Eighty & Sunny ☀️

A colorful, weather-reactive MCP Apps server. It renders rich Open-Meteo weather widgets directly inside an MCP host (the same SEP-1865 "MCP App" pattern as the NewsAPI / Taboola demos) — and ships a plain JSON tool for agents that just want the data.

The whole thing runs with no API key: it's backed by Open-Meteo, which is free for non-commercial use and needs no signup.

Background gradients, glow orbs and accent colors react to the actual weather code and day/night at each location — clear skies glow gold and blue, storms go deep purple, snow turns icy, night dims to indigo.

Tools

All four widget tools accept:

• city — city to show first. If it isn't one of the pre-fetched cities it's geocoded and added as the active tab. Blank → first configured city.

• units — "imperial" (°F, mph, in) or "metric" (°C, km/h, mm). Blank → server default. Every widget also has a live °F / °C toggle.

Restricted triggers

Like the NewsAPI/Taboola demos, the show_* tools are documented as restricted-trigger: only invoke them when the user wants to see a weather widget (e.g. "show me the weather in Paris", "open Eighty & Sunny", "hourly forecast for Tokyo", "full weather dashboard"). For plain data questions, prefer get_weather.

Related MCP server: Simple Weather MCP

How it works

host ──tool call──▶ server.py
                      │  geocode each city  (geocoding-api.open-meteo.com)
                      │  fetch forecast     (api.open-meteo.com)  ── current + hourly + 7-day
                      ▼
                 normalize → _last_payload (canonical METRIC values)
                      │
host ──resource read─▶ widget HTML  ◀── payload inlined into <script id="weather-payload">

• One place per call (rate-limit friendly). By default a tool call fetches only the active city (≤2 Open-Meteo requests), not all six. The other chips fill in from cache as you visit cities, so switching stays instant without hammering the API. Set WEATHER_PREFETCH=all to eagerly fetch every configured city up front.

• TTL cache + 429 backoff. Each city's forecast is cached for WEATHER_CACHE_TTL seconds (default 600); repeat requests make no API call. On a 429/5xx the server retries with backoff (honoring Retry-After) and, if that still fails, serves the last-good cached data instead of an error.

• Metric-canonical. All numbers are stored in °C / km/h / mm so the in-widget °F/°C toggle converts client-side — no refetch needed.

• Open-Meteo stays whitelisted in the widget CSP (connectDomains), so the app can also live-fetch if needed.

Run locally

uv sync
uv run python server/server.py            # http://localhost:8778/mcp
uv run python server/server.py --port 9000

Point your MCP host at http://localhost:8778/mcp (streamable HTTP).

Quick smoke test (no host needed)

uv run python -c "import asyncio, server.server as s; \
print(asyncio.run(s.get_weather('Paris'))['cities']['paris']['current'])"

Configuration

Everything is optional — see .env.example. Copy it to .env to override:

Deploy

Render (one click via render.yaml)

Push to GitHub, then in Render: New → Blueprint, point it at the repo. The included render.yaml builds with uv and starts the server on $PORT. No env vars are required (Open-Meteo needs no key); uncomment the ones in render.yaml to customize cities/units.

Docker / Azure Container Apps

docker build -t es-weather-mcp .
docker run -p 8080:8080 es-weather-mcp        # http://localhost:8080/mcp

azure.yaml is included for azd up (Azure Container Apps).

Credits

• Weather data: Open-Meteo (free, non-commercial).

• MCP Apps via @modelcontextprotocol/ext-apps.