What we have on you
Claude Code writes every session as JSONL into ~/.claude/projects.
The tracker reads exactly those files. Nothing else.
Two minutes and you will know. Get the repository, open the HTML, pick the folder. No account, no server, no telemetry.
The climb
On this leaderboard, first place is no compliment. Rank means damage. 25 tiers from pathetic to cosmic. Enter your monthly Claude spend and find out where you stand.
- The ladder is loading. It also works without JavaScript, as a plain list in the repository.
There are also eight leagues (Damp Twig, Kindling, Campfire, Bonfire, Wildfire, Inferno, Supernova, Heat Death) and achievements such as “Chronically Online Arsonist”, “Single-Day Extinction Event” or, if you take two weeks off, “Successfully Touched Grass (We Are Not Proud of You)”.
A title is worthless if nobody sees it. Sign up with a pseudonym, put the badge in your README and drag your colleagues in with a team code.
How it adds up
No black box. The whole chain sits in the source, here and in the dashboard. The leaderboard server recomputes the same thing independently, so nobody can cheat.
These are estimates
Data centre efficiency, batching, model size and cooling are not in the model. Anthropic sources renewable energy in part, which lowers the real footprint. Treat the number as an order of magnitude, not a measurement.
Why anchor on cost?
Because list price already scales with model tier. A flat rate per million tokens would count an Opus call the same as a Haiku call, and that would be considerably more wrong.
Local and server agree
Dashboard and leaderboard use the same constants. What your machine shows you is exactly what appears in the public ranking.
What leaves your machine
By default: nothing. And if something does, exactly this. Nothing more.
Sent, only after opt-in
Total cost in USD · GPU hours · number of active days · totals per day, week, month and year · optionally a self-chosen pseudonym and a team code.
Never sent
Prompts · responses · source code · file paths · project names · model names · token counts · session IDs · git branches · tool calls · your IP address in the clear.
Revocable at any time
Every entry gets a delete token. One click on “Remove me from leaderboard” wipes the entry and all its time series server side, with no questions and no waiting period.
All of it is checkable. The full source is open: dashboard, leaderboard server and the calculation chain. MIT licensed, including for use inside your company.
Frequent questions
Are my prompts or source code uploaded anywhere?
No. All analysis runs on your machine. Only when you actively click the leaderboard button and accept the confirmation dialog are aggregate numbers transmitted: total cost, GPU hours, active days and per-period totals. Prompts, file paths, project names, model names, token counts and sessions never leave your machine.
How are GPU hours and CO2 calculated?
Token counts from the local logs are priced with each model's list price to get API cost. That cost divided by 3.50 US dollars gives GPU hours, the on-demand price of an Nvidia H100 on AWS p5. One GPU hour equals 0.7 kWh, one kilowatt hour equals 0.4 kg of CO2, and one tree absorbs about 22 kg of CO2 per year.
Do I need Python or an installation?
No. In browser mode you open claude-tracker.html and use the folder picker to select ~/.claude/projects. Python mode is optional and only needed for the persistent archive, OpenAI Codex data and automatic opening. No external Python packages are required.
Why does Claude Code delete my old logs?
Claude Code cleans up JSONL files after roughly 30 days. Python mode therefore keeps a persistent archive that only fills gaps and never overwrites existing sessions, so your history survives even when the original files are gone.
What is the villain ladder?
A 25-tier title ladder based on estimated burned trees, from “Trying to Chop a Tree with a Dull Knife” below 0.01 trees to “Omnicidal Overlord of the Carbon Cycle, Sponsored by Nvidia” from 10,000 trees upward. Plus eight leagues and achievements for streaks, peak days and inactivity.
Can I use this at work?
Yes. The project is MIT licensed, so commercial and internal use are explicitly allowed. For teams there is a freely chosen team code that lets colleagues join a shared standing.
Does it work with OpenAI Codex?
Yes. Python mode also reads ~/.codex and normalises that usage into the same schema, so Claude Code and Codex sit side by side in one dashboard.
Become the very best. At your own expense.
Clone the repository, open claude-tracker.html, pick the folder. No account,
no server, no telemetry. After that you know your title, and you cannot un-know it.