# Tabular Score API

> Train a tree ensemble on your rows, then score at up to 570x scikit-learn's single-row speed on Mojo kernels

MCP: https://score.mojojojo.cc/mcp (JSON-RPC over POST). OpenAPI: https://score.mojojojo.cc/openapi.json.
Auth: `Authorization: Bearer mj_live_...`; create a key at https://mojojojo.cc/keys. Usage is billed from mojojojo credits.

## Pricing

- Rows trained or scored: $0.002 per 1k rows (100 free per month)

## Endpoints

- `POST /v1/models` (MCP tool `train_model`): Train a binary classifier on rows (list of objects) and the target column name; returns holdout AUC. Models are private to your account.
- `GET /v1/models` (MCP tool `list_models`): List your trained models.
- `POST /v1/models/{name}/predict` (MCP tool `predict`): Score rows with a trained model; returns probability of the positive class and a label per row.
- `DELETE /v1/models/{name}` (MCP tool `delete_model`): Delete a trained model.

## Limits and measured results

- Binary targets only; up to 200,000 rows and 200 feature columns; training is capped at 120 seconds. Rows with numbers and strings; missing values are imputed.
- Models are private to your account. The response includes a holdout AUC from a 80/20 split when you send at least 100 rows.
- Serving is exact: probabilities match scikit-learn to 1e-15 on the shipped benchmark models.
- Speed versus scikit-learn on 2 cores: 12-570x at batch 1, 2-8x at batch 64-1024, 1.0-4x at batch 8192 (gradient boosting is only 1.03x at large batch).
- Not for credit, hiring, insurance or medical decisions.