Weather-Prediction MCP Server + Agent

Day 3 homework: a weather-forecast MCP server (pattern borrowed from databricks-lakebase-app-day-3's Alpaca paper-trading MCP server) wired to a Databricks Agent Bricks agent.

Architecture

Agent Bricks agent --(MCP tool calls)--> mcp_server/weather_mcp_server.py --(REST)--> Open-Meteo (+ NWS alerts, US only)

• mcp_server/weather_mcp_server.py - FastMCP server exposing 5 tools over streamable HTTP (the transport Databricks' MCP gateway expects).

• mcp_server/weather_broker.py - adapter module: all HTTP calls and response parsing for Open-Meteo (geocoding + forecast) and the NWS alerts API. No API key required for either.

• mcp_server/app.yaml / mcp_server/requirements.txt - Databricks App config.

Related MCP server: weather-mcp

Weather API + auth

Open-Meteo (https://open-meteo.com/) - free, no signup, no API key, ~10,000 calls/day (non-commercial use). No Databricks secret needed. NWS alerts (https://api.weather.gov/) - also free, no key, US-only.

Tools

Setup steps

1. cd mcp_server && pip install -r requirements.txt

2. python weather_mcp_server.py - serves MCP on :8000 locally.

3. Sanity-check with an MCP Inspector or curl before deploying.

4. Push this repo to your own Git remote.

5. In Databricks: create a Git folder pointing at this repo, then Compute > Apps > Create app > Custom, name it e.g. weather-forecast-mcp, source = the Git folder's mcp_server/ subfolder. Deploy, copy the app URL.

6. AI Gateway > MCPs > Add MCP - register the app URL as an external MCP server (streamable HTTP).

7. Agents > Agent Bricks > Create agent - add the registered MCP server as a tool, paste in the system prompt below, deploy, and chat with it.

Agent system prompt

You are a weather assistant. Always call a weather tool before answering any question about current conditions or forecasts - never guess or make up weather data. Use get_current_weather for "what's it like right now" questions, get_forecast for future-looking questions, and predict_umbrella_needed when the user asks whether they need an umbrella, jacket, or similar. If a location can't be resolved or a tool returns an error, tell the user plainly and ask them to clarify the location rather than guessing. Only use get_severe_weather_alerts for US locations.

Demo

1. "Will it rain in Chicago tomorrow?"

2. "Should I bring a jacket to Austin this weekend?"

3. "Compare the weather in Miami and Denver on 2026-08-15."

4. "What's the weather right now in Dakar?"

(see attached transcripts/screenshots for full tool-call traces and answers)

Test Transcripts

7. Severe-weather alerts: US vs non-US

Test 7a: US City - Miami, Florida

>>> get_severe_weather_alerts('Miami')

Response:

{
  "location": "Miami",
  "country": "United States",
  "alerts": [
    {
      "event": "Heat Advisory",
      "severity": "Moderate",
      "headline": "Heat Advisory issued August 9 at 1:15AM EDT until August 9 at 6:00PM EDT by NWS Miami FL",
      "effective": "2026-08-09T01:15:00-04:00",
      "expires": "2026-08-09T18:00:00-04:00"
    }
  ]
}

Result: ✓ Returns active alerts from NWS API (when alerts exist)

Test 7b: Non-US City - Dakar, Senegal

>>> get_severe_weather_alerts('Dakar')

Response:

{
  "location": "Dakar",
  "country": "Senegal",
  "alerts": [],
  "note": "NWS alerts only cover US locations; this location is outside the US."
}

Result: ✓ Returns empty alerts with explanatory note for non-US locations

Test 7c: US City with no active alerts - Chicago, Illinois

>>> get_severe_weather_alerts('Chicago')

Response:

{
  "location": "Chicago",
  "country": "United States",
  "alerts": []
}

Result: ✓ Returns empty list when no alerts are active (still queries NWS API)

Run the tests yourself:

python test_severe_weather_alerts.py

For your agent playground, try asking:

• "Are there any weather alerts in Miami?"

• "Check for severe weather in Dakar, Senegal"

• "What weather alerts are active in San Francisco?"

Notes

• No secrets committed; Open-Meteo and NWS need no auth. If you swap in a key-requiring API (e.g. WeatherAPI.com), store the key as a Databricks secret following the _secret() / WorkspaceClient().secrets.get_secret() pattern from the Day 3 reference repo's alpaca_broker.py - never hardcode it.