multi-turn inc.
INDEX

index

Token Burn

AI coding agents burn tens of thousands of tokens a day. Most users find out only when they hit the limit. Token Burn puts that consumption on the menu bar.

Invisible consumption

A full day with Claude Code burns roughly 50,000 to 200,000 tokens. Add Codex and it goes higher. Most users never see this number. They find out when the limit hits.

"You have reached the 5-hour limit."

By then it is too late. The workflow is broken. You wait for the reset.

/cost in the CLI shows the current session, but not the full day. Multiple terminals, multiple projects, multiple agents are all burning tokens at once, and there is no way to see the total.

So I built Token Burn.

It lives in the macOS menu bar.

Token Burn menu bar icon

One icon, two readings:

  • Horizontal fill: remaining 5-hour quota
  • Vertical height: remaining 7-day quota

When quota is plentiful, the bars are full. As they drain, they shrink. Half a second is enough.

When the AI is actively calling, the bar pulses. Heartbeat.

Panel

Token Burn main panel

Top half: per-service remaining quota gauges. Bottom half: token burn chart. Switch between 1-hour, 24-hour, and 7-day scopes with a trackpad scroll.

Data is parsed directly from Claude Code's JSONL (~/.claude/projects/) and Codex's SQLite (~/.codex/logs_2.sqlite). Nothing goes to a server.

Remaining, not used

Most existing tools show "how much did I use?" — /cost, ccusage, TokenTracker. Token Burn shows "how much is left?" first.

"53% used" versus "47% remaining." The second one changes behavior.

Things encountered along the way

Claude Code stores OAuth tokens in the macOS Keychain, but deletes and recreates the entry on every refresh. The ACL resets each time, so Token Burn was asking for the Keychain password every 8 hours. Workaround: read from file (~/.claude/.credentials.json) first.

On macOS 26, an NSStatusItem created from a binary inside a .app bundle would not appear in the menu bar. Same binary outside the bundle worked fine. Workaround: LaunchAgent. Not elegant, but it works.

Token counts are based on local logs, so they do not match Claude Code /stats exactly. This is shown in the UI.

What comes next

The next thing I want to try is per-project token analysis. If you could see "this refactoring cost 120,000 tokens," the way you work with AI would change.


Open source. Available on macOS.

GitHub: multi-turn-inc/ai-usage-meter

Token Burn | Multi-turn Inc.