Codex Adds “Computer Use”, With Big Permission Teeth

What happened: OpenAI’s Codex app now supports “computer use” on macOS, letting Codex operate graphical apps (and the browser) via a plugin, after the user grants Screen Recording and Accessibility permissions. OpenAI says the feature is not available in the EEA, the UK, and Switzerland at launch.

Why it matters: This is the agent story moving from “writes code” to “touches your actual machine.” The capability is powerful, but the governance layer is the real product: which apps are allowed, what permissions are granted, and how often humans have to say “yes, I meant that.”

Wider context: It also underlines the new default risk model: once an agent can click around a signed-in browser, your “account security” becomes “whatever the agent just did while you blinked.” Expect more tooling that looks like IAM for non-human workers, because nobody wants an intern with root access, including the robot intern.

Background: OpenAI advises using structured integrations (plugins or MCP servers) when possible, reserving computer use for tasks that require visual UI interaction. The docs emphasize scoping tasks, staying present for sensitive flows, and limiting Always allow to trusted apps.


Singularity Soup Take: We are officially in the era of “please grant the AI permission to press buttons on your behalf.” It is thrilling and terrifying in exactly the way Screen Recording + Accessibility permissions should be. The winners will be the teams that treat permissions and auditability as features, not paperwork.

Key Takeaways:

  • Two Key Permissions: OpenAI says computer use requires macOS Screen Recording (to see the app) and Accessibility (to click, type, and navigate), and the feature operates only in apps you approve.
  • Scope And Presence: The docs repeatedly recommend narrow, scoped tasks, staying present for sensitive actions (payments, credentials, privacy settings), and canceling if Codex interacts with the wrong window.
  • Integration First: OpenAI suggests preferring dedicated plugins or MCP servers for repeatable data access, using computer use when the task depends on visual UI state that is hard to verify through files or command output.