Every CliAI account has three independent meters. Together they answer the question am I about to hit a limit? without ever surprising you. Here's exactly how they work.
Three windows
Total balance. A counter that decrements with each billable request. Refilled when you top up your account.
Daily window. A rolling 24-hour budget. Resets continuously β there's no "midnight reset" cliff.
Per-minute rate. A short-term rate limit. Prevents accidental spikes (a runaway loop in a script, for example) from draining your balance.
$ clai statsAccount: u_8a7cβ¦β Total balance ββββββββ 2,847 / 5,000β Daily window ββββββββ 37 / 200 (resets in 14h)β Per-minute rate ββββββ 1 / 5β Top tools today βββββ find (12), git (8), kubectl (6)If any of the three is exhausted, the next request returns a clear error and the panel shows which window blocked it β there's no mystery 429.
When caching saves you a request
CliAI caches generated commands by the hash of the prompt + shell + locale. Submit the exact same task in the exact same shell and the second time, you get the cached command β no LLM call, no balance drain.
$ clai find files larger than 100MBβ find . -type f -size +100M[cache: hit Β· balance unchanged]$ clai find files larger than 100MBβ find . -type f -size +100M[cache: hit Β· balance unchanged]The cache is per-account and lives on the server. It's invalidated when you change locale, switch shell, or refine the prompt β refinement is by definition a new request.
Whether cache hits count
By default, cache hits do not decrement your total balance β the daily and per-minute meters tick anyway, because they're rate-protective regardless of whether the work was real. If you have a strong opinion the other way (audit trails, fairness across accounts), there's a server-side flag your administrator can flip; check the docs.
Inspecting your usage
clai stats is the source of truth. The same numbers appear in the cabinet UI, but if you're doing anything programmatic β wrapping clai in a CI step, embedding it in a tool β clai stats --json is what you want.
See also
- Pricing β how the total balance is sized.
- From one-liner to script β when a request becomes a recurring CI step, the per-minute rate is the meter to watch.