User permission prompt before accessing apps

[Muhammad-Shayan-Haider]

Feature Request: Per-App Permission Prompts Before Computer Use

Problem

Currently, open-computer-use can access and interact with applications without explicit per-app user consent. Once macOS-level Screen Recording and Accessibility permissions are granted, the tool has broad access to any app on the system. This creates a security and privacy risk — the AI agent could potentially access sensitive applications (password managers, banking apps, email clients) without the user being aware or having a chance to intervene.

Proposed Solution

Implement a per-app approval flow similar to OpenAI's Codex computer use:

1. Permission prompt before each app access: When the agent attempts to interact with an app for the first time in a session, show a prompt asking the user to approve or deny access to that specific app.
2. "Always allow" option: Let users mark trusted apps so they don't get prompted repeatedly.
3. Manage allowed apps: Provide a settings view where users can review and revoke previously granted app permissions.
4. Sensitive action prompts: Optionally prompt before high-impact actions (e.g., submitting forms, deleting content).

How Codex Does It

In OpenAI's Codex app, the computer use feature:

• Shows a dialog: "Allow Codex to use [App]?" with options to allow once, always allow, or deny.
• Keeps a list of always-allowed apps in the Computer Use settings.
• Prompts separately for sensitive/disruptive actions.
• Keeps macOS system permissions (Screen Recording, Accessibility) separate from per-app approvals.

Why This Matters

• Security: Prevents unintended access to sensitive apps.
• User Trust: Gives users visibility and control over what the AI is doing.
• Industry Standard: Aligns with how leading tools (Codex) handle this concern.
• User Control: Users can revoke access to specific apps without losing full functionality.

Reference

• OpenAI Codex Computer Use docs — Permissions and approvals
• Related Codex issue: openai/codex#18896 (shows the approval mechanism in action)