mint

Configured accounts

One row per credential. The last four characters are the only part of a credential this page is ever sent.

What each provider actually reports

The three providers do not answer the same questions, and this table is the same data the gauges are drawn from. Where a cell says no, the matching figure on the dashboard reads "not reported" rather than zero.

How the numbers are made

Requests per minute

The last complete one-minute bucket from the provider's own usage report, divided by the configured limit. The minute in progress is never used: it reads low for most of its length and then jumps, which looks like traffic collapsing and recovering once a minute.

Input tokens per minute

Uncached input plus cache writes. Tokens read from the cache are excluded, because on every current Claude model except Haiku 3.5 they do not count toward the input limit. Including them would overstate the pressure on a well-cached workload several times over.

Cost

The provider's own cost report, not a price list multiplied by token counts. The two do not reconcile exactly and the providers say so, so this app stores what it was told rather than what it could infer.

The rollup across every account

The panel at the top of the dashboard is one row per model, summed across every enabled account, so the same model called from two accounts appears once. Three things it deliberately does not do. It does not present a total over a mixture of reported and unreported figures as though it were complete: every total says how many accounts it covers and is marked partial when that is fewer than all of them. It does not attribute cost to a model the provider did not attribute it to, because a share worked out from token counts would look like an invoice line and would not be one. And it does not roll up rate limits at all: two accounts at half their limits are not one account at its limit, so headroom stays on the cards, where a number and its own limit sit together.

Nothing here costs you anything

This app never sends a request to a model endpoint. Rate limit headroom could be read from the response headers of a one-token call, and that is deliberately not done: a monitor that spends the budget it monitors is a bug with a good excuse.

People

Sign-in is decided by my.thermals.app. What somebody may do once they are in is decided here, from this app's own table. The identity token carries no role and none is read from it.

Audit