Claude Code token usage history in one short read
The built-in view stops at a week and transcripts fade after 30 days. A daily token log fixes both.
By ClaudeGPT Si Newsroom · Editorial team
· 2 min read
Your Claude Code token usage history lives in two places: a short one in Claude Code, and a long one only if you keep it yourself. The quickest way to keep it is UsageAtlas, a free desktop app that logs tokens per day, per model and per tool, so last Tuesday and last month are always one click away.
The built-in view is short
Type /usage in Claude Code. The Session block lists the tokens your current session used. On a Pro, Max, Team or Enterprise plan, a breakdown below it shows what counted against your limits, and the d and w keys switch it between the last 24 hours and the last 7 days.
That is the whole built-in window. Anthropic's docs add that the figures are approximate and drawn from this machine's session files, so tokens spent on another computer never appear.
Why token history fades
Claude Code writes every session to a JSONL transcript in ~/.claude/projects/. When it starts, it clears out transcripts older than cleanupPeriodDays, 30 days by default. Those files are what most token counters read, so a month-old spike simply disappears from them.
A daily token log in three steps
- Install UsageAtlas on Windows, macOS or Linux.
- Keep coding. It writes daily totals in the background and reads the Claude credentials already on your machine.
- Open History and pick a day, a week or a range.
Each view shows tokens, requests, models and an estimated cost. Switch the tool filter to compare Claude Code tokens with Codex or Cursor over the same days.
Reading the history
Three patterns are worth looking for:
- One tall day. Usually a big refactor or a long agent run. Noise, unless it repeats.
- A steady climb. Your normal week is getting heavier, which matters before a plan decision.
- A late-night cluster. Time-of-day insights show when the heavy sessions land.
- A model shift. If one model's share of tokens jumps from one week to the next, check whether a setting or a new workflow changed which model runs by default.
Look at a few weeks before drawing conclusions. A single week can be skewed by one project, while a month of daily totals gives you a fair baseline to compare against.
| Question | Where to look |
|---|---|
| Tokens in this session | /usage Session block |
| Tokens this week | /usage, press w |
| Tokens on a day last month | UsageAtlas History |
| Tokens as a JSON file | ccusage daily --json |
When a CLI fits better
If you want token history as data rather than a dashboard, ccusage is a good open-source option. It prints daily, weekly and monthly tables from the same local files and supports date filters and JSON output. It is limited to the transcripts still on disk, while UsageAtlas keeps its own record and has no export.
Privacy, briefly
UsageAtlas reads usage data, not conversations: prompts and responses are not read, and no account is required. An optional free account saves daily totals to the cloud so your token history survives a new laptop.