Your LLM bill
is leaking.
Most teams run every call on their most expensive model. Frugon reads your real logs locally and shows which calls a cheaper model handles just as well, and which to leave alone. No upload. No telemetry.
$uvx frugon analyze ./logs.jsonl
Reads OpenAI / Anthropic-style request logs (JSONL).
No logs yet? frugon capture · Prefer it installed? pipx install frugon
Not a sales pitch. A measurement.
Share the receipt: frugon analyze logs.jsonl --report savings.html — hand it to your team or your boss.
Three small things, done correctly.
Frugon assembles existing tools. It does not invent. The boring parts are where credibility lives.
Counts tokens with tokencost. Prices with LiteLLM.
Nothing original here. That's the point. The LiteLLM pricing registry updates weekly so the math doesn't rot.
Quality anchored to RouteLLM / LMSYS bands.
We will not round up. If --measure contradicts the projection, the report says so plainly.
pipx, uvx, or pip.
macOS, Linux, Windows. Python 3.10 → 3.13. No accounts, ever. Nothing else to install.
Your data never leaves your machine.
Frugon collects nothing.
Token counting, pricing, and arithmetic all run on your machine. With --measure, your sampled prompts go to your providers using your keys. Nothing else. There is no Frugon endpoint in the path of your data.
- analyze modezero network calls
- telemetrynone
- accountsnever
- measure modeonly to your providers, never to us
Pay premium only when
premium is required.
Frugon is the diagnosis. Continuous routing is the treatment. We'll email you when it ships.
No spam. One email when it ships · unsubscribe in one click