Monitoring Claude usage limits

Last updated: 8 September 2026 · Deutsch

Two windows, plus one per model

A Claude subscription quota runs in two windows at the same time. The short one covers five hours (300 minutes), the long one seven days (10,080 minutes). Each arrives as a percentage of that window already used, together with the moment it resets. The card shows both, because either can be the one that stops you: a dense afternoon fills the five-hour window while the weekly one is barely touched.

On top of those there can be model-scoped weekly windows — Opus, for example. AI-Cockpit reads them as a list, not as a fixed set of fields: every entry the interface marks as a weekly-scoped limit becomes its own window, labelled with the model name the response supplies. Claude Code builds its own “Current week (Opus)” line out of the same list. If the interface sends no such entry, no such window appears — the app does not invent one to fill the space.

Where the figures come from

One address: https://api.anthropic.com/api/oauth/usage. It is reached with the OAuth sign-in you complete inside the app, not with an API key — there is nothing to paste for this card. The resulting token is stored in the macOS Keychain, and in the device keychain on iPhone and iPad. It leaves the device only as authentication back to Anthropic.

An undocumented interface, said plainly

Anthropic does not publish this endpoint. There is no reference page for it, no promise that it will keep its shape, and no guaranteed request rate. That is worth knowing before you rely on the numbers: if Anthropic changes the response, the card can go quiet, and no amount of care on this side prevents that.

What the app does instead is read tolerantly. Utilization, used percent and plain percent are all recognised as ways of expressing the same figure; a reset time is accepted both as an ISO timestamp and as Unix seconds; unfamiliar spellings of a window do not bring the whole response down. And when Anthropic answers with a rate-limit refusal, the app respects the wait it is given rather than hammering on at the usual interval — asking again immediately is what keeps a throttle alive.

How the forecast is calculated

The forecast answers one question: at the pace of the last hour and a half, when does this window run full? The app fits a straight line through the measurements from the last 90 minutes by least squares, which gives a rate in percentage points per hour. From there it is arithmetic: hours remaining equals the distance to 100 per cent, divided by that rate.

Four conditions keep that from becoming guesswork:

• History has to be switched on. Without stored measurements there is nothing to fit a line through.
• At least two measurements, spanning at least ten minutes. Below that the rate swings so hard that stating it would mislead.
• Measurements from before the last reset are dropped. A window falling from 90 per cent to zero would otherwise produce a negative slope and swallow every warning.
• Nothing is projected more than 24 hours out. Past a day the statement is worthless, so the app makes none — it still shows the rate, just no exhaustion time.

A flat or falling curve gets no exhaustion time either. The card also reports how many measurements the rate rests on, which is the honest way to say how much weight it deserves.

When it warns

The colour thresholds are not the same on every device, and that is deliberate. On the Mac a window turns orange at 75 per cent and red at 90. On iPhone and Apple Watch it is 80 and 95: you glance at a phone far less often than at a menu bar, and a warning that arrives too early becomes wallpaper.

The forecast can also send a notification before a window fills. That only happens in a narrow band — more than five minutes and less than four hours of predicted time left. Under five minutes it is no longer a warning, over four hours the projection is too vague to interrupt you with. It fires once per reset cycle, not once per refresh.

What it cannot show

• No token consumption. The interface reports percentages of a window, not tokens spent. Anything token-shaped on this card would be invented.
• No history from the provider. Anthropic returns the present state and nothing else. Every curve and every forecast rests purely on measurements this app collected on your own device.
• Nothing about the account holder. The response also carries the account name and e-mail address. Neither is evaluated, stored or displayed.
• The plan tier is in no token. The stored token holds an access token, a refresh token, an expiry and its scopes — nothing more. On the Mac the tier can be read from a local Claude Code file; on iPhone that file does not exist, so the badge can simply be missing. A missing badge is not a broken card.

The short answers to the questions this raises most often — empty cards, keys, folder access — are on the FAQ page. What happens to the sign-in token, what was reviewed and what was not, is set out on the Security page.