History
How many tokens your AI coding agents burn over time. Open it from the History pill of the TokenEater window.
It reads local logs only, nothing is uploaded, and it works without a usage login:
- Claude Code:
~/.claude/projects/**/*.jsonl - Codex:
sessions/andarchived_sessions/under~/.codex(orCODEX_HOME), while Codex tracking is on
What you see
- Ranges: 24h (hourly), 7d, 30d, 90d (daily). 7d is seven calendar days including today, like the Tokens History widget.
- Hero: active tokens, cached tokens and cache hit rate, with the delta versus the previous period.
- Tabs: Tokens by model, Projects, Sessions, Cache.
- Filters: Claude families (Fable, Opus, Sonnet, Haiku, Other) and each Codex model found in your logs.
How tokens are counted
- Active = uncached input + cache writes + output.
- Cached = cache reads, which set the cache hit rate.
- Codex reasoning is part of output. Codex Guardian and auto-review sessions are left out.
Changed in 6.0: cache writes now count as active, not cached. The same logs show a higher active total and a lower cache hit rate than in 5.x. Your usage did not change, the accounting did.
In a provider mode
The All / Claude / Codex switch also scopes History, but only token counts split by provider. Read projects, sessions and cache in All mode. See Claude and Codex.