Every top-level chat — a standalone chat or a project's main chat — has its own screen: a virtual display on macOS, a headless X session on Linux. T3 Code creates it the first time anyone opens that chat's computer view. A child chat has no screen of its own; it works on its parent's. Open one from the computer card in the conversation rail.
Watching is always safe. To use the screen yourself, choose Take control. While you hold control the pointer and keyboard follow you and agents in that chat are refused. Choose Stop controlling when done; the agent can act again immediately. Keys you are holding are released when you stop, close the view, or switch away.
The window list shows what is open on this chat's screen; selecting one raises it. Open app launches onto this chat's screen, not wherever you are looking.
Agents reach a chat's screen through the computer_* tools, automatically
scoped to the chat they run in — there is nothing to select. The screen is
shared, so an agent's input waits for you and is refused while you control it.
Turn this off with Settings → Integrations → Agent computer access (the same server setting applies to OpenBot). You keep the view either way; only agents lose it.
macOS asks for two separate grants, both for T3 Computer Helper. Granting one does not grant the other, and the view says which is missing:
- Screen Recording (Privacy & Security) to see the screen.
- Accessibility (Privacy & Security) to move the pointer, type, and place windows — without it, opening an app is refused rather than landing somewhere unexpected.
On macOS, every chat's screen is an extra screen on the one session you are signed into, which still has one pointer, one keyboard, and one frontmost app — shared between you at the Mac and every chat's agent. Nothing is isolated or sandboxed; it is more desks in the same room, not a room of your own.
On Linux, a managed screen is a separate X server with its own pointer, so it does not compete for input. See Computer on Linux.