# Weather-Prediction MCP Server + Agent

A Streamable HTTP MCP server backed by Open-Meteo, plus the prompt and registration metadata for a Databricks Agent Bricks weather agent. Open-Meteo requires no signup, API key, or committed secret.

## Submission links

- GitHub: <https://github.com/rajeshd101/databricks-mcp-demo>
- Databricks App: <https://weather-prediction-mcp-rajesh-1352785079224954.aws.databricksapps.com>
- Streamable HTTP MCP endpoint: `https://weather-prediction-mcp-rajesh-1352785079224954.aws.databricksapps.com/mcp`
- Final ZIP: [`evidence/databricks-mcp-demo-submission.zip`](evidence/databricks-mcp-demo-submission.zip)
- Grader summary: [`SUBMISSION.md`](SUBMISSION.md)
- Reproducible test guide: [`TESTING.md`](TESTING.md)

## Architecture

```text
User
  |
  v
Databricks Agent Bricks
  |  external MCP / Streamable HTTP
  v
Databricks App: weather_mcp_server.py
  |
  v
weather_adapter.py
  |----------------------|
  v                      v
Open-Meteo Geocoding     Open-Meteo Forecast
```

The server functions only validate tool inputs and shape success/error envelopes. `weather_adapter.py` owns all HTTP calls, API parsing, WMO-code translation, and recommendation logic.

## Tools

| Tool | Inputs | Result |
|---|---|---|
| `get_current_weather` | `location` | Temperature, feels-like temperature, condition, humidity, precipitation, and wind |
| `get_forecast` | `location`, `days` (1-16) | Daily high/low, feels-like values, condition, precipitation probability/amount, and wind |
| `get_travel_recommendation` | `location`, `date` (`YYYY-MM-DD`) | Forecast-backed umbrella, jacket, heat, and wind recommendations with explicit thresholds |

Locations may be a city/postal-code query or a `latitude,longitude` pair. All temperatures use Celsius, wind uses km/h, and precipitation uses millimetres.

Recommendation thresholds:

- Umbrella or waterproof layer: precipitation probability at least 40%.
- Warm jacket: daily low below 10°C.
- Light jacket: daily high below 18°C when the warm-jacket rule does not apply.
- Heat precautions: daily high at least 28°C.
- Strong-wind precautions: maximum daily wind at least 40 km/h.

## Local setup

Python 3.10 or newer is required.

```bash
python3 -m venv .venv
.venv/bin/pip install -r requirements-dev.txt
.venv/bin/pytest -q
```

Start the server:

```bash
.venv/bin/python weather_mcp_server.py
```

The Streamable HTTP endpoint is `http://localhost:8000/mcp`.

Generate three live Open-Meteo examples:

```bash
PYTHONPATH=. .venv/bin/python scripts/run_demo.py
```

Results are written to [`evidence/demo_results.md`](evidence/demo_results.md). They prove live adapter calls, not an Agent Bricks deployment.

## Databricks App deployment

The MCP server is designed to run as one Databricks App. Authenticate the Databricks CLI first:

```bash
databricks auth login --host https://<workspace-host>
databricks auth profiles
```

Create a workspace source directory, upload the project, create the app, and deploy it:

```bash
export DBX_USER='<your-workspace-email>'
export APP_NAME='weather-prediction-mcp'
export SOURCE_PATH="/Workspace/Users/${DBX_USER}/${APP_NAME}"

databricks workspace mkdirs "$SOURCE_PATH"
databricks sync . "$SOURCE_PATH" --watch=false
databricks apps create "$APP_NAME" --description 'Open-Meteo weather MCP server'
databricks apps deploy "$APP_NAME" --source-code-path "$SOURCE_PATH"
databricks apps get "$APP_NAME" -o json
```

Do not upload `.venv`; it is excluded by `.gitignore`. Databricks Apps installs `requirements.txt` and starts the `app.yaml` command. The deployed external MCP URL is:

```text
https://weather-prediction-mcp-rajesh-1352785079224954.aws.databricksapps.com/mcp
```

The app endpoint is permission-controlled by Databricks Apps. Grant the intended agent/user permission to use the app before testing the MCP connection.

## Agent Bricks configuration

The deployed server is registered as the governed Unity Catalog MCP Service:

```text
bootcamp_students.rajesh.weather_prediction_mcp
```

It is connected to Agent Bricks supervisor agent:

```text
supervisor-agent-2026-08-08-20-02-56
```

The agent uses these three tools:

- `get_current_weather`
- `get_forecast`
- `get_travel_recommendation`

Four recorded conversations—including three weather questions and an ambiguous-location guardrail test—are documented in [`evidence/agent_bricks_transcript.md`](evidence/agent_bricks_transcript.md).

## Error behavior

- Empty, unresolved, and invalid-coordinate locations return a clean `ok: false` response.
- Forecast lengths outside 1-16 and unsupported dates return validation messages.
- API timeouts, HTTP errors, malformed payloads, and unexpected internal failures do not expose stack traces.
- The agent prompt prohibits filling missing tool results with guesses.

## Verification status

- Automated adapter and MCP wrapper tests: see `tests/`.
- Live Open-Meteo adapter demonstration: see `evidence/demo_results.md`.
- Databricks App deployment: succeeded as `weather-prediction-mcp-rajesh`.
- Authenticated deployed MCP initialization: HTTP 200, MCP protocol `2025-06-18`.
- Deployed app screenshot: see `evidence/databricks-app-overview.png`.
- Active MCP Service and three-tool screenshot: see `evidence/Mcp_tools_screenshot.jpg`.
- Supervisor configuration, system prompt, MCP attachment, tool trace, and Toronto response screenshot: see `evidence/MCP_conversation_1.jpg`.
- Chicago forecast and Austin recommendation tool-trace screenshot: see `evidence/Mcp_conversation_3.jpg`.
- Austin grounded final answer and Springfield ambiguity guardrail screenshot: see `evidence/MCP_conversation_4.jpg`.
- Agent Bricks registration: completed through `bootcamp_students.rajesh.weather_prediction_mcp`.
- Agent behavior: three weather conversations and one ambiguity guardrail conversation recorded in `evidence/agent_bricks_transcript.md`.
- Evidence note: the Chicago conversation proves tool use but contains a “tomorrow” date-label mismatch and should be rerun for the cleanest correctness evidence.
- GitHub repository: published at <https://github.com/rajeshd101/databricks-mcp-demo>.

## Files

```text
agent/agent_config.yaml     External MCP tool record
agent/system_prompt.md      Agent Bricks instructions and guardrails
app.yaml                   Databricks App process configuration
evidence/demo_results.md   Three live API demonstrations
evidence/deployment.md     Deployment and protocol evidence
evidence/databricks-app-overview.png  Deployment screenshot
evidence/agent_bricks_transcript.md  Agent conversations and tool traces
evidence/Mcp_tools_screenshot.jpg  Active MCP Service and enabled tools
evidence/MCP_conversation_1.jpg  Supervisor configuration and conversation
evidence/Mcp_conversation_3.jpg  Forecast and recommendation traces
evidence/MCP_conversation_4.jpg  Recommendation answer and guardrail
scripts/run_demo.py        Reproducible live demonstration
tests/                     Adapter and MCP error-boundary tests
weather_adapter.py         Open-Meteo HTTP/parsing/recommendation layer
weather_mcp_server.py      Thin FastMCP tool layer
SUBMISSION.md              Grader-facing submission summary
TESTING.md                 Local and deployed test procedure
```

## Limitations

- Current conditions are modeled Open-Meteo data, not direct observations from a local weather station.
- Geocoding selects the first Open-Meteo match; users should add province/state and country for ambiguous names.
- This version does not provide official severe-weather alerts. Users should consult official local alert services for safety-critical decisions.
- Forecasts are limited to Open-Meteo's next 16 days.