vibescore-cli

Measure what a Claude Code project actually cost to build, then list it on VibeScore.

cd your-project
npx --yes vibescore-cli | pbcopy          # macOS
npx --yes vibescore-cli | Set-Clipboard   # Windows PowerShell
npx --yes vibescore-cli | xclip -selection clipboard   # Linux

Run it from the project you're submitting. Paste the result on the submit page. A human-readable summary is printed to stderr, so you still see what you are about to paste even when stdout is piped to the clipboard.

What it reads

Claude Code already keeps a transcript of every session in ~/.claude/projects. This reads those files and reports on one project: the folder you ran the command from, and anything nested under it — every transcript line carries the cwd it was written from, so that's exact, not a guess. Nothing to pick from a list of every project you've ever built.

Cost comes from ccusage, which maintains the per-model price table; the two are joined on session id. Computing it reads every transcript on the machine, not just this project's — a session that touched several directories has its cost split by each one's share of that session's tokens, which needs the session's whole footprint to get right. If ccusage cannot be run, everything else is still reported and cost comes back as null rather than 0 — an unknown cost and a free build are different claims to put on a listing.

Why not just ccusage

ccusage reports token usage per session, which leaves two gaps:

• It doesn't say which project a session belonged to. Every transcript line carries cwd, so scoping here is exact rather than you picking session ids off a list.
• It only records when a session ended. Transcript lines carry a timestamp each, so a build has a real start and end — including a single-session build, which ccusage alone can only call unmeasurable.

Time to ship

Two numbers, because they answer different questions:

• Active — time you were actually engaged. Gaps after Claude finishes and before you reply are counted only if you came back within 30 minutes; past that you had left. Claude's own turns always count, so a long autonomous run is work, not idle.
• Wall-clock — first activity to last, calendar time, no interpretation.

A real example: 45 hours of wall-clock on this project, 10 hours of active time. Both are true; they just measure different things.

Privacy

Nothing is uploaded. The command writes to stdout and you choose what to do with it.

Filesystem paths are deliberately not included in the output — only the project's bare folder name. Pasting a report does not disclose your directory layout.

The payload contains: folder label, session ids, message count, first/last activity, wall-clock and active minutes, token counts per model, and cost.

Licence

MIT