Frugon is a local-first LLM cost analyzer. Its strongest security property is structural: frugon analyze runs entirely on your machine — no network, no telemetry, nothing sent to us. The smallest attack surface is the one that never collects your data.

Please do not open a public issue for security problems. Report privately:

• Preferred: GitHub's "Report a vulnerability" button (this repo → Security tab → private advisory).
• Email: security@rodiun.io

Include the affected version, a description, steps to reproduce, and any suggested fix. You'll get an acknowledgment within 48 hours and a triage response within 7 days. We credit reporters who follow responsible disclosure, unless you'd prefer to stay anonymous.

Frugon is pre-1.0 and ships from main. Security fixes land on the latest PyPI release — please pipx upgrade frugon before reporting, in case it's already fixed.

Frugon handles potentially sensitive inputs, so these matter most:

• Local data handling — any path that could leak your logs / prompts / completions off-machine, or write them somewhere unexpected.
• --measure + API keys — --measure calls your own providers with your own keys; any path that could route a key or prompt anywhere other than the configured provider (e.g. a proxy/redirect side-channel) is in scope.
• capture proxy — any way the local logging proxy could forward credentials or data to an unintended destination.
• Supply chain — how frugon pins or fetches its dependencies and pricing data.

Out of scope: vulnerabilities in your own provider accounts, your local OS, or third-party models.

• No data egress in analyze — tokenize + price + arithmetic, fully local; asserted at the socket layer in tests (a regression that adds an outbound call breaks CI).
• --measure goes only to your providers, with your keys — never to us; keys are never logged or persisted (defense-in-depth socket-patching tests guard this).
• Open source — every claim above is checkable.

Built by Rodiun. MIT licensed.