TokenScale is a single static website with no servers, no framework, and no build pipeline in the usual sense. It's deployed by dragging one zip file onto a host. Here's the whole machine — because the "how" is half the point.
Most of the complexity in shipping software comes from disagreement about what's actually live. TokenScale removes that by making it literal. The exact zip that's on the server is kept in a folder called shipped/. That's the truth — not a branch, not a build number, not someone's memory. The next version-in-waiting sits in staging/. Two folders, no ambiguity.
Build a candidate. Test it. Ship it or reject it. Repeat.
TokenScale's whole job is to be right about prices — what 21 AI providers charge, turned into things you can picture (an email, a novel, a PhD thesis). So every night the numbers are re-verified against each provider's published rates and folded back into the site. A price move on a Tuesday shows up by Wednesday morning.
The code that does this is deliberately simple. What compounds is the record — a continuous, day-by-day history of what AI actually cost. Every verified day is one nobody can go back and collect.
The hardest bugs in any deploy aren't crashes — they're silent omissions: a file quietly dropped, a page that vanishes, a feature that reverts. So every build runs a guard: the new version must contain everything the last one had. If a single file would go missing, the build fails and nothing ships. Lessons learned the hard way, encoded so they can't repeat.
That's it. A countdown of AI prices, kept honest by two folders and a nightly habit. See the prices →