Quick answer
Claude has no single page that lists every window you have used. The usage panel in Claude settings shows the current 5-hour session and the weekly window. Per-day token history lives in Claude Code session files under `~/.claude/projects`, and a tracker such as UsageAtlas turns those files into 7, 30, and 90-day views.
Three places history hides
The usage panel answers one question well: how much capacity is left right now, and when does it come back. That is the question you ask before starting a long agent run. It is not the question you ask on Friday when you want to know where the week went.
| View | What it covers | How far back |
|---|---|---|
| Usage panel in settings | Current session and weekly windows | Only the open windows |
| Claude Code session files | Tokens per request, model, cache fields | Until you delete them |
| Tracker or dashboard | Daily and per-model rollups | As long as you keep the data |
Only the second row is a real record. The panel is state, not history, and it forgets a window as soon as the window closes. If you want to compare this week to last week, you need something that persists.
Read the session files
Claude Code writes one JSONL file per session under a folder per project. Each assistant message carries usage: input tokens, output tokens, cache creation tokens, cache read tokens, and the model name. Those five fields are the whole history, and everything else is arithmetic on top of them.
- Windows:
%USERPROFILE%\.claude\projects - macOS and Linux:
~/.claude/projects - Each subfolder name encodes the project path, so sessions stay grouped by repository
- Subagent calls appear as their own records, not inside the parent message
You can grep the folder, but summing by hand goes wrong in two predictable ways: cache read is already part of input, and subagent records are easy to skip. Claude Code token usage defines each field. UsageAtlas does the summing and keeps the history without an account.
Build a history you can compare
A history is only useful if two weeks are measured the same way. That sounds obvious and is routinely broken by deleting old sessions, switching machines, or changing how you count cache tokens halfway through a month.
- Keep the session folder. Deleting it deletes your only local record.
- Open the tracker weekly so no day falls outside the rollup.
- Note the weeks you changed model or plan, so a jump has a cause attached.
- Export a CSV at month end if you need to show spend to someone else.
One more habit helps: keep the definition of a token constant. Cache reads are cheap, cache writes are not, and a week that looks expensive may simply be a week with more cache misses. Read cache read tokens before you decide a model got pricier.
Questions, answered
Does Claude keep a history of my usage?
The product keeps current window state rather than a long archive. Per-request history exists in Claude Code session files on your machine, and in the invoice if you were billed through the API.
How far back can I see Claude usage?
A tracker reading local files can show as much as you have kept, commonly 7, 30, and 90 days. Anything older depends on whether you archived the JSONL files before deleting them.
Why does my history start mid-month?
Because it begins when you installed Claude Code or when the tracker first read the folder. Usage before that date exists only in invoices, not on disk.