# Seeing what a Claude Code session actually cost

Claude Code already writes every session to disk. Shiplings reads those files and totals the tokens by project and by model, so you can answer "what did this fortnight cost" without a dashboard or an API key.

*Updated 24 September 2026. Canonical: https://shiplings.app/claude-code-token-usage.html*

## Where the numbers come from

Claude Code keeps a transcript of every session under `~/.claude/projects/`. Those files record which model answered and how many tokens it used. Shiplings reads them. There is no API key to create, no billing endpoint to call, and nothing is sent anywhere.

Folders are grouped into **projects** you name, so two directories of the same codebase count as one thing rather than as two rows that both look half right.

## The model matters more than the total

A million Opus tokens and a million Haiku tokens are not the same story, so the split is kept rather than summed away. Figures are written to be read at a glance, one way everywhere: `4.2k`, `3.4M`, `1.2B`.

## Sessions, which is what makes it a number you can use

A total since install is trivia. A total for a piece of work is a fact. You open a session, name it, drag the projects it covers into it, and close it when the work ships. Closing it produces a card: tokens, agent time, days worked, what shipped, and a breakdown by project.

A session created after the fact can count from an earlier date, so naming it on Friday does not lose the fortnight behind it.

> What shipped is asked of GitHub, Vercel and Netlify for the session's own dates, rather than accumulated from what the app happened to be awake for. A card claiming 12 commits over a fortnight you know was 60 does not lose that line, it loses the whole card.

## And how much of your plan is left

What was spent is half the question. The menu bar menu also shows a gauge for each window of your Claude Code plan: the five-hour window, the week, and the weekly limit for a single model, each with what is left and when it resets. Codex gets the same.

Those figures are not in the transcripts, so Shiplings asks Claude Code itself, the way `/usage` does, under your own sign-in. It never reads that sign-in, sends no message to the model, and the question spends none of your plan. Because the answer comes from your account, it includes what you used on the web or on another machine.

## Questions

**Can it show how much of my Claude Code plan is left?**

Yes, free. The menu bar menu shows the five-hour window, the week and the per-model weekly limit, each with what is left and its reset time.

**Do I need an Anthropic API key?**

No. The figures come from the files Claude Code already writes on your own disk.

**Does it work if I use Claude Code through a subscription rather than the API?**

Yes. The transcripts are the same either way.

**Is it a cost dashboard?**

No. There are no graphs and no permanent counters. The numbers appear when a piece of work ends, on one card, and then get out of the way.