TokenBurn

htop for your LLM spend — proxy-only.

TokenBurn is a local-first HTTP proxy for LLM spend visibility and control.

Route OpenAI-, Anthropic-, and Gemini-compatible traffic through a local proxy. TokenBurn logs usage locally, attributes cost by model/provider/program/tag, and turns raw traffic into actionable waste reports.

No hosted backend. No account. No prompt egress by default.

Install

pip install "tokenburn[proxy]"
tokenburn proxy setup
tokenburn proxy start --background

After setup, clients that support base URL overrides can route through TokenBurn with no app-specific SDK integration.

What it does

• Proxy-based capture — intercepts LLM traffic at the HTTP layer
• Cross-language — works with Python, TypeScript, Go, curl, and anything else that can point at a base URL
• Cross-provider — OpenAI, Anthropic, Gemini
• Local logs — normalized JSONL logs under ~/.tokenburn/logs/
• Spend reports — model, provider, endpoint, program, and tag breakdowns
• Waste detection — highlights expensive patterns worth fixing first
• Shareable output — terminal and exported reports

Core commands

tokenburn proxy setup
tokenburn proxy start --background
tokenburn proxy status
tokenburn proxy stop

tokenburn report
tokenburn gain
tokenburn share --open
tokenburn doctor

Product direction

TokenBurn is proxy-only.

That means the product lives at the proxy boundary rather than inside application runtimes.

The product lives at the network boundary, not inside app SDKs.

Repo docs

• PROXY_TFF.md — proxy technical design
• SPEC.md — product and positioning
• keche.md — project operating brief
• CLAUDE.md — maintainer workflow notes

License

MIT