Weather-Prediction MCP Server + Agent Bricks Agent

A custom MCP server (FastMCP, streamable HTTP) exposing weather tools backed by Open-Meteo, plus a Databricks Agent Bricks agent that uses those tools to answer natural-language weather questions and make simple predictions.

Built on the Day-3 pattern (Alpaca paper-trading MCP server): thin @mcp.tool functions delegating all HTTP/parsing to an adapter module, deployed as a Databricks App and registered as an external MCP for an agent.

Architecture

Agent Bricks agent  --(MCP tool calls, streamable HTTP)-->  weather_mcp_server.py
                                                                   |
                                                                   v
                                                             weather_broker.py  --(HTTPS)-->  Open-Meteo API

• weather_mcp_server.py — FastMCP server; thin tools, served over streamable HTTP at /mcp.

• weather_broker.py — adapter: all HTTP calls + JSON parsing live here (same role as alpaca_broker.py). No MCP knowledge.

• The agent calls tools; the server calls Open-Meteo; results flow back as JSON.

Related MCP server: weather-mcp-agent

Weather API + auth

Open-Meteo (https://open-meteo.com) — no signup, no API key, ~10,000 calls/day for non-commercial use. Chosen because it needs zero credentials, so there are no Databricks secrets to manage for this project, and it bundles geocoding + current conditions + multi-day forecast in one keyless API. Works globally (not US-only). Endpoints used:

• Geocoding: https://geocoding-api.open-meteo.com/v1/search

• Forecast/current: https://api.open-meteo.com/v1/forecast

Because there is no key, there is nothing secret in this repo.

Tools

location accepts a city name ("Chicago", "Austin, TX", "London") or a raw "lat,lon" pair. when accepts "today", "tomorrow", or an ISO date.

The prediction tool does real reasoning (not a passthrough)

predict_umbrella_needed applies an explicit rule to the target day's forecast:

umbrella_needed = True if precipitation_probability >= 40% OR precipitation_sum >= 1.0 mm.

The probability catches "might rain" days; the millimetre floor catches days where the chance looks modest but meaningful rain is still expected. The tool returns the boolean and a reason string naming which threshold fired, so the agent (and the user) can see why. get_travel_recommendation layers on temperature and wind thresholds (hot ≥ 35 °C, freezing ≤ 0 °C, windy ≥ 40 km/h, wet ≥ 60% or 10 mm) to produce a headline plus a bring list.

Error handling

A bad location or an API outage returns a clean {"error": "..."} dict, never a stack trace, so the agent can react (ask the user to clarify, or report the outage) instead of failing. Example: get_current_weather("notaplace") → {"error": "Could not find a location named 'notaplace'."}.

Files

• weather_mcp_server.py — FastMCP server (5 tools; 3 required + 2 stretch)

• weather_broker.py — Open-Meteo adapter (all HTTP/parsing)

• test_weather.py — local sanity test (no MCP/Databricks needed)

• requirements.txt / app.yaml — Databricks App config

• AGENT_SYSTEM_PROMPT.md — the agent's system prompt + tool guidance

Setup

1. Test locally (optional, no credentials)

pip install -r requirements.txt
python test_weather.py           # hits Open-Meteo directly, prints results
python weather_mcp_server.py     # serves MCP at http://localhost:8000/mcp

2. Deploy the MCP server as a Databricks App

Push this folder to a Git repo, add it as a Git folder in Databricks, then: Compute → Apps → Create app → Custom, name it starting with mcp- (e.g. mcp-weather), and point it at this folder (which contains app.yaml). Databricks apps listen on port 8000 by default and expose the MCP endpoint at https://<app-url>/mcp. No secret configuration is needed (Open-Meteo is keyless).

3. Register it as an external MCP

AI Gateway → MCPs → Add MCP, paste the app's /mcp URL (streamable HTTP), name it weather-tools, and save. Databricks introspects and lists the 5 tools.

4. Build the Agent Bricks agent

Agents → Agent Bricks → Create agent (Custom LLM). Under Tools, add the weather-tools MCP server. Paste the system prompt from AGENT_SYSTEM_PROMPT.md. Evaluate, then deploy and chat.

Demo — 3 natural-language questions

These are the questions to ask the deployed agent (screenshot the tool-calling + final answers). Sample tool outputs below are from a real run.

1. "Will it rain in Seattle today — should I bring an umbrella?" → agent calls predict_umbrella_needed("Seattle", "today") → {"umbrella_needed": false, "precipitation_probability": 0, "precipitation_mm": 0.0, "reason": "No umbrella needed: only 0% chance and 0.0 mm expected, both below the thresholds (40% / 1.0 mm)..."} → agent answers: "No umbrella needed in Seattle today — 0% chance of rain, overcast but dry."

2. "What's the 3-day forecast for Austin, and is it a good idea to travel there tomorrow?" → get_forecast("Austin, TX", 3) then get_travel_recommendation("Austin, TX", "tomorrow") → recommendation "Very hot — hydrate and avoid prolonged midday sun" with bring: ["water"] (highs near 37 °C) → agent summarizes the three days and the heat advice.

3. "Which is warmer right now, Chicago, Miami, or Denver?" → compare_cities(["Chicago","Miami","Denver"]) → warmest: "Denver" (33.6 °C vs Miami 28.0 °C, Chicago 22.2 °C) → agent answers with the ranking and current conditions.

Notes / limitations

• Open-Meteo forecasts are model output; the prediction tools apply simple, transparent thresholds rather than a trained model — the point is explainable reasoning, easily tuned in weather_mcp_server.py.

• Temperatures are °C and wind km/h (Open-Meteo defaults); add temperature_unit / wind_speed_unit params in the adapter for Fahrenheit/mph if desired.

• Stretch ideas: severe-weather alerts (layer in the NWS /alerts API for US locations), historical lookup (Open-Meteo archive API), or a small dashboard app that logs agent queries.