See what your Claude usage actually costs. Zero install, data never leaves your machine.
A local-first dashboard for the token cost of everything you run through Claude on your machine: Claude Code (terminal), Claude Code in the Desktop app, Claude Cowork, and your own Agent-SDK apps. It reads the usage logs these tools already write to disk, prices them from published per-model rates, and shows you where the money goes, by day, model, surface, and session.
It is the only usage tracker that breaks out Claude Cowork cost, which lives in a separate sandboxed-VM session tree most tools never read.
Add a dark-mode screenshot at
docs/screenshot.pngand uncomment the line below before publishing.
Fastest, if you have Node:
npx @htrap94/claudemeterIt scans your local Claude folders, prices everything, and opens the dashboard. Nothing is installed permanently, nothing leaves your machine.
Option A. Open it in your browser. Download index.html, double-click it, and click “Choose your Claude folder.” Everything is computed in the page. No Node, no Python, no server, nothing uploaded. (Or host index.html on any static site / GitHub Pages and visit the URL.)
When the picker opens, select the folder with your Claude data:
| Pick this folder | You get | Platform |
|---|---|---|
~/.claude |
Claude Code (terminal), Claude Code (Desktop), Agent-SDK apps | all |
~/Library/Application Support/Claude |
adds Claude Cowork usage + session titles | macOS |
Add both to capture everything, including Cowork.
macOS note: both of those start with a dot or live under a hidden
~/Library, so the file dialog hides them. Press ⌘⇧. (Command-Shift-Period) in the picker to reveal hidden folders before selecting.
Option B. Instant, refreshable local dashboard (needs Python 3.9+, no packages).
python3 etl.py # scans the default locations, writes data.json
python3 -m http.server 8815 # then open http://localhost:8815index.html auto-loads a sibling data.json when served over HTTP, so this gives you a fast dashboard you can refresh anytime by re-running etl.py. Python 3.9+ ships with macOS; there are no dependencies to install.
- Total cost, tokens, API calls, sessions, and days at a glance.
- Claude Cowork spotlight shows a surface almost no tracker surfaces, isolated on its own.
- Cost per day, stacked by surface or by model, with hover detail.
- Insights cover the biggest cost driver, where spend concentrates (session power-law), cache-read efficiency, output verbosity by model, and week-over-week trend.
- Model and surface breakdowns.
- Top sessions by cost, sortable, so you can find the thread that ran up the bill.
It reads two transcript trees that Claude tools write locally:
| Surface | Source on disk | Entrypoint tag |
|---|---|---|
| Claude Code (terminal) | ~/.claude/projects/**/*.jsonl |
cli |
| Claude Code (Desktop) | ~/.claude/projects/**/*.jsonl |
claude-desktop |
| Agent-SDK apps | ~/.claude/projects/**/*.jsonl |
sdk-cli |
| Claude Cowork | ~/Library/Application Support/Claude/local-agent-mode-sessions/**/.claude/projects/**/*.jsonl |
local-agent |
It cannot see plain claude.ai web chat. That runs server-side and is never logged locally, so there is no on-disk token record to read.
Costs are estimates from published per-token API pricing applied to the token counts in your logs, a useful proxy for where your usage goes, not a copy of your actual subscription invoice.
Runs entirely on your machine. The browser build reads your files through a local folder picker and analyses them in the page: no network requests, no accounts, no telemetry, nothing uploaded. Your generated data.json is git-ignored and never touches this repo. The whole tool is one HTML file and one Python script you can read end to end.
Per assistant turn: input × rate + output × rate + cache-write × (1.25× input for 5-min TTL, 2× for 1-hour) + cache-read × 0.1× input, at each model's published $/Mtok. Claude writes one JSONL line per content block of a turn, all repeating the same usage object, so turns are de-duplicated by request id before pricing (skipping this is the #1 way these numbers get inflated 2–3×). Days are bucketed in your system's local timezone (the CLI reads it from your system and the browser uses your local zone; override the CLI with --tz).
python3 etl.py --tz America/New_York # bucket days in another timezone
python3 etl.py --out /tmp/usage.json # write elsewhere
python3 etl.py --root /some/dir # scan a custom root (repeatable)You can also Export data.json from the browser build to save the same artifact the CLI produces on that machine (both bucket days in your local timezone).
The pricing and aggregation logic lives once in etl.js (used by the browser) and is mirrored in etl.py (the CLI). A parity harness proves they produce byte-identical numbers so the two can never drift:
npm test # generates fixtures, runs etl.py and etl.js over them, diffs every field
node build.mjs # re-inline etl.js into index.html after editing itnpm test also accepts CLAUDEMETER_REAL_ROOT=/path/to/sample to cross-check against a copy of real transcripts. Any change to pricing or logic must land in both etl.py and etl.js or the harness fails. After editing etl.js, run node build.mjs to refresh the copy inlined into index.html.
MIT. See LICENSE.