| name | weather-bandit |
| description | Fetch deterministic, cross-validated weather forecasts from Open-Meteo and print a daily outlook as Markdown + JSON. Use when: (1) You need a weather forecast for a city, (2) You want to cross-check multiple weather models, (3) You want ensemble-derived probabilities (e.g. P(max ≥ 30°C)), (4) You want a conditional-climatology outlook for a target beyond the forecast horizon. Triggers on: "weather forecast", "weather for <city>", "how hot will it be", "rain probability", "weather bandit", "ensemble probability", "cross-validate models", "analog forecast". |
Prerequisites
Weather Bandit is a CLI tool. Try it without installing via npx:
npx weather-bandit --version
If npx can't find it, install it globally:
npm install -g weather-bandit
If the package is not yet on npm, build from source per the repo README
(git clone → pnpm install → pnpm -r build → npm install -g packages/cli;
requires Node.js 22+ and pnpm 10+).
Verify the installation:
weather-bandit --version
All examples below use the weather-bandit command. If you installed via npx
only (not global), prefix commands with npx instead, e.g.
npx weather-bandit outlook Berlin.
What Weather Bandit does
Weather Bandit fetches forecasts from the free, no-key
Open-Meteo API, cross-validates three deterministic
models (Best-match / GFS / ECMWF), derives per-day probabilities from a
30-member ensemble, and emits a structured daily "outlook" as Markdown + JSON.
It is fully deterministic and LLM-free — no API calls to an LLM, no auth, no
tracking.
Output modes
outlook prints human-readable Markdown to the terminal by default. Pass
--json for the structured Outlook object.analog prints a conditional-climatology outlook for a target period beyond
the forecast horizon. Pass --json for the structured AnalogOutlook object.
Prefer outlook for a quick look in the terminal. Prefer analog when the
target period is beyond the 16-day forecast horizon.
Commands
Print an outlook to the terminal
weather-bandit outlook
weather-bandit outlook "New York"
weather-bandit outlook Berlin --days 3
weather-bandit outlook Berlin --style briefing
weather-bandit outlook Berlin --style summary
weather-bandit outlook Berlin --style tables
weather-bandit outlook Berlin --json
Choosing a style
full (the default) appends a 24-row hourly table for today — roughly 4× the
size of summary for the same headline numbers. When you are relaying the
outlook into a conversation, pick the leanest style that answers the question:
- Conversational / "what's the weather today" →
--style summary (headline
line, daily table, climate Δ, probability bands; no hourly dump).
- "Should I bring an umbrella", "how hot", reasoning wanted →
--style briefing (adds prose + heat analysis, still no hourly table).
- User explicitly wants hour-by-hour →
full (the default). Only reach for
it when the hourly detail is the point.
- Post-processing the numbers programmatically →
--json.
Do not default to full just because it is the default — for a spoken-style
answer it wastes context on an hourly table the user did not ask for.
Analog forecast (conditional climatology)
Use analog when the target period is beyond the 16-day forecast horizon and
the question is "given how the season has gone so far, what's the weather
usually like for that target period?" Instead of averaging all 30 baseline
years equally, it finds the past years whose recent weather looked most like
right now (by RMSE of daily max temperature over a lookback window, default 21
days ending 3 days ago for ERA5 latency), selects the top-K most similar
years (default 8), and compares their target-period normal side-by-side with
the unconditional 1991–2020 normal.
weather-bandit analog Blavand --from 2026-08-21 --to 2026-09-03
weather-bandit analog Blavand --from 2026-08-21 --to 2026-09-03 --lookback 30 --top 5
weather-bandit analog Blavand --from 2026-08-21 --to 2026-09-03 --json
The output is a standalone Markdown artifact (analog years table, per-date
analog-vs-all-30 normal comparison, summary aggregates, and a skill caveat).
It is not a forecast — print it to the terminal or capture it as a plain
Markdown file.
When publishing analog output to a user, always repeat the skill caveat
verbatim so the reader does not mistake it for a forecast:
This is a conditioned climatology, not a forecast. Even the best analog was a
few °C off on average. Use as a sharper-than-average seasonal guide, and
re-run a real forecast once the target is within 16 days.
Flag reference
| Flag | Description |
|---|
[city] | City name; defaults to the first entry of the cities config (Berlin) |
-d, --days <n> | outlook: forecast days (default 7; ensemble capped at 16) |
--json | outlook/analog: print the structured Outlook/AnalogOutlook as JSON instead of Markdown |
--style <style> | outlook: Markdown style (full, briefing, summary, or tables; default full) |
--from <date> | analog: required target start date (YYYY-MM-DD) |
--to <date> | analog: required target end date (YYYY-MM-DD) |
-l, --lookback <days> | analog: lookback window length in days (default 21) |
--top <n> | analog: number of analog years to select (default 8) |
Tips
- The ensemble is queried without a
models= parameter on purpose — that
is the only way to get the 30 individual members. Do not "fix" this.
- ECMWF does not provide wind gusts;
windgusts_10m comes back as null and
renders as —. This is expected, not a bug.
- The deterministic models can disagree. The cross-validation section surfaces
agreements and disagreements; do not suppress disagreements when adding prose.
forecastDays is capped at 16 for the ensemble. Passing a larger value does
not extend the probability bands.- The
outlook command's Markdown has no frontmatter (clean for the
terminal).
- The tool adds nothing beyond Open-Meteo data plus cross-validation,
probabilities, and rendering. Do not claim accuracy the data does not
support; cite Open-Meteo as the source.
