Blog ·
Handing a VM to AI agents
How to set, per VM, how much agents such as Claude Code, Codex, Antigravity CLI, and GitHub Copilot CLI may do.
What agents can do
Agents such as Claude Code, Codex, Antigravity CLI, and GitHub Copilot CLI use MCP to see a VM's screen, click and type, and run commands. You decide per VM how much they may do.
| Level | What the agent can do |
|---|---|
| Off | Nothing. The VM does not even appear in the agent's VM list. This is the default. |
| View | Reads the VM list, status, screen, and accessibility tree. |
| Control | View, plus clicks and key presses, running commands, moving files inside the agent folder, start and shutdown, and saving checkpoints. |
| Full auto | Control, plus the risky actions you switched on, without a confirmation dialog. |
What stays protected
- Risky actions ask on your Mac. Running as administrator, force stop, and restoring or deleting checkpoints open a confirmation dialog every time at the Control level. The level can only be changed in the app, so an agent cannot raise it.
- Mac files stay in the agent folder. The agent reads and writes Mac files without asking only inside the folder set for each VM. Paths outside it still ask, even in Full auto, unless you turn that on separately.
- Activity is logged. The app shows which agent used which tool and when. Command contents, typed text, and output are not recorded.
- Take input back at any time. While an agent holds input, the viewer title bar shows “Agent in control”. Click it or start typing and control returns to you.
Connecting
In VM settings → AI Agents, a button adds the anyos MCP server to the installed Claude Code, Codex, Antigravity CLI, and GitHub Copilot CLI. One server handles every VM.