For Claude Code & Codex
Reduce Claude Code token usage by up to 60%.
A tiny CLI that strips logs, diffs, and test noise out of your agent's context — before the model reads it.
npx tokenpincher initthat's the whole setup
your prompts: untouched · originals kept on disk · savings are estimates
What it does
Same information. A fraction of the tokens.
PASS src/engine/compress.test.ts (4.1s)
FAIL src/lib/license.test.ts
✕ rejects expired subscription (211ms)
Expected: 403 Received: 200
at Object.<anonymous> (license.test.ts:88)
… 207 more lines of reporter noise, timing tables, stack tracesClaude reads all of it — again, every turn
213 passed · 1 failed: rejects expired subscription (403 vs 200)
Claude reads only this
14:02:11 INFO req GET /api/license/validate 200 12ms 14:02:11 INFO req GET /api/license/validate 200 9ms 14:02:12 INFO req GET /api/license/validate 200 11ms 14:02:14 WARN stripe retry (attempt 2) — timeout after 5000ms 14:02:15 ERROR license signature mismatch key=TP-9f… … 195 more lines, mostly repeats of the same requests
Claude reads all of it — again, every turn
182 INFO · 14 WARN · 4 ERROR — errors kept in full, repeats folded
Claude reads only this
The full output stays on disk — if the model wants the original, it pulls it back on demand.
Your prompts are ~4% of your token bill.
Setup, once
- 1
npx tokenpincher init - 2
every tool call is compressed, automatically
- 3
tokenpincher statsshows the receipts
Pricing
7 days, everything included. Then pauses until Pro.