Tokens and plan usage

Your plans pay for the models

Winskel never charges for AI usage. Claude Code and Codex run under your own accounts, so the tokens they use count against your Anthropic and OpenAI plans and their limits, as when you run them yourself.

Per-use billing is off by default

Winskel runs Claude Code and Codex only on your subscription sign-ins, and refuses a run that would bill per use, such as a sign-in with an API key. Provider API keys and endpoint settings in your environment are not passed to the agents it starts. To allow per-use billing on one Mac, open Settings, Advanced, This Mac, and confirm; you can stop it there again.

What adds to usage

• The agent's own work: your request, the files it reads and the commands it runs, as when you run it directly.
• Winskel's instructions and the short memory brief that go with each run, and any earlier chat you reference with @.
• A plan-first or review pass, when your Custom settings ask for one. Under Send these run on the same model.
• A repair turn when checks fail, and at most one move to a stronger model.
• Orchestrate: a planner, a model for each mission, and possibly an independent review. It may use more of your plans' usage than a single model, so Winskel asks before it starts.

What Winskel does not add

• Under Send, no separate planning or evaluating model runs beside the one doing the work.
• Under Send, choosing the model does not call a model.
• When an agent resumes after your answer, Winskel sends your answer to the same session; the memory is not sent again.
• Resuming verification of kept work checks the same code again without a model call.

Seeing your usage

Settings, Usage and the status bar show the latest usage windows and reset times Claude Code and Codex reported. A reading updates only when that agent runs, so it can be stale or missing. Winskel also keeps the token counts each provider reports for a run, and records Unknown when a provider does not report them. See provider usage.

Does Winskel lower token use?

Winskel has not published a measured comparison with running Claude Code or Codex directly, so it makes no claim that it lowers token use. What it adds to each run is listed above. To spend less of a plan, choose a lighter model in the Model menu, or exclude models in Preferences.