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

• Grader summary: SUBMISSION.md

• Reproducible test guide: TESTING.md

Related MCP server: weather-mcp

Architecture

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

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.

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

Start the server:

.venv/bin/python weather_mcp_server.py

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

Generate three live Open-Meteo examples:

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

Results are written to 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:

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

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

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:

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:

bootcamp_students.rajesh.weather_prediction_mcp

It is connected to Agent Bricks supervisor agent:

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.

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

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.