The distinguishing move is that the calibration is not hidden: `get_forecast` returns the bias-adjusted forecast and `list_cities` returns the calibration parameters, so an edge number can be traced back to the forecast and the station observation that produced it. Coverage is deliberately narrow — five cities — and the market and weather feeds are all public APIs.
An MCP server for Kalshi weather markets. It calibrates National Weather Service daily high-temperature forecasts per city, reads current Kalshi prices, and returns per-bucket probability, edge and net expected value — so the model gets a signal rather than a raw forecast.
- Calibrated signals for one city's Kalshi weather markets — `get_weather_signals`
- The same scan across every supported city at once — `get_all_signals`
- The bias-adjusted forecast behind a signal, so the input is inspectable — `get_forecast`
- The latest METAR observation from the station a market settles against — `get_station_observation`
- The supported cities and their calibration parameters — `list_cities`; the cities are `nyc`, `chicago`, `denver`, `miami` and `la`
- Three transports from one binary: `--transport stdio`, `--transport sse --port 8050`, `--transport streamable-http --port 8050`
The README's install line names `weather-edge-mcp` 0.2.1, which PyPI does not have. The documented Claude Desktop entry uses the module form, `python -m weather_edge_mcp`. Data comes from the National Weather Service forecast API, the Aviation Weather METAR API and the Kalshi public market API. An optional FastAPI surface can be started separately with uvicorn on port 8080, and the README notes stdio mode stays side-effect free.