Pick a field on the map and press Get recommendation.
You'll get the rate, its yield range, the economics, and how confident the model is.
Fertilizer products
Crop output prices
Each model declares a domain of validity — the crops and geographies where it has been trained or validated. Outside that domain a request still returns, but flags (e.g. zone_level) and a lower confidence grade say so.
Validated geographies
Every model is pluggable, carries a FAIR model card, and must clear the eight release gates before it is fully released. Built-ins are servable but not yet fully released (agronomist review + field-validation benchmark pending).
The observability & learning plane, read back: how much advice was served, how much returned as on-farm outcomes (adoption), how the models score against held-out field data, and whether enough outcomes have accumulated to retrain.
Model performance · field validation
Retrain loop
Advice served
Write actions that feed the data + learning planes. When the API has auth enabled these need a key with the right scope (outcomes, ingest, admin); it's remembered in this browser only.
Report an on-farm outcome
Ingest data
Retrain
A versioned, documented REST contract (and an MCP server for AI agents) over the same engine. FieldContext in, Prediction out — the recommendation carries its uncertainty, economics, reasoning, benchmark and provenance.
Endpoints
Explore
Interactive docs, the OpenAPI spec, and a technical console.
Auth: send X-API-Key; scopes gate crops/countries; the partner becomes the recommendation tenant. An MCP server exposes the same tools to AI agents.