People and agents
One shared desktop. Watch live, take control, and how your agent reaches you.
A screen is shared: your agent and you use the same desktop, with one mouse and one keyboard.
Watch live
Open the screen in the dashboard to see it as it happens. An agent can also hand you a link with POST /screens/{id}/viewer: a view-only live link that works once, must be opened within 60 seconds and ends after 60 minutes.
Take control
Press Take control in the dashboard and confirm, and your mouse and keyboard go straight to the screen, for example to type a password or solve a CAPTCHA.
While you have control, your agent's screenshots, actions, new commands, file calls, stop, restart and delete get 409 CONTROL_HELD. Nothing it does lands on top of your work, and it can't look while you type. It can still read the screen's status and history and follow commands already running.
Programs already running on the screen keep going. Kubi refuses calls and freezes nothing, so an agent running inside the screen isn't stopped by this. Stop it first if it should wait.
Control ends when you press Release, or after 5 minutes with no input from you. Any key or mouse move resets the timer, and you get a warning in the last 60 seconds. Closing the tab just lets the timer run out, so a reload doesn't hand the screen back.
Your clicks and typing count as input, so afterwards an action your agent planned on an earlier screenshot gets 409 STALE_OBSERVATION and has to look again. It never clicks blind over your work.
Nothing you do in the live view is saved to the screen's history. History shows when you took control and when it ended.
When your agent needs you
Kubi doesn't carry messages between your agent and you. When it needs a person (a CAPTCHA, a 2FA code, a yes or no), it asks you the way it always does, in its own chat or notifications, and it can send you a view-only link to show you the screen.
Several agents on one screen
There is no lock between agents: any key with the permission can call a screen at any time. Clicks and typing run one call at a time, in the order they arrive, and commands run side by side. Two agents clicking on one screen share one mouse, so give each agent its own screen.