Computer-Use & Risk Tiers

Jarvis can act on your computer directly — open apps, click UI elements, type, run admin actions — through its computer-use and vision layers. Because that power is real, it is gated by a four-tier risk system.

The four tiers

Tier	Behaviour
safe	Runs immediately, no prompt
monitor	Runs, but is logged for review
ask	Pauses and asks for your approval first
block	Refused outright

Priority: blacklist > whitelist > default

The tier for any action is resolved in a strict order: an explicit blacklist wins, then a whitelist, then the tool’s default tier. A whitelist entry downgrades an action to safe and records that it was approved by the whitelist.

The anti-confirmation-fatigue contract

A voice assistant that asks “are you sure?” before every action trains you to say yes without listening — and then a genuinely dangerous action sails through. Jarvis’s contract is the opposite: the whitelist silently approves routine work, so you are only ever asked when something genuinely warrants a decision. Fewer prompts, each one meaningful.

Elevation, per action

Administrative actions are elevated per action, never globally. Jarvis does not run as an administrator or a system service — it asks for elevation at the moment an action needs it, through your OS’s native prompt, and drops back down afterwards.

Cross-platform

Computer-use, app launching, UI-element clicking, the orb overlay, global hotkeys, and elevation have OS-specific implementations behind a shared capability seam. Windows is hardware-tested; the macOS and Linux desktop implementations are present in code but remain unverified on real desktop hardware. Where a capability is genuinely absent on your platform, it degrades to a logged no-op rather than crashing.