Weather MCP

MCP server that predicts rain for the next 7 days in any city and explains the prediction using SHAP values. Mirrors Part 3 of weather_shap.ipynb: XGBoost + SHAP TreeExplainer + live Open-Meteo data.

Tools

Data source: Open-Meteo (no API key required).

Install

cd weather-mcp
python -m venv .venv
source .venv/bin/activate
pip install -e .

This installs the weather-mcp console script, which is what the MCP clients below will launch.

Running locally (smoke test)

weather-mcp

The server speaks MCP over stdio — it will sit waiting for JSON-RPC. Use Ctrl-C to stop. Real testing happens through one of the clients below.

Connect to Claude Desktop

Edit ~/.config/Claude/claude_desktop_config.json (Linux) or ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

{
  "mcpServers": {
    "weather": {
      "command": "/absolute/path/to/weather-mcp/.venv/bin/weather-mcp"
    }
  }
}

Restart Claude Desktop. You should see a 🔌 icon listing predict_rain_7day, explain_rain_prediction, and city_model_info.

Connect to Cursor

Settings → MCP → Add new MCP server:

• Name: weather
• Command: /absolute/path/to/weather-mcp/.venv/bin/weather-mcp

Connect to VS Code (Continue extension or any MCP-aware client)

Add to your client's MCP config:

{
  "mcpServers": {
    "weather": {
      "command": "/absolute/path/to/weather-mcp/.venv/bin/weather-mcp",
      "args": []
    }
  }
}

Example conversation

You: Will it rain in Tokyo over the next 7 days?

(client calls predict_rain_7day(city="Tokyo"))

Claude: Tomorrow shows a 78% chance of rain, day 3 drops to 12%, …

You: Why is day 1 so high?

(client calls explain_rain_prediction(city="Tokyo", day_index=0))

Claude: The biggest driver is forecasted precipitation = 8.4 mm (+1.92 log-odds), reinforced by a higher-than-usual min temperature …

How it works

1. Geocode the city via Open-Meteo's geocoding endpoint.
2. Train an XGBoost binary classifier on ~2 years of daily archive data (precipitation, temp_max, temp_min, windspeed → did it rain that day, derived from WMO weather codes). The trained model is cached per location.
3. Predict by fetching the live forecast and running predict_proba.
4. Explain with shap.TreeExplainer — exact Shapley values for the trained model, surfaced as ranked factors per day.

Notes

• Models are cached in-process; restarting the server retrains on next call.
• A given location key rounds lat/lon to two decimals, so "Tokyo" and a query near Tokyo Tower share the same model.
• explain_rain_prediction will auto-run a 7-day prediction if you ask it before predict_rain_7day for that city.