Help & FAQ
Common questions, including privacy and the Windows sandbox.
Q · Privacy: does Fettle send my code anywhere?
Fettle itself sends nothing and stores nothing. It runs locally and reads your agents' source and trace files on your machine. The one network step is Codex: when you run fix/coach/patch, the relevant code is sent to OpenAI Codex under your own account and plan — exactly as if you used Codex directly. Use --dry-run to see precisely what would be sent, first.
Q · Will Fettle change my code without asking?
No. Codex only ever edits a throwaway copy. Nothing touches your real agent unless the change passes the benchmark gate and you pass --apply. A failing gate is rolled back automatically, leaving your code untouched.
Q · What data does Fettle keep?
Only what's already on disk — your agent files and its own reports. No telemetry, no accounts, no cloud storage.
Q · On Windows, Codex exits with an error / the sandbox fails.
The native Codex sandbox can clash with some Windows antivirus software (e.g. 360), whose system objects break its file scan. Re-run with --bypass-sandbox (Fettle already works on an isolated copy), or run Codex inside WSL. If Codex returns instantly with no output, check your Codex plan/usage — it may be a quota limit rather than a sandbox issue.
Q · How much does it cost?
Fettle is free and open-source. Codex usage draws on your ChatGPT plan (or API key) — the same billing as using Codex yourself.
Q · Which agent frameworks are supported?
Fettle is framework-agnostic. Anything with source and trace logs works — LangChain, LangGraph, the OpenAI Agents SDK, or a custom agent.
Q · What if a fix would make things worse?
It can't ship. The benchmark gate must pass on a concrete, machine-checkable definition of “fixed” before anything merges; otherwise it rolls back.
Still stuck? Head to ContactMe.