I built a macOS menu bar app in ~900 lines of Python that tracks Claude + ChatGPT limits — here's how

I got tired of getting blindsided by Claude Pro rate limits mid-session. The solution: a menu bar app that reads the same private API that claude.ai/settings/usage calls.

Here's what made it interesting to build:

The Cloudflare problem

Claude's API is behind Cloudflare. Regular requests or httpx get 403'd immediately. The fix: curl_cffi with impersonate="chrome131" to spoof the TLS fingerprint. This was the single biggest hurdle.

Browser cookie detection

Rather than making users copy-paste cookies, I used browser-cookie3 to read them directly from Chrome, Arc, Firefox, Safari, or Brave. Key insight: try Firefox first — it uses a plain SQLite database with no Keychain prompt. Chromium browsers need a one-time "Always Allow" dialog.

AppKit threading rule

rumps runs on the main thread. All UI updates (menu rebuilds) must happen there. My background fetch thread puts results in a queue.Queue, and a 0.25-second timer on the main thread drains it. Violate this and you get intermittent crashes.

Inconsistent API scale

The API returns five_hour (session) as a 0–1 fraction, but seven_day (weekly) as a 0–100 percentage. The fix: if raw > 1.0: already_pct else multiply_by_100. Spent embarrassing time debugging this.

One-line install

curl -fsSL https://raw.githubusercontent.com/yagcioglutoprak/AIQuotaBar/main/install.sh | bash

It also tracks ChatGPT usage in the same menu bar indicator.

MIT licensed, ~900 lines Python: https://github.com/yagcioglutoprak/AIQuotaBar

Happy to go deep on any of these implementation details in the comments.