quota-autopsy reads your local session transcripts and shows you, call by call: where the tokens went, which calls wasted them, and the config changes that stop it. Local-only. Zero dependencies. Free overview.
$ npx quota-autopsy Range 2026-09-07 → 2026-09-12 · 46 sessions · 1162 API calls TOTAL 276.58M tokens ≈ $187.80 ├─ input 4.20M $21.01 fresh tokens (full price) ├─ output 1.25M $31.23 generated tokens (full price) ├─ cache write 0 $0.00 1.25x–2x price └─ cache read 271.13M $135.56 0.1x price · saved ≈ $1220.07 Waste signals (all sessions) ⚠ 4 warm contexts re-sent uncached ≈ $3.39 ⚠ repeated file reads ~477.9k tokens ≈ $2.39 ⚠ oversized tool outputs ~937.4k tokens ≈ $4.69 Estimated avoidable spend: ≈ $10.47 → details + fixes per session with Pro
Real run: 46 sessions, 1,162 API calls (author's own data) — tool output stays in English
Anthropic publicly acknowledged users hitting Claude Code limits "way faster than expected" (Mar 2026). The community found the mechanisms: cache prefixes corrupted by context edits, TTL downgrades from 1h to 5m past the quota boundary, resumed sessions re-billing whole contexts uncached.
"One of our agents busted cache and did not notice. By the time we noticed, we went from 0% to 70% in three or four turns." — from the public issue threads
Counters tell you how much you burned. quota-autopsy tells you exactly where, why, and what to change — from the same transcripts your local Claude Code already writes.
Five patterns that cost real quota. quota-autopsy detects every one of them automatically.
Every burned call sits in your transcripts. quota-autopsy finds the exact turn — timestamp, tokens, dollars.
Cold starts get flagged: 610k tokens re-sent after 15h idle = $3.08. Batch your sessions instead.
Finds every 40k+ char dump — and the context tokens it permanently costs.
Dedupe by message.id — 3,600 duplicated rows collapsed. Numbers you can trust.
Runs 100% local. The only network call in the entire tool is the one-time license check.
The cached prefix rewritten while still fresh — full context paid at 1.25–2x instead of 0.1x. Shows the exact turn, gap, and dollars.
Calls that read zero from cache seconds after a full-context call. Two back-to-back full-context bills is the most expensive pattern in agent work.
Watches cache writes shift 1h → 5m mid-session — the documented post-quota penalty that compounds future burn.
Files read 3+ times, 600KB terminal dumps and screenshots that re-inflate your context on every later turn.
Throttled calls that still billed, 150k+ contexts, sidechain spend — each with a concrete fix.
Continued/forked sessions copy whole histories. Same API call = same charge — 3,600 duplicate rows collapsed in our own dogfooding corpus.
All you did today was refactor. The numbers don't add up — and you can feel it.
One finding: 610.5k tokens re-sent after a 15h idle gap. Right there. Visible.
From tomorrow on, that $3.08 no longer exists. Sleep better.
The free overview shows you the size of the problem. Pro shows you exactly where it is:
What a Pro report looks like (real output)
$ quota-autopsy show 50f4d9 (Pro) 2026-09-12 · 151.31M tokens · ≈ $91.91 · cache hit 99.0% [MED] 1 oversized tool output ≈ $0.36 ~72.8k tokens • Read …/Screenshot 2026-09-09.png — 269k chars fix: trim at the source — pipe through head/tail/grep, paginate listings, ask for a summary instead of raw dumps [INFO] 2 cold starts after idle (normal) • 09-11 17:52 re-sent 610.5k fresh tokens after 15h 31m • 09-11 01:31 re-sent 552.0k fresh tokens after 26h 42m Most expensive calls 1. 09-11 17:52 input 610.5k $3.08 2. 09-11 01:31 input 552.0k $2.79
Slide to your situation — the estimate uses the share found in our real 46-session corpus.
Estimate, not a promise — but every percent is real money.
Streams ~/.claude/projects/**/*.jsonl — 46 sessions (99 MB) in 0.42 s. No network, no telemetry, no account.
Per-call token buckets, cache behavior, tool outputs, sidechains — priced with a table you can edit for any provider or router.
Terminal overview, per-session autopsy, Markdown/HTML export for the team, JSON for your own tooling.
Counters tell you how much. Only quota-autopsy tells you where and why — and what to change.
| quota-autopsy | /cost | ccusage | |
|---|---|---|---|
| Call-by-call: where & why | ✓ | – | – |
| Fixes per finding | ✓ | – | – |
| Cache timeline (TTL, re-creates) | ✓ | – | – |
| Report export (Markdown/HTML) | ✓ | – | – |
| Data source | local transcripts | local session | Anthropic API |
| Price | Free · Pro $24 | free | free |
One-time. No subscription. Local license check with 30-day offline grace. macOS · Windows · Linux.
Pays for itself the moment you remove one found waste — and the overview is free anyway.
No. The tool reads local transcript files and writes only its own config under ~/.quota-autopsy/. The single network call it can make is the license check against Gumroad when you activate Pro — nothing else, and no telemetry.
macOS, Windows and Linux — anything with Node 18+. It reads whatever Claude Code transcript directory exists on your OS.
Yes. Costs come from an editable pricing table (~/.quota-autopsy/pricing.json, see examples/). Detectors work on any model name.
Token counts are exact (deduplicated by API response id — streaming partials and session-split copies counted once). Dollar figures are estimates from a published price table, shown as “≈”, and cross-checked against Claude Code's own cost-state rows when present. An independent second implementation in tools/verify.py recomputes every metric in the test suite.
Yes. Activation doesn't consume uses — install and activate everywhere you work. The check re-verifies locally with a 30-day offline grace.
Counters answer “how much?”. This answers “where did it go, why, and what do I change?” — per-call forensics with fixes, not a counter.
Report it — a mislabeled pattern is a bug. Refunds are handled through Gumroad, but most issues are fixed in the next release, and updates are free with your license.
One email when a new detector ships. No spam, unsubscribe any time.
Collected via Netlify Forms — no third-party tracking.