Your usage, where you work.
See actual remaining allowance, regular reset time and the last successful update. Missing windows stay missing; failed reads stay visibly unknown.
OPEN SOURCE. ON YOUR DEVICE.
Your AI usage.
Your next reset.
One glance.
Know what’s left. Get a nudge before you run low. Catch an unexpected reset while the conversation is still unfolding.
AI Usage Sentinel
Codex 5-hour72%
Regular reset in 2h 14m
Weekly61%
Regular reset in 3d 7h
Possible reset signal
↗Early Signal · 38% confidence
Illustrative interface. These are sample values, not live account data.
Settings preview
No cookies collectedOptional, consent-based analyticsEvery signal has a sourceFive interface languages
01 / A QUIETER COMPANION
Less checking.
More knowing.
A reminder at your pace.
Set up to five remaining-quota thresholds. One alert per stage per cycle, with a one-hour pause when you need some quiet.
Notice the unexpected.
A quota jump before its previous reset time becomes account evidence. Normal resets stay quiet. Banked, purchased and complimentary events keep their own labels.
02 / FIVE STAGES. YOUR CHOICE.
A little heads-up.
Before zero.
Add, remove or edit distinct thresholds from 0–99%. If several stages are crossed in one update, Sentinel sends one consolidated alert. Restarting the app won’t repeat delivered reminders.
Try the example. Move the allowance down, then reset the cycle.
REMINDER DEMO%
Browser-only example. No account connection or system notification.
03 / EARLY DOESN’T MEAN CERTAIN.
A signal.
With the evidence
still attached.
OpenAI official updates, GitHub and Reddit reports are checked independently. Related reports are grouped; duplicate text and authors don’t inflate confidence. Your own quota jump can strengthen an existing signal.
Reset Watch & Forecast ↗How confidence is calculated ↗
Rumor10–29%
Limited evidence. An invitation to look closer.
Early Signal30–59%
Fresh community reports. No official confirmation.
Default alerts start at 25% confidence.Likely60–89%
Stronger corroboration, still not official proof.
Confirmed90–100%
Official evidence. Read the source for eligibility.
04 / DIFFERENT AGENTS. HONEST DATA.
One place to look.
Choose a provider. The app shows one selected agent at a time and keeps its history local. Integration support depends on what each vendor officially exposes.
Codex
Automatic local readUses the official local app-server after Codex sign-in. Actual returned quota windows; no inference prompt.
Claude
Official status-line bridgeImports quota-only JSON from Claude Code’s documented status line. Requires setup and a session publishing rate-limit data.
Gemini
Local JSON / manualAccepts authorized local quota exports or manual values. No unsupported personal-usage endpoint is assumed.
Grok
Local JSON / manualAccepts local quota evidence. API rate limits are not substituted for a SuperGrok subscription allowance.
Public reset monitoring currently focuses on OpenAI/Codex. Other agents’ news feeds are not yet integrated. Stale exports never become fresh usage.
Read the agent setup guide ↗YOUR USAGE BELONGS TO YOU.
Local by default.
Account history stays local by default. Optional daily analytics requires consent; coarse quota bands require separate consent. Public news requests carry no account history or credentials. The official Codex client handles its own sign-in.
Privacy & security details ↗05 / SMALL APP. OPEN INVITATION.
Make room
in your menu bar.
Version 1.9.0 · Free, open source, no account required by Sentinel. An official signed-in Codex installation is needed for live Codex usage.
macOS
macOS 14+ · Apple Silicon & Intel
SwiftUI / MenuBarExtra
Ad-hoc signed, not notarized. Follow macOS Privacy & Security’s explicit Open Anyway flow. Keep system protections enabled.
Windows Beta
Windows 10/11 · x64
Native Qt system tray
Extract the whole folder, then run UsageSentinel.exe. Unsigned preview; review the release and your system’s trust prompt.
Linux Beta
Linux x64 · glibc 2.35+
Qt tray / compatible desktop
Extract and run UsageSentinel/UsageSentinel. Some desktops need Qt libraries or a tray extension; a window is available as fallback.
Only after consent: Cloudflare derives country from your IP; D1 stores no IP. This counts download clicks, not completed downloads. / 同意後才回報，不影響下載。 Privacy & public insights / 隱私與公開統計 ↗
Release notes & checksums ↗Windows / Linux setup ↗Build from source ↗
Windows/Linux are beta ports: CI builds and smoke tests do not replace real desktop notification, sleep/wake and login validation. macOS has richer charts and menu styles. See the platform guide for differences.
A FEW THINGS TO KNOW.
Does it show every ChatGPT limit?
No. Live Codex usage contains only the windows the official provider returns. Other ChatGPT personal limits may be unavailable. Missing data is shown as unknown, never as an invented percentage.
Does an early signal mean I will get a reset?
No. A community report can be account-specific, mistaken or part of a limited rollout. Notifications include confidence and source links. Sentinel does not predict a next global-reset date.
Will notifications always appear immediately?
Detection runs on successful polls, normally every five minutes and after wake or opening the macOS menu. Sleep, Focus, Do Not Disturb and desktop notification support can delay or hide alerts.
Can I contribute or request another agent?
Yes. Provider adapters, translations, bug reports and platform testing are welcome. Link official documentation and keep private account data out of issues. Contribution guide ↗