My first MCP, or how I talked to SQL Server in plain Spanish

By late July 2026 everyone was talking about MCP (Model Context Protocol) and I had no idea what it was. The fastest way to understand something is to build it, so on July 30 the MCP Relational AI Gateway was born.

MCP in one sentence

It’s a standard that lets an AI model use external tools. You write a server once and any compatible client (Claude, Cursor, whatever) can use it. Think of it as USB for agents.

What the gateway does

You connect to SQL Server or PostgreSQL, pick a database and ask things like “what were the 10 most rented movies?”. The model writes the SQL, runs it through the MCP tools and, when it makes sense, draws a zoomable chart. You can switch between Gemini and Claude without touching code.

That same day I piled on features: a database picker on the login screen, an animated collapsible sidebar (of course), and I removed a console and some logs that were just in the way. Everything runs in Docker.

What I learned

• The model makes fewer mistakes when you give it the schema instead of letting it guess.
• Separating “what the model decides” from “what gets executed” is key. The model proposes, the server validates and runs.
• MCP isn’t magic: under the hood it’s JSON-RPC calls with a well-defined schema. But that well-defined schema is exactly what makes it work with any client.

This project is what made me think “ok, I want to build an MCP server people actually use”. That came a couple of months later with mcp-banxico.