← All work

Teaching agents the forecast.

Overview

A Model Context Protocol server exposing five weather tools to AI agents — forecast, current, historical, location search and astronomy — with an evalite test suite scoring the answers.

MCP is how AI assistants get real capabilities, so this is a ground-up server implementation: five typed tools over the WeatherAPI, from forecasts to sunrise and moon phase. The interesting part is validation — an evalite suite runs real model conversations against the server and scores whether the tools get picked and the answers hold up, with an interactive UI for inspecting each eval.

View on GitHub ↗

Stack

  • TypeScript
  • Model Context Protocol
  • Node
  • Evalite

Details

  • Year2026
  • KindMCP server · TypeScript