Pacer shows how much of each weekly limit you've used. Tell it when you want to use your limit by, for example "all of it by Friday", and how to spread it. Pacer then tells you when you're running hot and will run out early, or when you have room to lean in.
What you get
- A menu bar icon that means something. One ring per tool, filled to the share of your weekly limit you've used. A notch marks where your plan says you should be right now. If the fill passes the notch, you're ahead of plan. A ring turns orange when that tool will run out before your finish day.
- A popover with each tool's status, how much you can still spend today, and a week chart of planned versus actual use per day.
- One plan for both tools: finish day, target percentage, shape (even, front-loaded, back-loaded), spending days, and working hours. Each tool can be tracked or turned off.
- Nudges. At most one "ease off" and one "lean in" notification per tool per day, and only during your working hours.
Where the numbers come from
| Weekly and 5-hour limits | Daily breakdown |
|---|
| Codex | GET https://chatgpt.com/backend-api/wham/usage, the request behind /status in Codex. It uses the ChatGPT sign-in Codex stores in ~/.codex/auth.json. If that fails, Pacer falls back to rate_limits events in ~/.codex/sessions/**/*.jsonl. | Measured. Pacer combines rate_limits events from the logs with the live readings it saves while running. |
| Claude | GET https://api.anthropic.com/api/oauth/usage, the request behind /usage in Claude Code. It uses the sign-in Claude Code stores in your Keychain (Claude Code-credentials). | Estimated. The exact weekly total is spread across days in proportion to token counts in ~/.claude/projects/**/*.jsonl. |
Things to know:
- Claude Code doesn't write limit percentages to disk, and current Codex builds no longer write them to every session log, so Pacer reads both from the network every 5 minutes. Pacer reads the Claude Keychain item through
/usr/bin/security and the Codex auth.json file, and never refreshes or rewrites either. If a sign-in expires, run claude or codex once to refresh it.
- Both usage endpoints are undocumented and may change.
- Pacer reads only token counts, timestamps, and rate-limit fields. It doesn't store or send your prompts, code, or replies.
- Pacer has no account, server, or analytics.
How pacing works
For each tool, Pacer lays your plan over the current limit window. It takes the spending days between the window start and your finish day, weighted by the plan's shape and spread evenly across your working hours. That gives an expected-usage curve.
- Pace: compares your actual usage with the curve now after a full planned spending day. The chart shows the curve from the start.
- Plan versus reset: plans default to each tool's own reset. The card labels the expected percentage as "Plan". "Under plan" compares usage with that curve; unused capacity remains available until the provider resets.
- Catch-up: if a chosen finish day passes with capacity left, Pacer rebalances it across the remaining spending days before that tool's reset. If no spending hours remain, all leftovers become available today, including on off days, with an expiry countdown. Catch-up nudges can run on off days, still only during your working hours.
- Forecast: after a full planned spending day, assumes your ratio to the plan holds. That gives your projected usage at the finish, or the time you'll hit 100%. Before then, a short burst or off-day use is too little evidence for a run-out time.
- Allowance: divides what's left over the rest of the plan in the same proportions. That gives "left today" and "per day from here". When you're ahead of plan, today only gets what keeps you on the plan by tonight, so an overspend comes out of today first.
- Today's guidance: when there's still room in today's allowance, the status shows how much is left even if usage is ahead of the hourly curve.
- Fresh-week guidance: names the first spending day's target and whether a front-loaded or back-loaded plan tapers or builds after it.
The math is in Pacer/Model/PaceMath.swift and covered by PacerTests/PaceMathTests.swift.
Build
Requirements: macOS 15 or later, Xcode 16 or later, and XcodeGen.
xcodegen generate
xcodebuild -project Pacer.xcodeproj -scheme Pacer -destination 'platform=macOS' test
open Pacer.xcodeproj # then Run
Pacer isn't sandboxed, because it reads ~/.claude and ~/.codex and calls security. That means it can't ship on the Mac App Store.
To render the popover, setup, and main window to PNGs for design review:
TEST_RUNNER_PACER_SNAPSHOTS=1 xcodebuild -project Pacer.xcodeproj -scheme Pacer -destination 'platform=macOS' test -only-testing:PacerTests/SnapshotRender
The images land in $TMPDIR/PacerSnapshots.
To build the downloadable disk image, run scripts/make-dmg.sh. It builds Release and writes site/downloads/Pacer.dmg with a drag-to-Applications window. It signs with the HMDFV Inc. Developer ID; scripts/release.sh notarizes it.
The app icon is drawn in code. To change it, edit scripts/render-icon.swift and run swift scripts/render-icon.swift from the repo root. The README banner is scripts/readme-banner.html; run scripts/render-banner.sh to re-render it.
Updates and releases
Both the app and the site use Sparkle. The app reads site/appcast.xml from the website once a day. An update found right after launch opens Sparkle's window; one found later waits as an "Update available" tile in the popover. Check for Updates… is in the popover's ⋯ menu and in Settings.
To ship a release, bump CFBundleShortVersionString and CFBundleVersion in project.yml, commit, then run:
scripts/release.sh "What changed, in a sentence or two."
It builds the disk image, notarizes it when a ramihmd-notary notarytool profile exists, signs it for Sparkle, creates the GitHub release, adds the release to site/appcast.xml, commits, pushes, and deploys the site. The Sparkle signing key lives in the login keychain; back it up with generate_keys -x (the tool is in Xcode's SourcePackages at artifacts/sparkle/Sparkle/bin).
Roadmap
- Calendar-aware plans: lower a day's share based on hours of meetings.
- Persist scan offsets so a relaunch doesn't rescan a week of logs.
License
MIT
The Claude and OpenAI marks in Pacer/Assets.xcassets come from Simple Icons (CC0). They're trademarks of Anthropic and OpenAI, used here only to label each tool. Pacer isn't affiliated with or endorsed by either company.
