A tool to answer the question my usage meter wouldn't

Claude Code’s built-in usage display tells you what percentage of your weekly cap you’ve consumed. That’s useful. It’s not the number I actually wanted: am I going to blow the cap before it resets?

Percentage-used tells you where you are. It doesn’t tell you where you’re going. At 60% with four days left, that looks fine, unless you burn the same amount tomorrow as you have the rest of the week combined. Which is exactly the kind of session I was worried I’d been having.

So I built UsagePace. Python, standard library only.

How it works

Claude Code already writes per-message usage logs to disk: token counts, model, timestamp. UsagePace reads those directly. No API key, no network access. All local.

The trickier part: raw token counts aren’t interchangeable across models. An Opus message costs roughly five times what a Sonnet message costs against the cap. So UsagePace cost-weights each message before aggregating (multiply by a model weight, sum across the window) and you get a proxy for “how much cap have I consumed” rather than “how many tokens have I sent.”

From there it’s a simple projection: usage over the last N days, extrapolated to the reset date. Calibrate the weights once against your real usage number and it tracks from there.

What it tells you

A single number: projected cap usage at reset, as a percentage. Over 100 and you’re trending hot. Under 100 and you have room.

It also breaks down usage by day and by model, so you can see whether a spike was Opus-heavy work or just volume. Useful for figuring out whether to reach for Opus or stick with Sonnet for the rest of the week.

The limitations

It’s a proxy. Anthropic’s real cap formula is weighted server-side and unpublished. The weights in UsagePace are an approximation calibrated against observed behavior, close enough to be useful and not precise enough to bet on if you’re right at the edge.

The postscript

The gut feeling that made me build this was that I was burning through my cap too fast. First run of UsagePace: projected comfortable headroom. The feeling was wrong.

That’s still a useful result. Knowing the anxiety was noise is worth something.