I would like an explicit setting that controls when ChatGPT may begin operating applications on my computer.
My goal is predictable control over when desktop interaction starts. On Windows, Computer Use can take over the foreground, mouse and keyboard while I am working. I want to choose how frequently that takeover requires my confirmation.
In my current desktop app, Computer Use settings show an “Any App” switch and an “Always-allowed apps” list. I would find the following four modes much clearer:
| Mode | Behaviour |
|---|---|
| Off | Disable general desktop access through Computer Use. |
| Ask before first use in Session | Ask once before the first use. Approval remains valid for the current session. |
| Ask always | Ask before each new period of desktop control. |
| Ask never | Allow desktop control without an additional access prompt, within applicable permissions and policies. |
For this proposal, I would define a session as one ongoing task within a chat on a particular computer. Approval would expire when the task ends, is restarted, or I revoke it. It would not automatically carry over to other chats or tasks.
“Ask always” should apply when ChatGPT takes control again after returning control to me. Once I approve a continuous interaction, it could perform the necessary clicks and keystrokes without asking about every individual action.
For example, ChatGPT might use Edge, finish that interaction, and return control to me. If it later needs to operate the desktop again, “Ask always” would require another confirmation before it starts.
The selected mode should apply consistently to native desktop control, including browsers operated as ordinary Windows applications without a browser extension. Selecting “Ask always” should also take precedence over previously saved “Always allow” decisions.
The approval dialog should identify the application and briefly explain the intended interaction. Existing confirmations for consequential actions would remain separate from this desktop-access setting.
Would others find these modes useful? Is there currently a supported way to configure this behaviour?