Hey Product Hunt, I'm Ali.

I opened DevTools on my Claude Code usage page, double-clicked the total, typed 401,833,011,420 and took a screenshot. It looked exactly like a real one. That's the problem with every "X billion tokens this month" post: nobody can tell the real ones from the fakes, so nobody believes any of them.

GitHub never had this problem with contribution graphs, because GitHub counts the commits and you never touch the number. tokenproof does the same for agent work.

How it works:

It runs on your Mac as a small local relay between Claude Code or Codex and the provider. Your Anthropic or ChatGPT sign-in keeps working.
It only reads the usage block off each response. Prompts and code pass through in memory and are never written down.
Every count is signed by a key in the Secure Enclave, and Apple App Attest vouches that the key belongs to the unmodified app on a real Mac.
The server re-checks every signature before your card moves. Only daily totals per model ever leave your machine.

You get one URL with a card, badge and heatmap for your README, portfolio or LinkedIn. Anyone who clicks can check the signatures without trusting you.

Two things to be upfront about:

Usage from before you install shows up as self-reported history, kept apart from the verified number, and every widget shows the verified share. My own card says "2% verified" right now because most of my usage happened before the app existed.
The source isn't public yet, so for now you're trusting that the app does what I say. What you can check without trusting me: watch it in Little Snitch or LuLu (it only talks to Anthropic or OpenAI, GitHub to bind your key, Apple to attest it, and tokenproof.pages.dev to publish and check for updates), and open the local ledger and summary.json, which hold models, counts and ids, never any prompt text.

Mac only for now. Free for 7 days, then $6.99 once.

What I'd love from you: try to break it. If you can see a way to get a fake number through this chain, tell me here.