Claude Pro's 5-hour window visibility, reduced to one worst-case badge on your toolbar

It comes from GET /api/organizations/{org_uuid}/usage on claude.ai, the same JSON the Settings page renders. The extension calls that endpoint with your existing logged-in cookies, parses the five_hour.utilization float, normalizes the 0-to-1 vs 0-to-100 scale, and writes it to chrome.action.setBadgeText. There is no second source of truth and no local estimation involved. If the number on the badge disagrees with the bar at claude.ai/settings/usage, you are looking at a stale poll, not a different metric.

If your account belongs to two organizations, say a personal Pro and a Team workspace, the rate limiter on Anthropic's side trips per org. You can be at 14 percent on org A and 96 percent on org B; the next prompt to org B will 429. Cycling between two badges or showing the active tab does not help, because the org you happen to have open is not always the one closest to throttling. ClaudeMeter computes worstPct(snaps, 'five_hour') across all your memberships in a single reduce, so the badge always tells you which way you are about to fall over.

They are. In extension/background.js the badge turns orange at >=80 percent and red at >=100 percent. In src/bin/menubar.rs the title pill turns orange at >=90 percent and red at >=100 percent. Same release, two different warning thresholds. The browser nudges you a notch earlier because the toolbar is a smaller, more peripheral pixel surface and we wanted the orange to fire while you still had room to wind down a session. The Mac title sits on a larger, more attention-grabbing surface and gets the tighter band.

Once per minute, fixed. POLL_MINUTES = 1 in extension/background.js sets a chrome.alarms tick that calls refresh() every 60 seconds, and the macOS app uses POLL_INTERVAL = Duration::from_secs(60) in src/bin/menubar.rs. There is no exponential backoff and no on-focus poll. If the page disagrees with the badge by more than a percent or two, wait up to 60 seconds and the badge will catch up.

Not on the pixel itself. The badge is one number, the worst case, with no org tag. Click the toolbar icon and the popup renders one row per account with the email next to each five-hour bar, so you can see which membership tripped the worst-case. The Mac menu bar dropdown does the same with one row per snapshot. Keeping the badge to a single number is intentional: at a glance you want one signal, not a cycle.

That happens. The endpoint is internal and the scale is inconsistent across buckets and across releases. The clamp is at extension/popup.js:6-11, u <= 1 ? u * 100 : u, applied before any worst-case reduction. The same clamp lives in the Rust side as part of the formatter. If you write your own visibility tool against this endpoint and skip the clamp, an org at 0.94 will look like 0.94 percent and the other org at 94.0 will dwarf it in your badge. Always normalize first, reduce second.

Not on the badge text itself. The badge is a percent. The popup row shows resets_at as a relative label like '5-hour · 3h' or '5-hour · 22m', computed from the live ISO timestamp the server returns. Hover the toolbar icon and the title attribute shows the same number expanded. The reset is sliding, not anchored to your first prompt, so the label moves on every poll if you are still active.

A cycling badge alternates between five-hour, seven-day, and other windows on a four-second loop, which means at any single glance you might be reading the wrong axis. ClaudeMeter pins the badge to one bucket, five-hour, because that is the one most likely to throttle a synchronous chat. The seven-day and weekly buckets matter, but they live in the popup, not the badge. Different design choice, same source of truth.

No. The fetch uses credentials: 'include' which means the request rides your existing claude.ai session cookies. If you log out, the next /api/organizations/{org_uuid}/usage call returns a 401 or 403, refresh() catches it, and the badge text becomes '!' on a gray background. Open claude.ai, log back in, click the toolbar icon's refresh button, and the badge recovers in under a second. There is no separate auth flow and no API key to re-paste.

Because the cookies live in your browser. The endpoint authorizes per-session, not per-API-key, and Anthropic does not publish a programmatic credential for this data. A server poller would either need you to paste cookies into a remote service, which is the manual step ClaudeMeter exists to remove, or it would simply not work. Running the poll inside the same browser that already holds the session is the cheapest path to server-truth visibility, which is why the extension is the default surface and the macOS app is the long-running one.