Why
An AI assistant that answers from memory can get dates, distances and risks wrong. An MCP server gives it tools to fetch fresh numbers from the source instead of guessing, and to show where the numbers come from.
What I built
- Research first. 12 documents on APIs, market, the MCP spec, architecture and roadmap, and 42 APIs researched.
- An MCP server in TypeScript. 11 tools with input validation, and source, license and timestamp on every answer.
- Safety. The server refuses requests to command satellites.
- A website. Next.js with 12 pages and 13 API routes, where research briefs export as .md or .bib.
- Agent skills. A skill for Claude and one for OpenClaw with four agents.
- Hosting. On my own server with systemd, nginx and Let's Encrypt, with a public health check.
The result
The server is live with 11 tools, and the code is open on GitHub under the Apache 2.0 license. The answers in the illustration above are real answers from the server on 31 October 2026.
The same technique can connect an AI assistant to your CRM, your bookings or your prices, so it answers with your numbers instead of guessing.
Tools and tech stack
- TypeScript
- Node.js
- MCP SDK
- zod
- Next.js
- Tailwind CSS
- nginx
- Let's Encrypt
Want this for your business? See Generative engine optimization (GEO) or AI automation services.







