Skip to main content
When your organization’s Desktop policy allows the Shell, agents in app conversations can run commands on your Mac or Linux computer, for example tests or a build in a trusted project. The computer strip shows Shell on.

Allow commands and file changes for a chat

The first command in a conversation asks Run a command? with the full command. The first file write asks Write a file? with the path. Moving an item to the Trash and running a script ask the same way.
Run a command? dialog showing npm test -- --watch=false, with Not now, Always allow, Always in this chat and Allow for now

The first command in a conversation asks first. Shown with sample data.

Stopping a conversation, or a policy change, makes every conversation ask again, except for what you chose Always allow for. Always allow is kept for your account on this computer only, and a policy that turns the ability off still wins. Settings > Shell & files lists each one, for example Run commands: always allowed, with Ask each time to take it back. Signing out, Forget this Mac or a new xpander address clears them all. Always allow is in Omni Desktop 0.1.0-internal.715 and later (Settings > About shows yours). An earlier build, such as one an Air-Gapped install may still serve, offers the other three choices only.

What a command can reach

Each command runs in its own sandbox, which xpander builds from your trusted folders for that one command. The sandbox uses the operating system’s own isolation, so the shell needs no Docker and no virtual machine: In both, a command can reach only this: Shell network is on by default. Your organization can turn it off for everyone (Shell can reach the network in the Desktop policy), and you can turn it off for this computer in Settings > Shell & files. Commands then still run, without network access.

When the shell does not run

  • The policy turns the Shell off, or you turned Shell commands off in Settings.
  • Files are off for this computer, in the policy or your Settings.
  • No folder is trusted yet.
  • The policy is read only, so the shell cannot write anywhere.
  • The computer cannot run commands safely, for example a Linux computer without bubblewrap (command -v bwrap prints nothing). The strip reads Shell unavailable and agents work with your files directly. See Where it runs.
Each command is recorded with its outcome and duration, never the content of your files.