Policy before execution
Principal/resource rules, tool overrides, capability modes and connector presets determine whether a request may run, must ask locally or is denied.
Security model
Useful AI access without pretending machine automation is harmless.
RunOnMine is a policy and approval boundary, not a sandbox. It is designed to make machine authority explicit, narrow resource scope and fail closed when dangerous authorization or audit state is unavailable.
Filesystem boundaries
File operations are constrained to explicit selected roots and use descriptor-relative traversal rather than a check-then-open pathname race.
Browser isolation
Owned browser sessions use a disposable profile and network controls. Remote connectors cannot opt into private-network browser access.
Credential handling
Secrets live in operating-system credential stores or the documented encrypted headless Linux fallback. Managed connector credentials are not printed into normal output.
Tamper-evident audit
Audit records form a tamper-evident chain and summarize sensitive arguments instead of recording raw tokens, cookies or stdin content.
Emergency Lock
Lock is an explicit local stop control for the agent, managed connectors, pending approvals, OAuth sessions and temporary connector credentials.