#Token budgets

lazy stats budget shows how much of your model usage allowance is left and where the tokens went. The same view appears on the dashboard (the Token budget box), on a Lazy Teams project's Usage page (the "Token usage" link on the project page), and to the builder agent through the lazy_usage_limits tool, which uses it to plan batches of work.

lazy stats budget          # human-readable
lazy stats budget --json   # machine-readable

#What it shows

Credential windows. For each credential lazy has seen a usage-limit reading for — for example a Claude subscription's 5-hour and 7-day windows, or a ChatGPT subscription used by Codex — the percentage used and when it resets, the tokens lazy saw spent on that credential inside the window, and from those an estimate:

• tokens left — the unused percentage × tokens spent per percent so far;
• typical turns left — tokens left ÷ the median size of a recent agent turn on that harness.

These are estimates for planning. The provider meters usage by its own rules. Only tokens lazy can tie to that exact credential are counted, and anything spent on the same account outside lazy makes the estimate come out lower than the real headroom.

When lazy cannot make an estimate it says why instead of guessing: the provider reported no percentage, the window has reset since the last reading, the window's length is unknown, lazy no longer has a record of the spend in the window, or too little of the window is used to extrapolate from.

By harness, by day, by task (last 7 days). Tokens and turns from each agent turn's recorded usage. A harness with no usage window reading (for example Cursor, or Pi on a local model) is marked tokens only: its spend is shown and no budget is invented for it.

Figures are tokens and percentages only; lazy never shows money amounts.

#Planning with it

Before starting a batch of tasks, the builder reads the budget, estimates the batch as tasks × likely turns × typical turn size, and — if it would not fit before the window resets — proposes starting the most important work first and holding the rest. Cluster tasks do the same for their waves of subtasks. This is advice to you, not a gate: to have lazy actually hold new turns near a limit, configure [usage_pause] in lazy.toml (see lazy stats limits).

#Who sees what

On a single-person install you see every credential. On Lazy Teams, a member sees the project's shared credential and their own, never another member's. A task agent sees only the credential its own turn spends and, in the task list, only its own task; the harness and day totals and the typical turn size are project-wide figures that name no credential.