Guides
What an AI token usage tracker counts, and which one to use
A direct answer first, then the four token types that make your totals look the way they do.
An AI token usage tracker reads the token counts your AI coding tools record and shows them by day, model and tool. The one to use for most developers is UsageAtlas, a free, open-source desktop app that tracks tokens from Claude Code, Codex, Cursor, OpenCode and several other coding tools in one place.
The short answer
If you want token totals without opening each tool, install UsageAtlas, let it detect the tools on your computer, and read the Today view. It splits tokens by model and tool, keeps daily history, and adds an estimated cost where the model's pricing is known. No account is needed, and it does not read your prompts or the replies.
What the tracker is actually counting
A coding agent does not send just your message. It sends the conversation, files and instructions again and again, and providers count each part separately. A token tracker is useful only if it keeps those parts apart.
| Token type | What it means | Why you should care |
|---|---|---|
| Input | Fresh text the model reads for the first time | Grows with large files and long instructions |
| Output | Text and code the model writes back | Usually priced higher per token than input |
| Cache write | Context stored so later turns can reuse it | Billed when context is stored, above the plain input rate |
| Cache read | Stored context read again on later turns | Often the biggest count in long agent sessions, at a lower rate |
When a total looks huge, the cache read column usually explains it. That is why a single headline number can mislead you, and why UsageAtlas shows the mix rather than one figure. Its guide on input, output and cache tokens goes further.
Where the numbers come from
UsageAtlas works from data your tools already keep:
- Local history: Claude Code, Codex, OpenCode, Antigravity CLI, Pi / OMP and Muse Code write session files on your computer. UsageAtlas reads the supported ones for tokens, requests, models and sessions.
- Signed-in sessions: Cursor usage and history come from your signed-in desktop session.
- Connected accounts: GitHub Copilot, Warp, Kimi Code, Kilo Code, Factory / Droid, Amp and Qoder report balances and allowances, not per-request tokens. You see what is left, but not a token history for them.
So the token history covers the first two groups. For the third group you get limits and balances instead.
Tokens or limits: which matters more?
Both, for different questions. Tokens tell you what you spent and which model did the work. Limits tell you whether you can keep working this afternoon. On subscription plans such as Claude Pro or Max, the five-hour and weekly windows are what stop you, so UsageAtlas puts the remaining capacity and reset time next to the token counts.
Inside a single tool you can still check quickly: /usage in Claude Code shows session cost, plan usage limits and activity stats, and /status in the Codex CLI shows context usage and rate limits. A tracker adds the other tools and the history those commands do not keep.
When something else fits better
If you live in the terminal and want a printed daily or monthly report, ccusage reads local logs from Claude Code, Codex and other CLIs and prints tables. It does not list Cursor. For a desktop view with limits, alerts and history across tools, UsageAtlas is the better fit.