Tokometer
Today and the week
One sentence on what to do about today, the week against an even spend, tokens, the 5-hour limit, and what all of it would have cost at API prices.
A token speedometer
for your Dock.
Tokometer shows how today is going against your week across every Claude Code and Codex subscription, from a small piece of glass between your icons. Underneath, a Rust proxy hands your sessions to whichever account has room.
Download for Mac Free · macOS 13 or later · Apple silicon · updates itself
Two numbers, both honest.
Daily pace is today's use against today's share of the week. The share is what was left at midnight, spread evenly over the time to each account's reset, and it stays fixed for the day. 100% means today used exactly its part. A heavy day goes past 100%, and the days after it get a little less.
Week is how much of all your weekly allowances is used. Close the ring every day and every week ends at exactly 100%: nothing wasted, nothing borrowed.
- Blue room to use
- Green on pace
- Orange borrowing from later days
- Red out: at a limit, or signed out
One click for the rest.
Accounts
How many Claude Code and Codex accounts you have, when each one's week renews, and the one thing an account needs when it needs one: sign in, renew, or spend a saved reset.
The calendar
A square a day, a month at a time, darker for the busier days. Swipe back through earlier months; point at a day for its tokens, its cost and its pace.
Rust, measured.
Tokometer runs CLIProxyAPI-Rust, a single-binary rewrite of the Go CLIProxyAPI. It serves your subscriptions to Claude Code, Codex and any OpenAI or Anthropic client, and when an account reaches its 5-hour or weekly limit, new sessions move to the account whose week renews soonest. We ran both side by side on the same Mac.
| Rust | Go 6.9.5 | ||
|---|---|---|---|
| Requests per second | 85,700 | 58,700 | 1.5× more |
| Latency, 99th percentile | 1 ms | 4 ms | 4× lower |
| CPU per request | 30 µs | 71 µs | 2.4× less |
| Memory at idle | 9.4 MB | 42 MB | 4.5× less |
| Memory after 20,000 requests | 12.4 MB | 51 MB | 4× less |
| Binary | 9.1 MB | 39.8 MB | 4.4× smaller |
How these were measured. Apple M3 Pro, 18 GB, macOS 26.5. Each proxy ran with an empty configuration and no accounts, on its own port. Throughput, latency and CPU come from ApacheBench (ab -k -n 20000 -c 32) against /v1/models after a 1,000-request warm-up; CPU is the process's CPU time over the run divided by requests. Memory is resident size from ps. Every figure is the median of five runs of each, alternating, on a Mac in ordinary use. These measure the proxy's own overhead, not a model's response time, which is set by the provider.
The app is held to the same standard. The widget redraws only when a number changes, reads the Dock's layout only when something in the Dock changes, and watches no mouse movement while it sits in the Dock. At rest it uses less than 0.1% of one core and 18 MB, and the proxy beside it 4 MB.
Really in the Dock.
The Dock holds square tiles and nothing else, so apps that put wide widgets "in the Dock" usually draw a Dock of their own and hide Apple's. Tokometer stays in the real one. It keeps three empty spacer tiles next to the Trash and lays its glass exactly over them, working out where they are from the Dock's own outline. It asks for no permission.
Quit it and the tiles go with it. Move the Dock, change its size, open apps: the widget follows.
What it costs, at API prices.
The proxy counts tokens by model, keeping cache reads and writes apart, and Tokometer prices them with LiteLLM's public price list, refreshed weekly. Your subscription isn't billed this way; the figure is what the same use would have cost without one.
Make every week count.
Open the disk image, drag Tokometer to Applications, and open it. Sign in to Claude Code or Codex from the shelf; point your tools at http://127.0.0.1:8317.