Skip to main content
Open a computer-capable machine’s cloud desktop to look at a Linux app, use Chromium, or watch an agent work. The dashboard’s Desktop view uses authenticated WebSocket streaming.

Choose a template

Use a published computer or devbox-computer template for visible human watch and takeover. Coding templates include browser and desktop testing tools; agents and devbox start their testing display lazily on the first actual tool call.
desktop start explicitly starts the same lazy guest session. Tool discovery alone does not start it.

Open the desktop

Open the machine’s Desktop view in the dashboard, or mint a view-only link:
Watching disables keyboard, pointer, and clipboard input at the guest server. The viewer scales the screen, supports fullscreen, and reconnects when the connection drops. Watching follows the machine’s idle policy.

Computer use

With a credential profile configured:
Inside the machine, discover installed driver guidance before using its tools:
Use Chromium as a browser for AI agents. Use browser tools for web content and desktop tools for native apps, browser chrome, and dialogs. Observe the current screen or tree, act on current references, then verify the effect. Refresh references after navigation or UI changes. Check installed help for supported options and error dispositions. Use task-owned browser sessions. For direct Chrome CDP sessions in the guest, keep the persistent profile at /root/.soml-chrome-stealth.

Ask for human help

Use watch and takeover when a human needs to interact. During private human control, managed observation returns human_private_control. Poll the control state before resuming agent actions.