Model archive
Every cycle, kept, and actually queryable
Operational forecasts and reanalyses from nine centres, rechunked into Zarr so you can ask for one variable over one basin instead of pulling a 4 TB run.
Why this is not just a file server
Rechunked for the question you ask
GRIB is chunked by field. Time-series questions want the opposite. We store both layouts, so a 40-year series at one point is 200 ms, not 200 hours.
Subset before you transfer
Variable, level, time range, bounding box, ensemble member. The API returns exactly that, in Zarr, NetCDF or GRIB2.
Complete cycles
Not a rolling ten-day window. Every operational cycle since 2016 is retained, including the ones the originating centre has deleted.
Ensembles included
All members, all perturbations, not just the control run and the mean.
Reanalyses too
ERA5 and its land and single-level products, regional reanalyses, and the CMIP6 subset most people actually use.
Metadata that survives
CF conventions, original GRIB keys retained as attributes, and a provenance record for every rechunking step.