content}`)
            .join("\n\n")
        }\n\nPlease create a structured summary.`,
      },
    }],
  })
);

async function main() {
  const transport = new StdioServerTransport();
  await server.connect(transport);
  console.error("Notes MCP server running...");
}

main().catch(console.error);
Question · Self-check
Why does an MCP server with stdio transport use console.error instead of console.log for log messages?

Step 4 of 4Reflect

Building your own MCP server gives you full control over integrating your data and services. The key point: stdout is reserved for JSON-RPC -- logs must go through stderr. In the next section, we will look at MCP in the enterprise context.