Burnd

burnd — find what's burning your Claude Code budget

Built by Garvit Surana — Class XII ISC, Guwahati, India.
Status: feature-complete, no longer actively developed. Everything below works
and stays on npm. The paid tier was removed in 0.1.0 — all 10 detectors and every
command are free. I wrote up honestly why this project didn't work as a business:
read the postmortem.
I spent $14,502 on Claude Code in one month and had no idea where it went. Built this to find out.

Burnd reads your local ~/.claude/projects/*.jsonl session files and finds 10 patterns
that waste your Claude Code money — with dollar estimates and concrete one-line fixes.

Free. Open source. Nothing leaves your machine.

Quickstart

That's it. Output looks like this:

Full dashboard (10 detectors, spend chart, per-project breakdown):

What it detects

Nine per-session detectors plus one that compares across sessions:

burnd estimates dollar waste from token-consumption patterns. Treat the numbers as
directional, not as an invoice.

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 10 detectors, and estimates dollar waste from token consumption.

Two design choices are worth calling out:

Per-user baselines, not absolute thresholds. computeUserBaseline() builds a
histogram of your session-start hours, finds the 10-hour window where you actually work,
and defines "off-focus" relative to that. Cost thresholds are percentiles of your own
sessions. A detector tuned on a $14k/month user still fires correctly for a $20/month user.

Fail-expensive on unknown models. When Anthropic ships a model the pricing table
doesn't know, burnd assumes the most expensive…