OpenAI Codex Now Controls Windows PCs Autonomously for Testing and Bug Hunting

OpenAI announced on May 29, 2026 that Codex Computer Use is now available on Windows 11, bringing the same autonomous screen‑control capabilities that launched on macOS in April to Microsoft's operating system, according to The Verge. Codex can now visually interpret what is on your screen and interact with any Windows application by seeing, clicking, and typing with its own cursor.

The feature is toggled on in Codex's settings. Developers can target specific programs using commands like @computer to give Codex full desktop access, or @Paint to restrict it to a single application, The Decoder reported. This is not just a port of the Mac version — it is a full Windows‑native implementation with permissions controls and background task support.

One of the most significant additions is background computer use. Codex can now operate apps while you are away from the computer (via mobile remote control), running QA passes, testing flows, and hunting bugs autonomously. Unlike macOS where Codex runs in true background mode, on Windows it occupies the active desktop, as detailed by.² Previously, computer use required the developer to be present and watching. Now, you can start a task, walk away, and review the results later.

Permission controls are built in. According to OpenAI's developer docs, Codex asks for permission before it can use an app for the first time. During a task, it can only see and take action in the apps you explicitly allow. This is a meaningful security consideration — an AI agent with screen control is powerful, but it should not have unrestricted access to everything on your machine.

Alongside the Windows launch, OpenAI announced that you can now start and monitor Codex tasks from the ChatGPT mobile app on iPhone and Android, The Verge confirmed. This means you can kick off a long‑running test suite or bug hunt on your desktop, then check progress from your phone while away from your desk.

The mobile integration is part of a broader push to make Codex a persistent coding agent that works across devices. As The Decoder noted, the rapid Codex expansion is part of OpenAI's plan to build a unified platform for work and daily life — a strategy that could eventually fold ChatGPT's capabilities into the Codex app.

The Windows Computer Use feature opens up several practical workflows for developers. The most immediate is automated QA testing: Codex can see the interface, click through flows, type into fields, and record what fails — essentially acting as an autonomous tester for desktop applications, as OpenAI's use case documentation outlines.

Beyond QA, developers on Reddit are discussing use cases like having Codex review pull requests by actually running the changed code, testing Windows desktop apps that lack API‑based testing infrastructure, and debugging UI rendering issues by watching the screen for visual regressions. The shift from API‑level testing to screen‑level testing is significant — it means Codex can now validate the actual user experience, not just the backend logic.

The Windows Computer Use launch puts OpenAI in direct competition with Claude Code's Dynamic Workflows (launched the same week with Opus 4.8) and GitHub Copilot CLI, which Microsoft is aggressively pushing internally. Each tool takes a different approach: Codex controls the entire desktop through screen vision, Claude Code orchestrates subagents within a terminal, and Copilot CLI focuses on command‑line integration.

For Windows developers, Codex now has a distinct advantage. Claude Code runs on Windows through WSL, but it does not have native Windows app control. Copilot CLI is terminal‑only. Codex is the only major AI coding assistant that can click through a Windows desktop app's UI to test it — a capability that matters for teams building desktop software, legacy Windows apps, or anything with a GUI that cannot be tested through an API. The question is whether this level of autonomy is something developers actually want, or whether it introduces more debugging complexity than it solves.