tokenwise

The efficiency layer for AI coding spend. Stop flexing how much you burned — flex how much you shipped per dollar.

Every AI-coding tracker ranks you by tokens torched, so the "winner" is whoever wasted the most. Tokenwise flips it: one honest score for value shipped per dollar, across Claude Code and Codex.

npx tokenwise-cli

Published as tokenwise-cli on npm; the installed binary is tokenwise (npm i -g tokenwise-cli, then tokenwise).

  tokenwise · efficiency, not waste

  EFFICIENCY SCORE  (local estimate · last 30 days)
  92/100  ████████████████████████

  Cost / commit    $0.41
  Cache hit rate   73%
  Spent            $128.40  (1,204 turns)
  Shipped          312 commits  +18420/-6310 lines

Commands

Command	What it does
npx tokenwise-cli	Show your spend + efficiency score (default report)
tokenwise score	Print just your efficiency score
tokenwise share	Write a shareable card to ./tokenwise-card.svg
tokenwise submit --handle @you	Preview the exact anonymized aggregate, then post it to the leaderboard

Options: --days <n> (default 30), --all, --all-projects, --handle <@you>, --json, --dry-run.

Scope: by default the score is for the current repo — usage is scoped (by each session's working directory) to the repo you run it in, so "cost per commit" lines up with that repo's commits. Add --all-projects for your global spend across everything.

submit posts your anonymized aggregate straight to the leaderboard (Supabase PostgREST, anon key bundled). The official score is computed in the database, so nobody can fake it. Use --dry-run to preview without uploading. Self-host by overriding TOKENWISE_SUPABASE_URL / TOKENWISE_SUPABASE_ANON_KEY.

Privacy — this is the whole point

This CLI is open source (MIT) precisely because it touches your most sensitive data: prompts, code, file paths. So you can read it and verify the promise:

Only anonymized numeric aggregates ever leave your machine. Never prompts, never code, never file paths or branch names.

Run tokenwise submit to print the exact JSON that would be uploaded before anything is sent. Project paths appear only as a one-way SHA-256 hash. Cost is computed locally. The leaderboard and card are opt-in.

How it works

1. Reads local logs from Claude Code (~/.claude/projects/**/*.jsonl), Codex (~/.codex/sessions/**/*.jsonl) and Copilot CLI (~/.copilot/, experimental).
2. Computes real cost from a bundled pricing table (Claude Code leaves cost out, so we calculate it — including cache-read/creation rates).
3. Pairs spend with your local git activity (commits, lines kept) to estimate efficiency: cost per merged PR, per kept line, cache reuse.
4. The official ranked score on the leaderboard is computed server-side from these aggregates (anti-gaming) — the CLI shows a local estimate.

Sources

Tool	Support	Notes
Claude Code	✅ full	the anchor — local JSONL with token usage
Codex CLI	✅ experimental	token tracking only in recent builds
Copilot CLI	⚠️ experimental	best-effort local-log parsing; format varies
Cursor	❌ not supported	billing is server-side; local data has no token/cost. (Teams: a Cursor Admin-API path is on the roadmap.)

Copilot/Codex sessions don't record a working directory, so they only count under --all-projects (the per-repo score needs a cwd).

MIT © Bluu — part of tokenwise.dev