Use with an AI agent

Every prompt below tells an assistant where to read first: the STAC catalog, which is the source of truth for every dataset, and llms.txt, which explains how the datasets are shaped and opened. Copy one, paste it into an assistant that can fetch URLs and run code, and add what you want to make. Text in angle brackets is yours to fill in. Each dataset's catalog page has a prompt for that dataset alone. Every prompt here was tested cold in Claude Code and Codex; an assistant that can only browse will get as far as reading the catalog.

Set up a coding agent

One line. The agent fetches agent-setup/prompt.md, installs the Python path, verifies it with one read, and asks what you want to build.

Fetch and follow the setup instructions at https://dynamical.org/agent-setup/prompt.md

The minimum

Any assistant. The STAC catalog is the one thing it must know.

Use the free, open weather and climate data published by dynamical.org. The STAC catalog at https://stac.dynamical.org/catalog.json is the source of truth: read it first, then the collection.json of any dataset you use. Each collection lists the variables with units, the dimensions, the extent, the license, and the icechunk-https asset to open, and its examples show how to open it. No credentials are needed. Resolve storage URLs from STAC at run time; never hard-code them. Select the variables, the region, and the time window before loading anything. Then ask me what I want to make.

With the conventions

Any assistant that can fetch URLs. Adds llms.txt, so it also learns how the datasets are shaped.

Help me work with the open weather and climate data from dynamical.org. Read these before writing any code:
1. https://dynamical.org/llms.txt — what the datasets are, how they are structured, how to open them, and the conventions.
2. https://stac.dynamical.org/catalog.json — the catalog. Each dataset's collection.json lists the variables with units, the dimensions, the extent, the license, and the icechunk-https asset.
Open a dataset with dynamical_catalog.open("<dataset-id>") from the dynamical-catalog package (Python 3.12+), or with icechunk>=2 and xarray from the collection's icechunk-https asset. Resolve storage URLs from STAC at run time; never hard-code them. Select the variables, the region, and the time window before loading anything; the archives run to terabytes. Forecast datasets are indexed by init_time (the model run, UTC) and lead_time; analysis datasets by time. Choose the dataset by domain, resolution, and forecast horizon, and tell me which one you chose and why.

A chart for a presentation

No code required of you. Routes the assistant through the point API, which returns JSON.

I am making a presentation and want a chart of the 2 m temperature forecast for <city> for the next 7 days, from dynamical.org's free open weather data, with the source credited. Read https://dynamical.org/llms.txt first. The simplest route is the point API documented at https://dynamical.org/api/: POST https://api.dynamical.org/v1/forecasts with a dataProductId of noaa-gfs-forecast, the city's latitude and longitude, variables ["temperature_2m"], and validTimeStart now and validTimeEnd 7 days from now (UTC). It returns JSON for the latest model run with no credentials. Convert the times to the city's local time, label the units, save a clean chart as a PNG I can drop into a slide, and caption it "NOAA GFS forecast via dynamical.org, run <initTime>".

Compare a forecast with its analysis

Analysts. Scores a forecast dataset against the model's own analysis.

Using dynamical.org's open data, compare the NOAA GFS 2 m temperature forecast with its analysis at <latitude, longitude> for valid times in the last 30 days. Read https://dynamical.org/llms.txt and https://stac.dynamical.org/catalog.json first. Use noaa-gfs-forecast (indexed by init_time and lead_time) and noaa-gfs-analysis as the reference (indexed by time; it is the model's own estimate of what happened, not observations, so say so). Open a dataset with dynamical_catalog.open("<dataset-id>") from the dynamical-catalog package (Python 3.12+), or with icechunk>=2 and xarray from the collection's icechunk-https asset. Resolve storage URLs from STAC at run time; never hard-code them. Select the variables, the region, and the time window before loading anything; the archives run to terabytes. Keep only init_times whose valid times (init_time + lead_time) fall in the window, drop pairs where either side is missing, and report bias, RMSE, and sample count by lead time with a plot.

A multi-model ensemble

Practitioners. Combines several models into one forecast with spread.

Build a multi-model ensemble forecast of 10 m wind speed for <location> for valid times from now until 5 days from now (UTC), from dynamical.org's open data. Read https://dynamical.org/llms.txt and https://stac.dynamical.org/catalog.json first, then choose from noaa-gfs-forecast, ecmwf-aifs-single-forecast, ecmwf-ifs-ens-forecast-15-day-0-25-degree, noaa-gefs-forecast-35-day, and dwd-icon-eu-forecast-5-day by domain, variables, and lead times. Open a dataset with dynamical_catalog.open("<dataset-id>") from the dynamical-catalog package (Python 3.12+), or with icechunk>=2 and xarray from the collection's icechunk-https asset. Resolve storage URLs from STAC at run time; never hard-code them. Select the variables, the region, and the time window before loading anything; the archives run to terabytes. Use each one's latest init_time and only the leads inside the window, compute wind speed from u and v, and align on valid time at the coarsest common step. Weight models equally and members equally within a model; show the mean, spread, and each model, and report any window a model cannot cover.

Connect the MCP server

Coding agents that can add tools. The server reads the live catalog.

Add dynamical.org's MCP server so you can search its open weather-data catalog directly: URL https://mcp.dynamical.org/mcp, streamable HTTP transport, no authentication. Its tools are search_catalog, get_dataset_info, get_access_pattern, and list_recent_runs. Configure it for this session or this project only, keep my existing servers, and ask before making any user-wide change. If you cannot add MCP servers, read https://dynamical.org/llms.txt and https://stac.dynamical.org/catalog.json instead. Then ask me what I want to build.

Migrating code that still reads data.dynamical.org? The migration guide has a prompt for that.

What the agent reads

stac.dynamical.org/catalog.json — every dataset's variables, dimensions, extent, license, and the Icechunk asset to open.
dynamical.org/llms.txt — the conventions, the two ways to open a dataset, and a one-line index of the datasets.
dynamical.org/agent-setup/prompt.md — the setup instructions behind the pill: read, install, verify, hand off.
mcp.dynamical.org/mcp — the same catalog as MCP tools, for agents that can add them.
The point API — forecast and analysis time series as JSON, no data libraries needed.