burnd
May 13, 2026 · View on GitHub
Built by Garvit Surana — Class XII ISC, Guwahati, India.
I spent $14,502 on Claude Code in one month and had no idea where it went. Built this to find out.
📖 Read the autopsy → 8-minute breakdown on dev.to of the 8 patterns ranked by what they actually cost me, with detector source.
Burnd reads your local ~/.claude/projects/*.jsonl session files and finds the 8 patterns that waste your Claude Code money — with dollar estimates and concrete one-line fixes.
Free. Open source. Nothing leaves your machine.
Quickstart
npx getburnd
That's it. Output looks like this:
burnd v0.4.1 — scanning ~/.claude/projects/
TOP 3 LEAKS THIS WEEK
1 Long Bash output 23 sessions est. \$31.40 wasted
fix: pipe through head/tail to cap output at 500 lines
2 Late-night sessions 11 sessions est. \$18.20 wasted
fix: your 01:00-04:00 sessions cost 2.4x more per task
3 Repeated file reads 8 sessions est. \$9.60 wasted
fix: one file read 19 times in a single session — use Edit not Read
total estimated waste: \$59.20 → run 'npx getburnd serve' for full dashboard
Full dashboard (8 detectors, spend chart, per-project breakdown):
npx getburnd serve
# open http://localhost:4711
What it detects
| Detector | What it catches | Typical monthly saving |
|---|---|---|
| Long Bash output | Test runners / builds dumping 10k+ bytes into context | $20–40/mo |
| Repeated reads | Same file read 3+ times in one session (worst case: 31×) | $10–20/mo |
| Tool error storms | Agent thrashing on broken environments | $30–50/mo |
| Tool overuse | Bash dominating 70%+ of calls when Read/Grep would be cheaper | $15–25/mo |
| Late-night sessions | 00:00–05:00 sessions cost 2.5× more per task completed | $50–200/mo |
| API retry storms | Hidden in system records, invisible from the billing UI | $5–15/mo |
| Skill over-firing | Skills with broad triggers eating 40%+ of tool calls | $10–30/mo |
| Project outliers | Projects costing 3× your session median | $20–40/mo |
In my own data: $76/month in flagged waste. Fixed in a weekend.
How it works
Claude Code writes a .jsonl file for every session to ~/.claude/projects/. Each line is a structured event — tool calls, responses, token counts. Burnd parses these files locally, runs the 8 detectors, and estimates dollar waste from token consumption patterns.
It never connects to the internet. No telemetry, no account, no API key.
Privacy
- Runs 100% locally — no server, no uploads, no accounts
- Never reads your code, only the session metadata Claude Code writes itself
- Parser source is in
src/cli/src/— audit it yourself - MIT license — fork it, self-host it, do whatever
About
Built by Garvit Surana — Class XII ISC, Guwahati, India. After my Anthropic bill hit $14,502 in one month with no good tooling to understand why, I built this. Portfolio: garvit-surana.vercel.app.
- Landing page: getburnd.vercel.app
- npm: npmjs.com/package/getburnd
- Bugs / feature requests: open an issue
- Email: garvitsurana10@gmail.com
Contributing
Each detector is ~50 lines of TypeScript implementing the Detector interface. src/cli/src/detectors/long-bash-output.ts is the simplest one — copy it to add your own. PRs welcome.
License
MIT