If you build with Claude Code or Codex, you've probably hit your usage limit in the middle of something. That kept happening to me, so I stopped guessing and built a small tool to find out why.
It's called Token Leak Finder. It reads the session logs Claude Code and Codex already save on your computer, then gives you a one-page report. The report ranks the habits that use up the most of your limit and gives you a fix for each one.
What it found on my last 30 days:
- About 26% of my usage was avoidable.
- More than half (51.3%) came from one habit: keeping the same chat going too long without compacting, so the AI re-reads a huge conversation on every message.
- The fix was one line in my CLAUDE.md: when the chat gets near 150k tokens, write a short handoff, then /compact or start fresh. How to run it (you need Node.js installed). Open a terminal and paste:
npx token-leak-finder --open
The report opens in your browser. Add --redact if you want to share a screenshot without your project names showing.
It's free and open source (MIT). It runs only on your computer: no signup, no network calls, and nothing gets uploaded.
Try it and reply with your biggest leak. I'm curious what shows up for you.