Watch the agent's desktop and take control

The session header carries five tabs — Desktop, Code, PRs, Auto QA and Scheduled — each opening its pane next to the chat (click again to hide it), and a ⋯ menu with the rest: Terminal, Context, Snapshot, Fork…, Pull to folder… (sessions with a copied folder), Session settings, Stop and Delete. Resume shows in the header while the box is stopped.

Desktop opens the box's screen next to the chat. While the agent is working the view is read-only so you don't fight over the mouse; Take control hands it to you until the agent's next turn. When the agent is idle the desktop is always interactive: log into a site for it, open a program, arrange windows.

The agent works the same screen through the built-in desktop MCP server: screenshot, zoom, the mouse (left_click, right_click, double_click, left_click_drag, scroll…), the keyboard (type, key, hold_key), wait, and the recording tools (start_recording, annotate_recording, stop_recording, narrate_recording). Every tool, its parameters and limits, and what it returns is in the MCP tools reference.

The agent's cursor glides rather than jumps: every move, click, scroll and drag with a target eases from where the cursor is to where it goes, about 100 ms for a short hop and up to 300 ms across the screen, at ~120 positions a second. You see it in the Desktop pane and in recordings, and the program under the cursor gets the same run of motion events a hand-moved mouse produces — hover effects fire, and drag-and-drop that needs to see the pointer travel works. SESSIONBOXER_MOUSE_GLIDE=0 in the MCP's environment (in the box's .mcp.json) makes the cursor jump instead.