Sg Weather

A small MCP server that wraps NEA's 2-hour Weather Forecast API and reports rain conditions for any of the ~47 NEA forecast areas covering the whole of Singapore. Check the entire island, or pass your own list of areas to filter to just the places you care about. Built for use with Poke.com automations over streamable HTTP.

Data source: data.gov.sg dataset d_3f9e064e25005b0e42969944ccaf2e7a (NEA / Meteorological Service Singapore, 2-hour forecast).

Tools

check_rain

The main tool for a rain-alert automation. Fetches the live 2-hour forecast for all of Singapore and reports whether any checked area shows rain, showers, or thundery weather.

Optional inputs:

• areas — list of NEA area names to filter to (case-insensitive). Omit to check the whole island. Names not in the live feed are ignored and reported back under unmatched_areas, so the list is safe to tune per user.
• response_format — markdown (default) or json.

Key fields in the response:

• rain_expected (bool) — true if ANY checked area is wet
• wet_areas — list of {area, forecast} currently wet
• forecasts — areas grouped by forecast condition (keeps the whole-island payload compact instead of repeating the forecast string ~47 times)
• checked_area_count — how many areas were checked
• unmatched_areas — requested names that don't exist in the feed
• valid_period — forecast validity window
• retrieved_at — UTC timestamp

The live NEA feed is always the source of truth for area names — if NEA adds or renames areas, the server keeps working without a code change.

list_all_areas

Lists every NEA area (~47) with its current forecast. Useful for discovering exact area names to pass to check_rain, or for a Singapore-wide view. Set names_only: true for the smallest possible payload (just the names).

Deploy to Render

1. Push this folder to a GitHub repo.
2. In Render: New > Web Service > connect the repo (or use the included render.yaml via New > Blueprint).
3. Render auto-detects Python. Build: pip install -r requirements.txt. Start: python server.py. No environment variables needed — Render's PORT is picked up automatically, and render.yaml pins PYTHON_VERSION so runtime upgrades on Render's side can't silently break the service.
4. After deploy, your MCP endpoint is: https://<your-service>.onrender.com/mcp

Uptime

• Health checks: render.yaml sets healthCheckPath: /health, a dedicated liveness endpoint that always returns 200 {"status": "ok"}. Render uses it to verify deploys and restart the service if it stops responding. It deliberately does not call the NEA API, so an upstream outage never gets the service restarted or a deploy marked as failed. Don't point the health check at /mcp — the MCP endpoint returns non-2xx to plain GETs, which fails every deploy (this bit us before).
• Free-tier spin-down: Render's free plan spins the instance down after ~15 minutes without inbound traffic, so the next request (e.g. your Poke automation) waits out a cold start of up to a minute. Poke's scheduled calls will still succeed — MCP clients tolerate the delay — but if you want instant responses, either upgrade the service to a paid plan (no spin-down) or point an external uptime monitor (UptimeRobot, Better Stack, etc.) at https://<your-service>.onrender.com/health every 5–10 minutes. /health answers HEAD requests too, which is what most monitors send.

Connect in Poke.com

1. In Poke, add a custom MCP connection pointing at the /mcp URL above.
2. Create an automation on a schedule (e.g. every 30 minutes) that calls check_rain — pass the areas you care about (e.g. ["City", "Bukit Merah"]), or omit areas for island-wide coverage.
3. In the automation logic, message yourself when rain_expected is true, including wet_areas so you know which spots.

Upgrading from the earlier version? The tool was previously named check_rain_central and defaulted to a hardcoded central-area list. It is now check_rain and defaults to all of Singapore — update your automation to call check_rain and pass your preferred areas.

Local test

python -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt
PORT=8123 python server.py
# then POST an MCP initialize to http://localhost:8123/mcp

Note: the NEA API is only reachable from an environment with open outbound internet (like Render), not from restricted sandboxes.