Find where your Claude Code tokens are leaking — with $ impact and fixes.

npx github:yurukusa/cc-token-diet

The npm package is not yet published. Use the GitHub-direct form above for now; an npx cc-token-diet short form will be added when published.

Reads your local Claude Code conversation logs (~/.claude/projects/), detects specific waste patterns, and tells you which sessions cost the most and what to change. Nothing leaves your machine.

Other tools tell you how much you spent. cc-token-diet tells you where the waste is and how to stop it:

• Cache write ratio — if you're rewriting the cache too often (>30 %), you're paying 12.5× what a stable context would cost.
• Verbose output — sessions where the model averages >3K output tokens per turn, usually because it's been asked to "explain everything."
• Runaway sessions — long single sessions (>100 assistant turns) that bloat context exponentially. These are where the money quietly disappears.

Each finding comes with a concrete fix and the Token Book chapter that covers it in depth.

📉 cc-token-diet report (last 7 days)
────────────────────────────────────────────────────────
Sessions analysed: 312   Assistant turns: 36973
Input:   457.8K   Output:         7.36M
Cache read:   5.21B   Cache write:  87.15M
API-equivalent spend:  $10007.22  (Opus 4.x pricing)
  ↳ On Max/Pro you pay the flat subscription; this is context intensity
  ↳ On Sonnet divide by ~5. The patterns stay the same either way.
Cache hit ratio: 98.4% (higher is better, aim >85%)

🔥 Waste patterns:
  ❗ 87 runaway session(s) (>100 assistant turns) (≈ $2726.01 wasted)
     Fix:  Long single sessions bloat context exponentially.
           Use /compact or restart for each logical task.
     See:  Token Book ch3 (context management)

🌡️  Hottest sessions:
   $513.26  2026-04-19  projects-cc-loop         955 turns / 157min
   $499.34  2026-04-16  -home-namakusa           898 turns / 3898min
   $431.08  2026-04-20  projects-cc-loop        1103 turns / 180min

This is real output from the author's own logs. A 7-day window surfaced 87 runaway sessions worth ~$2,700 of context intensity that could have been trimmed by splitting into smaller sessions. Your numbers will differ — but the shape is how you know what to change.

# default — last 7 days
npx github:yurukusa/cc-token-diet

# wider window
npx github:yurukusa/cc-token-diet --days 30

# machine-readable
npx github:yurukusa/cc-token-diet --json > report.json

Runs 100 % locally. No network calls. No analytics. The tool only reads the .jsonl logs Claude Code already writes to ~/.claude/projects/. If you want to verify, the source is one file: cli.mjs.

Cost estimates use public Anthropic Opus 4.x API pricing (input $15/MTok, output $75/MTok, cache read $1.50/MTok, cache create $18.75/MTok). Actual subscription users (Pro/Max) aren't billed per token — treat the number as "what your session would cost on the API" and watch it relative to your plan's headroom.

If you mostly run on Sonnet, divide by ~5. The patterns still surface the same inefficiencies either way.

• cc-cost-check — calculates cost per commit/hour/day from aggregate stats
• cc-cost-forecast — projects API-equivalent spend to month-end
• cc-token-diet — finds specific waste patterns in your sessions and tells you how to fix them

Use them together: forecast your pace, audit your waste, fix the leaks.

The full token-saving playbook (10 chapters, 40+ documented symptoms, real incidents with hook examples):

• Token Book (¥2,500) — https://ko-fi.com/s/fd44ef09a7
• Free first chapter — https://zenn.dev/yurukusa/books/token-savings-guide

cc-token-diet shows you this month's waste. The patterns that cause it keep changing as Claude Code evolves. CC Safety Lab Founder (¥500/mo Ko-fi Membership) delivers a monthly digest: 3–5 curated incidents from GitHub with fixes, 1 new safety hook, 1 measured token-saving technique with actual $ data, and early access to Token Book updates. Issue #1 ships 2026-05-01. Founder price locked for charter members.

MIT