Safety Model

OpenMind is built around explicit local approval. The agent can request tools, but the CLI enforces policy.

Diff before write

write_file and edit_file always print a diff before asking for approval.

Confirm commands

run_command, git status, git diff, and git commit prompt before running.

Protected paths

.env, private keys, certificates, and credentials are blocked by path checks.

Dangerous commands

sudo, filesystem formatting, reboot, shutdown, and registry deletion are blocked.

Important Boundary

OpenMind is not an OS sandbox. Approved commands run with your shell user's permissions.