I used 2.29 billion Claude Code tokens in 30 days, so I built a leaderboard that checks the numbers

TokenTrust is like TrustMRR, but for AI developers: show how much you're using AI and let other people find you.

Every few days someone on X posts a screenshot of their Claude Code usage. Three billion tokens. Thousands of dollars of API-equivalent spend in a month. And the replies are always the same: "is that real?"

Usually nobody can tell. The screenshots come from ccusage or a dashboard, and the leaderboards that collect them rank whatever number you send in. Nothing stops someone from adding a zero.

I wanted the opposite: a public leaderboard where the number is checked before it's ranked. Revenue already has this. TrustMRR shows startup revenue pulled from payment providers like Stripe instead of founder tweets. So I built TokenTrust, the same idea for AI coding agents.

Where the numbers come from

Claude Code and Codex both keep a local log of every session. Each model response in those logs carries the provider's own ids (msg_ and req_ for Anthropic, resp_ for OpenAI), the model name, a timestamp, and exact token counts: input, output, cache reads and cache writes.

TokenTrust's CLI reads those logs and uploads one row per response, with only these fields: agent, model, timestamp, token counts, provider ids, and salted hashes of the session id and git remote. No prompts, no responses, no file names, no paths. Running push --dry-run prints exactly what would be sent.

How it tries to stop fake numbers

Because every response has a provider id, the server can check things a screenshot never could:

• Ids have to match the provider's id format.
• The same response can't be counted twice.
• A response id another account already claimed is rejected.
• Output faster than parallel agents could stream in that window gets flagged.
• A large share of byte-identical token counts gets flagged.
• A first upload that's mostly old history is shown but not ranked.

If you have an Anthropic organization, you can skip logs entirely and connect a read-only Admin API key. Those numbers come straight from Anthropic's usage report.

What it can't do yet

I want to be upfront about this: uploaded logs are checked, not attested by the provider. Someone determined enough could still fabricate well-formed logs with plausible ids. The fraud rules make faking a lot more work; they don't make it impossible. If you have ideas for closing that gap, I'd love to hear them in the comments.

Joining takes about 30 seconds

1. Sign in with GitHub or X at tokentrustt.dev
2. Paste one command (macOS, Linux or Windows)
3. Your usage syncs on its own after each Claude Code or Codex session

Right now there's exactly one person on the board: me, at 2.3 billion tokens and about $850 of API-equivalent usage in the last 30 days. That's the number to beat.

Originally published on Medium.