The quota side
The 5-hour session window claude.ai reports for your account, and the rate-limit windows Codex CLI publishes for itself. Taken as the server states them, not rebuilt from your own logs.
A meter for the limit, not the receipt
Know when
you'll run out.
Prove what
you used.
A token count says what you already spent. TokenWatch says how much of your limit is left, and roughly when you will hit it. The number sits in the menu bar and changes colour, so you do not have to ask. Thirty days stay on the record.
7-day trial · $1.99 once · No subscription
Current sessioni
5h
0%
Resets in 2h 15m
1h 47mi
To limit
Opus
62%
Sonnet
15%
Fable
9%
All models
28%
At this rate the weekly limit runs out on Sunday
Code, 5h
$211.98
Code, today
$423.97
Weekly resets
5d 11h
What a log cannot tell you.
Your transcripts hold every token you spent. They do not hold what your plan allows, so they cannot say how much is left or when you get stopped. TokenWatch reads that side too.
One limit, several places
claude.ai, Claude Code and Claude Desktop draw on the same session limit and the same weekly limit. Anthropic documents it. One dial covers all three, and says what Claude Code put into the window you are in now.
Answered before you ask
The percentage sits in the menu bar and changes colour as the window fills, so one glance is the whole check. The warning lands at your own threshold, while there is still room to finish.
Two things measurement does not settle. That mix came off one Max account over a month, where Claude Code drove nearly all of the movement, and yours will look different. And a heavy Claude Code run can take an hour or two to appear, because the server's own figure lags the work. The app says so on the dial rather than leaving you to wonder why it is not moving.
See it run.
The menu bar, the panel, then the record window. Stills below, if you would rather not press play.
The menu bar · one pill per agent, live
The countdown · how long you have, not just how much is gone
Codex analytics · every model, every tool call
Claude Code · read from ~/.claude, no key
Codex CLI · same plate, its own rate-limit windows
The record · thirty days after the server moves on
Widgets · one agent per widget
Alerts · your thresholds, your quiet hours
Data access · folders you grant, revocable
The parts list.
Everything above runs on one read of your own machine and one call to your own account. This is what comes with it.
Pace
Burn rate measured from your own recent readings, a countdown to the next reset, and the weekly limit projected forward to the day it runs out at this rate.
Record
Thirty days archived, an hourly heatmap of when you actually work, this week against last, and CSV export for Claude sessions and Codex history.
Alerts
Your own warning and critical thresholds, 80% and 95% until you change them. Quiet hours, so nothing fires in a meeting or at 3am. A notification when the window opens again.
Weekly model limits
Every model Anthropic reports a weekly limit for gets its own meter and its own reset, under the name the API gives it. A new model needs no app update.
Inside the window
What the current window is made of: the split by model, what those same tokens would have cost without prompt caching, and a line when it is running at twice your own recent average.
Claude Code
Tokens by model, sessions, active days and messages, parsed from ~/.claude. Cache read and cache write get a line each, often 95% of the traffic. What Code has spent today and inside the window you are in is priced from those same local logs, with no key.
Codex CLI
Whichever rate-limit windows Codex CLI publishes, with their resets, plus cache-hit ratio, daily sessions and per-project tokens, parsed from ~/.codex.
Widgets
Small, medium and large, for the desktop and Notification Center. Each one points at Claude, Claude Code, Codex, or two side by side.
Cost tracking
Optional, and separate from the Claude Code figures above, which need no key. An Anthropic Admin API key adds your organisation's billed spend from the cost report: today, this month, the models driving it, and a budget alert.
macOS fit
Siri, Spotlight and Shortcuts. Launch at login. English or Turkish, switched in Settings.
Any one agent is enough on its own. Without a claude.ai subscription you lose the dial, the ETA and the projection, and TokenWatch still reads your local Claude Code and Codex activity.
No server in the path.
TokenWatch has no account, no analytics and no backend. It reads your claude.ai figures through the session you are already signed in to, and your local logs where they already sit.
Your own session
The app opens claude.ai in a window and you sign in the way you always do. No password to hand over, no token to paste, and nothing read out of a credentials file on disk.
No developer server
Nothing passes through anything of mine, and nothing is collected. Your Mac talks to Anthropic for the readings, to your identity provider while you sign in, and to Apple for the purchase.
Keychain
Your Admin API key sits in the macOS Keychain, encrypted at rest. Never plain text, never UserDefaults.
Sandbox
Outgoing network only. Local logs need a folder you pick yourself, and one click in Settings takes it back.
One week free.
Then $1.99, once.
No subscription, no ads, no tracking. One purchase covers every Mac you own and every update that follows.
7-day trial
Every part unlocked. No card, and nothing to cancel.
$1.99 one time
A single in-app purchase, all your Macs, forever.
All future updates
New parts ship free to everyone who unlocked.
Bought it earlier
Anyone who paid for TokenWatch before keeps everything. Nothing to do.
macOS 14 Sonoma or later · Universal binary, Apple Silicon and Intel