Bypass permissions is the mode with no brakes. In bypassPermissions Claude acts with no prompts at all. No confirmation before edits, none before commands. It is the fastest mode and the most dangerous, so it deserves a clear rule of thumb about when it is acceptable.
What it actually does
Every other mode has a checkpoint. Default asks before edits and commands. Accept-edits asks before commands. Plan changes nothing. Bypass removes all of that. Whatever Claude decides to run, it runs.
You can start a session in it from the CLI:
claude --permission-mode bypassPermissions
Where it can make sense
The mode fits situations that are both controlled and disposable, where a mistake cannot hurt anything you care about:
- A throwaway container or sandbox with nothing valuable in it
- An isolated environment you can rebuild in seconds
- A tightly scoped automated job where the environment itself is the guardrail
The common thread is that the surroundings, not the prompts, keep you safe.
Where to avoid it
Do not use it on anything you would miss:
- Your main working machine with real projects on it
- A repository with a remote you can push to
- Anything with access to credentials, production, or shared systems
Without prompts, a single wrong command runs with no chance for you to say no. That is fine in a sandbox and frightening on a real box.
A safer default
Most of the speed people want from bypass is available from accept-edits plus a good allow list, which still stops before commands. Reach for that first. Keep bypass permissions for the rare, isolated case where the environment can absorb any mistake.
Comments
Be the first to comment.