# Weather Prediction MCP Server + Databricks Agent

Assignment 3 for the DataExpert.io Databricks AI Boot Camp.

This project exposes weather capabilities through a **FastMCP Streamable-HTTP server** backed by **Open-Meteo**, then connects a **Databricks Agent Bricks** agent to those tools.

## Architecture

```text
User
  |
  v
Databricks Agent Bricks
  |
  | MCP tool calls
  v
Weather Prediction MCP Server (Databricks App)
  |-- get_current_weather
  |-- get_forecast
  `-- get_travel_recommendation
            |
            v
      weather_adapter.py
            |
            | HTTPS
            v
 Open-Meteo Geocoding + Forecast APIs
```

The MCP layer is deliberately thin. All HTTP calls, geocoding, response parsing, WMO weather-code mapping, validation, and recommendation logic live in `weather_adapter.py`.

## Weather API and authentication

**Open-Meteo** is used for geocoding and weather forecasts. This lab requires **no API key**, so there are no API secrets to store or commit.

## MCP tools

### `get_current_weather(location)`
Returns current temperature, feels-like temperature, conditions, humidity, precipitation, cloud cover, and wind.

### `get_forecast(location, days=5)`
Returns 1-16 daily forecasts with high/low temperatures, conditions, maximum precipitation probability, precipitation total, and maximum wind.

### `get_travel_recommendation(location, date)`
A derived prediction/recommendation rather than an API passthrough:

- umbrella when precipitation probability >= 40% **or** precipitation > 0.02 in
- jacket when daily low < 55°F
- heat caution when daily high >= 90°F
- wind caution when max wind >= 25 mph

The tool returns forecast values, booleans, exact threshold logic, reasons, and a human-readable recommendation.

## Error handling

- Blank/unresolvable locations return a clean `status: error` result.
- Invalid coordinates, invalid forecast days, and invalid/out-of-range dates return clean errors.
- HTTP failures and invalid upstream JSON are translated into user-safe errors.
- Unexpected MCP failures are logged server-side but return a generic message instead of a stack trace.

## Project structure

```text
weather-prediction-mcp-agent/
├── weather_mcp_server.py
├── weather_adapter.py
├── app.yaml
├── requirements.txt
├── agent/
│   ├── system_prompt.md
│   ├── agent_config.json
│   └── demo_questions.md
└── tests/
    ├── test_weather_adapter.py
    └── test_server_contract.py
```

## Run locally

```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install pytest
pytest -q
python weather_mcp_server.py
```

FastMCP serves the MCP endpoint at:

```text
http://localhost:8000/mcp
```

## Deploy MCP server as a Databricks App

1. Push this repository to GitHub and open it as a Databricks Git folder.
2. In **Compute -> Apps**, create a **Custom** app such as `weather-prediction-mcp`.
3. Point the app source to the repository root containing `app.yaml`.
4. Deploy it.
5. Copy the Databricks App URL. The MCP endpoint is the app URL plus `/mcp`, for example:
   `https://<your-app>.aws.databricksapps.com/mcp`

`app.yaml` runs:

```yaml
command: ["python", "weather_mcp_server.py"]
```

The server binds to `DATABRICKS_APP_PORT` automatically.

## Register the MCP and build Agent Bricks agent

1. In Databricks, register the deployed app's `/mcp` URL as the MCP service/tool source.
2. Confirm Databricks discovers:
   - `get_current_weather`
   - `get_forecast`
   - `get_travel_recommendation`
3. Create an Agent Bricks agent.
4. Add the registered weather MCP server under **Tools**.
5. Paste `agent/system_prompt.md` as the system prompt.
6. Run the three prompts in `agent/demo_questions.md`.
7. Capture screenshots showing each natural-language prompt, its tool call, and the grounded answer.

## System prompt / guardrails

The supplied system prompt requires the agent to use tools for weather facts, never fabricate readings, explain tool errors rather than guess, clarify ambiguous locations, and explain the threshold behind derived recommendations.

## Required demonstration prompts

1. `What is the weather in Chicago right now?`
2. `Will it rain in Austin over the next 3 days?`
3. `Should I bring an umbrella and jacket to New York on <a date within the next 16 days>?`

S


## Notes

Open-Meteo forecasts are forecasts, not guarantees. The agent is instructed to avoid presenting predictions as certain and to use only values returned by tools.