Library and platform

whileai is two things that share one import, and it helps to know which one you are using.

	Library	Platform
What it is	A scientific Python package. Simulate, grade, measure, select, export.	A hosted service. Models, storage, training, serving, tracking.
Where it runs	Your machine. Offline with `simulator=False`.	Our GPUs and storage, behind the While API.
Needs a key	No. Bring your own model provider keys if you use one.	Yes. `whileai login`.
Calls	`simulate`, `grade`, `pass_at`, `delta_report`, `optimize`, `export_`, `verify.`, `judge_trust`, `hack_scan`	`push_rows`, `datasets`, `train`, `serve`, `cut`, `hf_publish`, `training_run`, `agents`, `whileai login`, the terminal verbs
Docs	The Library tab	This tab

What the platform adds

Hosted models. A situation writer and a judge you did not have to stand up, on your key, with a daily allowance on the trial tier. Leave agent= off simulate() and the hosted Qwen answers; leave judge= off grade() and hosted Phi-4 grades.
Storage. data.push("name") stores a run as a dataset with its holdout split, provenance and the analyzer report. Datasets are what training reads.
Training and serving. wai.train(dataset, method="sft" | "dpo" | "grpo" | "rm") runs on hosted GPUs; wai.serve("name", run) puts the adapter behind an OpenAI-compatible endpoint you can hand to local_model.
Tracking. Agents, behaviors, runs and verdicts as objects the terminal verbs read: whileai agents, whileai verdict <id>, whileai promote <id> <version>.

Everything the platform stores can be pulled back as JSONL, and every number it reports is computed by the same library code you can run locally.