Imported from magnus919/agent-skills (tempest/SKILL.md). Install upstream with npx skills add magnus919/agent-skills --skill tempest. Copyright stays with the author (MIT).

tempest — Hyper-local weather from your Tempest station

Drive a WeatherFlow Tempest station from the terminal. Two transports, both first-class: the documented REST API (swd.weatherflow.com/swd/rest, personal-use token) for conditions, forecast, and history — officially the primary data source — and the hub's unauthenticated UDP broadcast on port 50222 for real-time, lowest-latency readings on your LAN. The bundled CLI decodes the positional observation arrays and every UDP message family, keeps --json output metric-native, and converts units only for human display.

Setup

1. Create a personal access token: sign in to the Tempest web app (tempestwx.com), then Settings → Data Authorizations → Create Token. (This is the documented non-graphical auth method; OAuth exists for web apps but is not what a CLI uses.)
2. Export it:

export TEMPEST_TOKEN="<YOUR_TOKEN>"

The token travels to the API as a query parameter (?token=...) per the official docs — the CLI handles this. If the env var is not set, the CLI falls back to reading TEMPEST_TOKEN= from ~/.tempest.env (handy for agent subprocesses that skip shell profiles). --help and --dry-run never need a token. UDP listening never needs one either — the hub broadcast is unauthenticated and LAN-only.

Essential Commands

stations — discover your stations and devices

tempest stations                       # names, station ids, device types, serials
tempest stations --json | jq '.stations[] | {station_id, name,
  devices: [.devices[] | {device_id, device_type, serial_number}]}'

Every station response nests a devices array: device_type is ST (the Tempest all-in-one), AR/AIR, SK/SKY, or HB (the hub — it has no observations; always filter it out before querying observations). Run this first when you don't know your ids.

current — latest conditions

tempest current                                       # human-readable, converted
tempest current --json                                # metric-native, jq-ready
tempest current --station-id 12799 --device-id 60526  # pin exact hardware

With one station it auto-selects and picks the best sensor (ST, then SKY/SK, then AIR/AR, skipping HB). Output .observation carries the decoded positional array as named fields with _unit companions.

forecast — current conditions + daily + hourly

tempest forecast                       # current + 5-day daily + next 12 hours
tempest forecast --days 7 --json
tempest forecast --station-id 12799 --days 3

The better_forecast response nests daily/hourly under a forecast wrapper key, and it is unit-selectable (units_temp=c|f and friends, default metric) — the CLI reads the response's units before converting anything.

obs — historical observations

tempest obs --device-id 60526 --days 1       # last UTC day (day_offset)
tempest obs --device-id 60526 --days 7
tempest obs --device-id 60526 --json

--days N maps to the API's day_offset (whole UTC days). The underlying endpoint also accepts time_start/time_end epoch ranges (one-minute resolution guaranteed up to 5 days) — use raw calls for those; see references/rest-api-and-auth.md.

UDP broadcasts from your hub (port 50222, listen-only)

tempest udp listen                          # live stream until Ctrl-C
tempest udp listen --timeout 30             # auto-stop after 30s
tempest udp listen --timeout 60 --json      # one JSON object per datagram
tempest udp listen --show-all               # include hub_status/device_status

Requires being on the same LAN as the hub (routed connectivity is not enough — broadcasts don't cross routers). No token involved. The listener decodes every message family, dispatching on type before touching array positions:

Multi-step pipeline recipes

Discover, then observe

# Stage 1 -> stage 2: stations --json emits integer ids that current consumes
tempest stations --json | jq -r '.stations[].devices[]
  | select(.device_type == "ST") | .device_id' | head -1
tempest current --device-id <DEVICE_ID> --json

Rain watch: yesterday's total, then live rain events

tempest obs --device-id 60526 --days 1 --json \
  | jq '{samples: (.observations | length),
         day_rain_mm: .observations[-1].local_day_rain_accumulation}'
tempest udp listen --timeout 600 --json | jq 'select(.type == "evt_precip")'

obs --json ends with decoded observations carrying local_day_rain_accumulation (mm, number); evt_precip datagrams decode to {type, serial_number, timestamp} — both stages emit typed fields the next stage can consume.

Unit-aware forecast slice

tempest forecast --days 7 --json \
  | jq '{units_temp: .forecast.units.units_temp,
         highs_f: [.forecast.forecast.daily[] | .air_temp_high * 9 / 5 + 32],
         rain_hours: [.forecast.forecast.hourly[]
                      | select(.precip_probability > 30) | .local_hour]}'

The jq math here is safe only because it checks units_temp first — see gotcha 2.

JSON output and jq processing

--json output is metric-native — the raw wire units (m/s wind, mm rain, °C temperature, MB pressure) with _unit companion fields naming each. Convert at the consumption edge:

tempest current --json | jq '{temp_c: .observation.air_temperature,
  temp_f: (.observation.air_temperature * 9 / 5 + 32),
  wind_mph: (.observation.wind_avg * 2.237),
  rain_in: (.observation.rain_accumulation / 25.4)}'

Global flags work in any position: tempest --json current --device-id 60526 and tempest current --device-id 60526 --json are identical. --quiet silences the progress logs (data on stdout, logs on stderr). --dry-run prints a plan object and exits 0 without touching the network.

Known Gotchas

1. Observations are positional arrays, not objects. Raw obs rows have no field names; meaning comes from the index (obs_st: 0 epoch, 2 wind avg m/s, 4 wind direction, 6 pressure MB, 7 temperature °C, 12 rain mm, 16 battery V, 17 report interval). Reading index 6 as temperature gives you a plausible-looking wrong number — decode with the CLI or the layout tables in references/observation-layouts-and-units.md.
2. /better_forecast is unit-selectable, not Celsius-locked. It defaults to metric but honors units_temp=f, units_wind=mph, units_pressure=inhg, units_precip=in. It reports what it used in response.units. Converting an already-Fahrenheit response doubles it (25.4 °C → 77.7 °F → 172 "°F"). Always read units before converting; the CLI does this for you.
3. UDP message families differ structurally — dispatch on type first. obs families nest rows under obs; rapid_wind carries one array under ob; evt_precip/evt_strike carry one array under evt; hub_status/device_status carry named fields with no payload array. Iterating rapid_wind's ob element-wise is the classic TypeError; the bundled decode_message() shows the correct dispatch.
4. UDP obs_st rows stop at index 17; REST rows run to 21. The four Nearcast/analysis fields (18–21) exist only in REST responses. Decoders must tolerate both lengths — the CLI emits None for missing tails.
5. Pressure is MB (millibars), numerically hPa — not kPa. It is also station pressure (raw sensor). The Tempest app's "relative pressure" adds an elevation adjustment; don't compare raw station pressure against the app and conclude the sensor drifted.
6. Forecast timestamps are epoch integers, never ISO strings. day_start_local, sunrise, sunset, hourly time are epoch seconds; hourly objects carry local_hour (0–23) and local_day (day of month). There is no local_time or time_string field — code expecting one silently falls back to its default branch.
7. The forecast nests under a forecast wrapper key. data["daily"] is always empty; read data["forecast"]["daily"] and data["forecast"]["hourly"] (the CLI's --json preserves the full response, wrapper and all).
8. Hubs (HB) have no observations. They only relay. Auto-selection skips them; if you call the API directly, filter device_type == "HB" out before hitting /observations/device/{id} (documented 404 otherwise).
9. UDP is LAN-only and unauthenticated. Broadcasts don't cross routers and can't be token-gated — anyone on the network can read your station. WeatherFlow officially positions REST/WebSocket as primary and UDP as the off-grid/backup interface.
10. obs_sky UDP day-rain is always null. Local-day rain accumulation (index 11) is null in UDP SKY broadcasts; REST supplies the real value. Don't build day-rain totals from UDP SKY rows.

When to use

• The user owns or manages a WeatherFlow Tempest / Air / Sky station and asks about its readings, forecast, or history.
• Parsing or integrating with the hub's local UDP broadcast (port 50222).
• Rain/wind/lightning monitoring scripts, dashboards, or home-automation hooks fed from the station.

When not to use

• Generic city forecasts or users without a station — every endpoint requires the user's own Tempest station and a personal-use token; use a public weather service instead.
• Shakespeare's play The Tempest, or any literary/meteorological-theory question — this is a station-data CLI, not an encyclopedia.
• Other vendors' hardware (Netatmo, Ecowitt, Davis, Ambient) — different APIs entirely; no endpoint here will accept their devices.
• Commercial/network-wide data products — those need WeatherFlow's TempestONE agreements, not a personal token (see the remote developer policy).

Reference Files

Available Scripts

• scripts/tempest — the CLI: stations, current, obs, forecast, udp listen; global --json, --dry-run, --quiet, --verbose accepted in any position; offline dry-run plans for every command.
• scripts/test_tempest.py — offline suite: canned UDP datagram bytes fed to the decoder (no sockets), mocked REST transport, both pytest and unittest runners.

Prerequisites

• Python 3.8+ with requests (the only dependency).
• TEMPEST_TOKEN for REST commands (free, personal use; created in the Tempest web app). UDP listening needs no token, only line-of-sight to the hub's LAN.