Skip to content

Root Mode

Normally every Agent, the Terminal and Finder act as the aos user: no sudo, Protected Paths need an Approval, and the paths you lock stay locked. Root Mode is a second way to run the whole Machine in which all of that acts as root — Agents can run anything that needs sudo, every file and folder is unlocked, and Protected Paths are not enforced.

It is a separate Realm, not a setting you toggle inside a session. The Machine runs one Realm at a time, chosen at start from root_mode: in config.yml, so every switch restarts the daemon and reloads the Desktop. When Root Mode is running you will know it: a red ROOT badge in the menu bar, a banner across the Agent app, and a root@…# prompt in the Terminal.

  • Agents run as root. sudo is unnecessary — they are root. They can install packages, edit anything under /etc, and manage system services directly.
  • Nothing is locked. The Protected Paths list, your own 🔒 locks, the .env guard and the “don’t discard uncommitted work” git guard are all switched off. Finder shows your locks as not enforced in Root Mode.
  • The Terminal and Finder are root too. id prints uid=0(root), /root is your home, and every folder opens.

Risky actions still follow your Autonomy level. On confirm-risky an Agent is still stopped before an rm -rf or a package removal — being root removes the permission wall, not the are-you-sure one. Set Autonomy to auto if you want none.

Root Mode has its own history, kept in its own database, separate from your normal (Standard) Realm:

  • Its Tasks (chats), Audit Log, Memory, Notifications, Services, window layout and Browser profile are all its own.
  • In Standard Mode you see none of it, and the other way round. The daemon only ever serves the Realm that is running, so the Desktop and the CLI cannot show the other Realm’s history — there is no query that reaches across.
  • Even a root Agent is kept out of Agentic OS’s own state (/var/lib/aos, /etc/aos) by the kernel, so it does not stumble into Standard’s history or your API key.

It is like incognito, but not the same: Root Mode’s history is kept across restarts. When you want it gone, aos root clear (or Clear Root Mode history in the System pane) deletes that database and its command outputs, and the next Root start begins empty.

Root Mode is a privacy boundary between the two histories — not a security boundary against a root Agent. An Agent running as root can, if it really tries, get around anything on the machine, Agentic OS included: it can reach its own Approvals, undo its own confinement, and touch files outside any Realm. The separation keeps the two histories from mixing; it does not contain a root process that is determined to escape.

And it does not undo what happened. What root did to the server is real and stays. A package it installed is still installed, a file it wrote under /etc is still there, a service it started keeps running. Clearing Root Mode’s history erases Agentic OS’s record of the work, not the work. There is no Checkpoint or Replay that walks it back — Checkpoints do not cover what root does outside them.

Flip Start in Root Mode in the System pane, or run aos root on. You will read a warning of exactly the consequences above, confirm that you understand, and enter your Desktop password. The Machine restarts into Root Mode. The switch is refused while a Task is running and names the Task; cancel it and try again.

Because the Desktop password is already root on a native install, Root Mode does not hand an attacker anything they could not already reach — it hands your Agents the same authority, all the time, with the guardrails off. Turn it on for a job that needs it, and turn it back off when the job is done: leaving Root Mode stops every service it started and returns the Machine to its locked, aos-user self.

If this is more standing authority than you want in one place, the Docker Compose install keeps Root Mode’s blast radius inside a container instead of your server.