Car dash for your Claude usage

The hardest part was owning the whole thing end to end, and the speed bumps that come with that. Halfway in I hit a blocker I thought had killed the live version of the product, and I pivoted to other in-product solutions before I found a way back to what I wanted in the first place.

Here is what actually happened. There is a first party usage endpoint that returns your real limits on demand, in about 200 milliseconds, costing zero tokens. I had ruled it out early, because reaching it needs your sign in and my hard rule was that this never touches credentials.

What I had missed is that I was treating two different things as one. The rule I actually cared about is that the web page never touches credentials, and that rule exists for reasons specific to web pages. A page is untrusted by default, it runs whatever it is served, and you cannot inspect it before it executes. None of that describes a program you install yourself and can read first.

So I split the rule. A small helper on your machine reads the sign in that Claude Code already saved there, in process, and it never logs it or serves it. What reaches the browser is percentages and reset times with nothing else attached.

That cleared the roadblocks all at once. Any browser, no folder picker, nothing written to your settings, no restart. It polls on a schedule instead of waiting for you to type, and because it reads the limiter directly it counts your usage everywhere, including claude.ai and your phone.

It did cost me a way in. There had been a third route, where you dropped in a folder of past sessions, and it was carrying real weight while live was still uncertain. Once live worked I had three overlapping promises where two would do, and the first thirty seconds is where people decide whether this is for them, so I cut it for the experience rather than because it was useless.

Demo stayed, and it is not a mockup. It is a generated five and a half hour session running through the same code as live data, with a scrubber you can drag through a cache miss, a hot context window and a reset. The generator knows the true conversion rate, so any drift between that and the recovered one is a bug, and it has caught real ones.

Designing around what you don't know

Fresh numbers are not the same as certain numbers. A helper can be stopped, a sign in expires, the endpoint rate limits you, and for anyone not running a helper the 3.4% problem is exactly what it always was.

So when the data goes dark the tanks stop estimating. They become outlines holding the last real reading as a floor, labelled UNREAD, and the panel tells you why. When work is happening that it can see but cannot quantify, it says UNMEASURED and counts the turns instead. Before there is enough movement to compute an honest pace the speedometer reads MEASURING PACE, because I would rather tell you I am still measuring than hand you a zero you might believe.