A Claude Code usage tracker that predicts the 429 reads a different number than ccusage

Wrong question, because there are two non-overlapping tools and you want both. ccusage (github.com/ryoppippi/ccusage) walks ~/.claude/projects/*.jsonl and totals the tokens your CLI sent. ClaudeMeter (github.com/m13v/claude-meter) calls https://claude.ai/api/organizations/{org}/usage with your browser cookie and reads the percentage of plan that Anthropic counted those tokens against. ccusage answers 'what did this machine spend?'. ClaudeMeter answers 'when does the next prompt get a 429?'. Different inputs, different ledgers, both useful at the same time.

Because the two numbers are not the same number. ccusage sums input_tokens + output_tokens out of the JSONL transcripts on disk. The 429 is decided by the server-side seven_day or five_hour utilization on Anthropic's side, which factors in per-model weights, the peak-hour multiplier, attachments, browser-chat usage on the same plan, and tool-call overhead. The local token count and the server utilization can drift by 30 or 40 points and both still be correct readings of what they actually measure. ClaudeMeter shows the second one alongside ccusage so you can see the gap and stop guessing.

It ships a Manifest V3 browser extension that runs in the same browser you're already logged into. The extension fetches /api/organizations/{org}/usage with credentials: 'include', so the existing claude.ai cookie travels automatically. No copy-paste from DevTools, no embedded sign-in webview, no Anthropic API key (which only sees console spend, not plan quota). The seven lines that do this are at extension/background.js lines 5 to 12 in the open-source repo. Same private endpoint that claude.ai/settings/usage itself reads to draw your usage page.

It is your own session calling a route your browser already calls every time you open Settings > Usage. There is no second account, no scraping, no parallel session. The READ rate is one request per minute per browser, well under what your normal browsing of claude.ai produces. The endpoint is undocumented, which means Anthropic can change it, not that they prohibit it. If they rename a field, the open-source repo gets a same-day patch and you pull the next brew release. The risk profile sits between 'reading my own bank statement on the bank's site' and 'using a chrome extension that customizes my Gmail'. Worth it for the 429-prediction; not worth it for someone uncomfortable with that.

Not strictly, but it is the recommended path. With the extension loaded, the menu bar app gets snapshots over localhost:63762 and never touches your keychain. Without the extension, the app falls back to decrypting the Chrome Safe Storage cookie itself, which works but pops a Keychain prompt the first time on Arc, Brave, and Edge (Chrome silently approves). Most people use Route A because it skips the prompt and works across multiple browsers without configuration.

The browser extension yes, the macOS menu bar app no. The README is explicit: macOS only. On Linux or Windows you can install the extension and read the same 5h/7d numbers from the browser toolbar popup, but you don't get the always-visible menu bar chip or the localhost bridge. If you want a status-line readout on those platforms, the extension's chrome.action API still exposes the latest snapshot and a small companion script can mirror it into i3blocks/Polybar/Powershell.

No. ccusage is the only thing that breaks down spend by Claude Code session, by model, by project. If you want to see 'this refactor cost me 1.2M Sonnet tokens' you keep ccusage. ClaudeMeter does not parse JSONL at all and cannot show you which session burned which model. The two tools are complements: ccusage tells you what your Claude Code traffic weighed in tokens, ClaudeMeter tells you what fraction of the plan ceiling Anthropic counted that against. Run both. They poll different sources and answer different questions.

Eight Window structs and one ExtraUsage block, typed in src/models.rs lines 18 to 28. The Window structs are five_hour, seven_day, seven_day_sonnet, seven_day_opus, seven_day_oauth_apps, seven_day_omelette, and seven_day_cowork. Each Window has a utilization float (0.0 to 1.0+) and an optional resets_at timestamp. The ExtraUsage block carries is_enabled, monthly_limit, used_credits, utilization, and currency for the April 2026 metered-billing fall-through. The menu bar surfaces five_hour and seven_day in the title; the dropdown lists every Window, the extra-usage dollar line, and the resets_at countdown for each.

Open ClaudeMeter on a Mac later. The full path is brew install --cask m13v/tap/claude-meter (one command), then load the unpacked extension from the cloned repo into Chrome/Arc/Brave/Edge (about 60 seconds), then visit claude.ai once. The menu bar lights up within a minute with two percentages. If you only want the browser side, the extension folder is at github.com/m13v/claude-meter/tree/main/extension; load it unpacked and the toolbar badge starts ticking after one refresh of claude.ai.