Accuracythe one others get wrong
Real dollars, not guesses
Most trackers price cache tokens wrong and under-bill by half. JP reads the 5-minute vs 1-hour cache split and charges each correctly — so your total matches your invoice to the cent.
Claude · Codex · Gemini & more
Know exactly what your AI agents cost.
JP is a highly opinionated tortoiseshell who sits in your menu bar and tallies every token your coding agents burn — in real dollars, across every tool, as it happens. She's very particular about accuracy. You'll be glad she is.
- ✓Works offline
- ✓No account
- ✓7-day free trial
- ✓One-time price
XcodeFileEdit◆$4.21
Today · Mon
$4.21
▲+$0.32
last hour
Last 14 days$58.40
Claude Code$3.10
Codex$1.11
Max plan · 5-hour window62%est.
resets in 2h 14m
Tallies every agent you run
Claude CodeCodexGeminisoonCursorsoonCopilotsoonAidersoonOpenCodesoonWindsurfsoonClaude CodeCodexGeminisoonCursorsoonCopilotsoonAidersoonOpenCodesoonWindsurfsoon
How JP works
Four small jobs, done impeccably.
1Track
It just sees everything
JP reads the session logs Claude Code, Codex and friends already write to your Mac. No setup, no API keys, no account — open the app and your spend is already there.
2Cost
Real dollars, done right
Every token is priced from the live model rates — including the cache read/write split most tools get wrong. The number in your menu bar is the number on your bill.
3Limit
Know your headroom
See how much of your Claude Max or Codex window you've burned, with a clear 5-hour and weekly gauge — before you hit the wall mid-refactor.
4Alert
A gentle tap, never a surprise
Set a daily budget and JP gives you a quiet nudge when you cross it. No dashboards to babysit — glance at the menu bar and get back to work.
What's inside
The spend-and-limits brain your agents never came with.
Limits
Plan headroom at a glance
Live 5-hour and weekly gauges for Claude Pro / Max and Codex, with honest (official) vs (est.) labels so you always know what's measured and what's modelled.
History
Memory that outlives the cleanup
Claude Code wipes its logs after 30 days. JP keeps your history forever in a tidy local store — per day, per model, per project — and exports to CSV.
Attribution
Per-project & per-branch
See which repo — and which git branch — is eating your budget. Finally answer 'what did that feature actually cost me?'
Budgets
Quiet, useful alerts
A daily or weekly cap with a gentle notification when you cross it. No noise, no nagging — just a heads-up before the bill surprises you.
Native
Light, fast, at home on macOS
A few megabytes of memory, a crisp menu-bar readout, and a daylight design that doesn't fight your Mac. No Electron, no web view, no fan spin.
Honest numbers
A meter you can actually trust.
Other apps blur the line between a measured number and a guessed one. JP never does. Every plan figure is stamped official when it comes straight from the provider, or est. when it's modelled from your token counts. You always know which is which.
Never leaves your Mac
- Your prompts & code
- Your transcripts
- Your token history
- Your API keys
All JP ever sends
- A salted license check
- (only when you buy)
- No analytics by default
- No account, ever
Plan usageClaude Max 20×
5-hour window62%est.
Weekly · all models38%official
Weekly · Opus71%official
Why two labels? Providers only publish exact limits through a signed feed. When JP has it, you get the real number. When it doesn't, you get a careful estimate — clearly marked, never disguised.
Pricing
Pay once. It's yours.
No subscription, no account, no creeping monthly fee for a menu-bar app. One calm price, free updates, and a cat that earns its keep.
$12.99one-time
For one Mac · +$6 per extra Mac
Download — start your 7-day free trial✓ 7-day trial, no card✓ 14-day refund✓ macOS 14+
- ✓Every provider — Claude, Codex, Gemini & more
- ✓Real-dollar cost, priced to the cent
- ✓5-hour & weekly plan gauges
- ✓Unlimited local history + CSV export
- ✓Per-project & per-branch attribution
- ✓Budget alerts
- ✓Free updates
Apple Silicon & Intel · signed & notarized by Apple · Homebrew cask available
Questions
Good questions, honest answers.
How is this different from the free usage tools?+
Free CLIs and menu-bar tools mostly show token counts and rough percentages — and several price cache tokens incorrectly, under-billing by as much as half. JP leads with accurate dollars, keeps history past the 30-day log cleanup, adds per-project attribution and budget alerts, and wraps it in a design you'll actually enjoy glancing at.
Isn't this just another notch app like the cat-and-island ones?+
Different job. Those are agent control panels — approve prompts, jump to terminals. JPtheCat is a cost-and-limits meter: it answers "what am I spending and how much headroom is left?" It's the app you keep open alongside everything else.
Who is JP, exactly?+
A highly opinionated tortoiseshell with impeccable taste and zero patience for wasted tokens. She's lovable, a little bossy, and extremely picky about who she works with — which is exactly why she's so good at keeping your AI spend honest.
Is my data private?+
Completely. JP reads the local session logs your tools already write and never uploads your prompts, code, transcripts, or token history. The only network call it ever makes is a salted license check when you purchase — and analytics are off by default.
Which tools does it track?+
Claude Code and Codex at launch, with Gemini, Cursor, Copilot, Aider, OpenCode and Windsurf rolling in. If a tool writes usage to your Mac, JP aims to read it — no API keys required.
One-time or subscription?+
One-time. $12.99 buys a license for one Mac, with free updates. Extra Macs are $6 each. No subscription, no account.
Does it work offline?+
Yes. Everything is computed on-device, so JP works on a plane with no internet. Your license is verified with a long offline grace window, so you're never locked out of your own data.
How accurate is the cost?+
It's priced from the same per-model rates as your invoice, including the 5-minute vs 1-hour cache-write split and cache-read discount. Where a provider exposes an official figure, JP shows it; otherwise it's clearly marked as an estimate.