agentwaste

Your AI coding agent is wasting tokens. Find out where, and how to stop it.

Documentation · Quickstart · How it works · Contributing

Try it in 10 seconds

uvx --from git+https://github.com/nikghodki/agentwaste agentwaste

That's it: no install, no account, no config. It reads the Claude Code logs already on your machine and prints a report. No history yet? Add --demo.

Want the agentwaste command permanently? Install it once:

uv tool install git+https://github.com/nikghodki/agentwaste      # or: pipx install git+https://github.com/nikghodki/agentwaste

The examples below use the installed command. Without installing, put uvx --from git+https://github.com/nikghodki/agentwaste in front of them.

Why

Usage trackers tell you how much you used. agentwaste tells you why, and how much of it was avoidable.

On one developer's month of Claude Code use, about a third of all input cost was re-caching the conversation after stepping away for more than five minutes. Nobody noticed, because nothing showed it.

Whether you pay per token or hit subscription limits, that's waste you can fix in minutes once you can see it.

What it finds

Every finding comes with its cost, the worst offenders and a specific fix. Exact definitions →

Share your grade

agentwaste --card my-card.html

A one-file card with your grade and top waste categories, ready to screenshot and post. It contains aggregates only: no paths, project names, commands or code. A test enforces that on every change.

More options

agentwaste --days 7               # last week only (default 30; 0 = everything)
agentwaste --usd-per-mtok 3       # dollar estimates at $3 per million input tokens
agentwaste --json > waste.json    # machine-readable, for scripts and dashboards
agentwaste --demo                 # try it on synthetic sessions

Full reference: CLI · JSON output

Private by design

• Local and read-only. No network requests, nothing uploaded, your logs are never changed.
• No dependencies. Plain Python 3.10+, so there's nothing hidden to audit.
• Honest numbers. Costs are in token-equivalents that weight cache writes, cache reads and fresh input the way they're actually billed. Self-hosted models are reported separately, since their cost is GPU time, not tokens.

Roadmap

• Claude Code support, six detectors, report, JSON, shareable card
• Codex support
• pip install agentwaste / uvx agentwaste from PyPI
• Live warnings while you work ("your cache expires in 1 minute")
• Before/after comparison: did the fix actually help?

Ideas and votes welcome in issues.

Contributing

Contributions are very welcome, especially new log sources (Codex, Gemini CLI, OpenCode) and new detectors. Start with CONTRIBUTING.md:

git clone https://github.com/nikghodki/agentwaste && cd agentwaste
python3 -m venv .venv && .venv/bin/pip install -e ".[dev]" && .venv/bin/pytest

The original AgentSight SDK lives on in legacy/agentsight.

License

MIT