🌩️ Weather Alert & Forecast Explorer — Athena MCP Agent

An MCP server that exposes live open weather data and renders it as an embedded interactive widget inside Athena.

Built for the Athena AI Interview Challenge — subject: Weather alert and forecast explorer.

Live endpoints

No authentication — the endpoint is public so any MCP host can connect directly.

Real open/public data sources

No API keys, no auth, no scraping — both are official public feeds.

Alert coverage is stated honestly in the payload (alertCoverage) rather than silently returning an empty list outside the US.

MCP tools

explore_weather

Opens the interactive explorer for any location on Earth.

{ "location": "Tampa, Florida", "days": 7 }

Returns a text summary (for the model), structuredContent (for programmatic use), and the interactive widget.

list_active_alerts

Lists currently active NWS alerts, filterable by state code and severity — used to find where severe weather is happening, then drill in with explore_weather.

{ "area": "TX", "severity": "Severe" }

The widget — interactions

The widget ships three meaningful interactions (two were required):

1. Day selector — click any of the 7 forecast day cards to pivot the entire hourly view to that day. Keyboard accessible (Enter/Space).

2. Metric toggle — switch the hourly chart between Temperature, Rain %, and Wind. The chart re-renders as a line or bar series with an axis scaled to the selected metric.

3. Expandable alerts — click any active NWS alert to expand the full official description, protective-action instructions, onset/expiry window, urgency and certainty.

Hovering the chart also scrubs a per-hour readout (temp / precip / wind).

Selecting a day or expanding an alert additionally emits a follow-up prompt to the host so the conversation can continue from what the user clicked.

Design notes

Host-convention hedging. MCP hosts disagree on how a tool returns UI. Each explore_weather result therefore carries the widget in every known convention at once — an mcp-ui embedded resource in the content array, plus _meta.ui.resourceUri (Anthropic Apps) and _meta["openai/outputTemplate"] (OpenAI Apps SDK). Unknown _meta keys are ignored by spec, so this costs nothing and removes a single point of failure.

Self-contained widget. The full dataset is inlined as JSON at render time and every interaction runs client-side. The widget makes zero network calls and needs zero host cooperation to be fully interactive — important because iframe CSP and host-messaging support vary widely. Host messaging is layered on as enhancement, never as a dependency.

Graceful degradation. Hosts with no widget support still get a complete text summary plus structuredContent, so the model can answer regardless.

Safety. NWS alert text is third-party input and is rendered exclusively via textContent / DOM construction — never innerHTML.

Run locally

npm run dev      # http://localhost:3000/mcp

Smoke test:

curl -s http://localhost:3000/mcp \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"explore_weather","arguments":{"location":"Miami, Florida"}}}'

Open the widget standalone in a browser:

http://localhost:3000/widget?location=Tampa,Florida

Stack

Zero runtime dependencies. Hand-rolled JSON-RPC 2.0 over Streamable HTTP (POST /mcp, with SSE framing when the client's Accept header asks for it), deployed as a Node serverless function on Vercel.