Build a Custom MCP Server

MCP (Model Context Protocol) servers let you extend Claude Code with custom tools. If your team has an internal API, a proprietary database, or a custom workflow, wrap it in an MCP server.

Minimal TypeScript Server

import { McpServer } from "@anthropic-ai/mcp";

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

server.tool("get_user", "Fetch user by ID from our internal API", {
  userId: { type: "string", description: "The user ID" },
}, async ({ userId }) => {
  const res = await fetch(`https://internal.api/users/${userId}`);
  const user = await res.json();
  return { content: [{ type: "text", text: JSON.stringify(user, null, 2) }] };
});

server.start();

Register It

Add to .claude/settings.json:

{
  "mcpServers": {
    "internal-api": {
      "command": "npx",
      "args": ["tsx", ".claude/mcp/internal-api.ts"]
    }
  }
}

What to Build MCP Servers For

• Internal APIs: give Claude access to your company’s services
• Databases: safe read-only queries against staging/dev databases
• Monitoring: pull metrics from Datadog, Grafana, or CloudWatch
• Ticketing: read/update Jira, Linear, or GitHub issues
• Documentation: search internal wikis or Confluence

Keep It Read-Only

Start with read-only tools. Don’t give Claude write access to production databases or ticketing systems until you’ve validated the patterns with read-only access.

Tip

Use descriptive tool names and parameter descriptions. Claude picks the right tool based on these descriptions, so “get_user_by_email” works better than “query1”.