OpenAI Bring Codex Coding Agent to Windows with Screen Control

Bringing its native Windows client to feature parity with macOS, OpenAI has officially rolled out “Computer Use” screen control and remote orchestration to the Codex desktop application for Windows.

The blockbuster update completely transforms Codex from a traditional, text-restricted coding assistant into an active, autonomous engineering workstation. Moving far beyond basic inline code autocomplete, the Windows-native agent can now actively look at a user’s screen, move the mouse cursor, click buttons, and type inside external software applications.

The update is rolling out immediately via the Microsoft Store and is accessible to users authenticated through ChatGPT Plus, Pro, Business, or Enterprise accounts.

1. How “Computer Use” Operates Natively on Windows

While the original February 2026 macOS release allowed Codex to run hidden, sandboxed operations in the background, the fundamental architecture of Microsoft’s operating system requires a different execution strategy. On Windows, Computer Use operates strictly within the foreground of the active desktop.

The Foreground Takeover

When an agent is triggered to interact with a graphical user interface (GUI), it actively takes control of the local mouse pointer and keyboard. Because it literally interacts with the visible environment, developers will watch Codex manually open windows, click through drop-down menus, and type text in real time.

Real-World Engineering Use Cases

Instead of manually copy-pasting code or errors back and forth, developers can use the @ trigger system to pull desktop applications directly into a Codex thread context:

UI Bug Reproduction: A developer can prompt: "@Chrome Open localhost:3000, click the checkout button, find why the modal crashes, and fix the underlying script." Codex will open the browser, reproduce the error state, view the crash parameters, edit the source code, and re-run the UI loop to verify the fix.
Cross-App Prototyping: Designers and engineers can point Codex toward wireframes inside tools like Figma, instructing the agent to actively extract visual layouts and translate them into functional React components inside an IDE like Visual Studio or VS Code.

Security Note: To use this feature, developers must navigate to Codex settings, open the Computer Use tab, and manually click install to authorize the desktop-interaction plugin. The feature is strictly permission-scoped; it activates only when a user explicitly tags an app or mentions a computer-use command.

2. The OS-Level Sandboxing Environment

To protect local machines from destructive agent loops or unwanted filesystem contamination, OpenAI has backed the Windows application with strict security infrastructure:

The PowerShell Sandbox: When Codex runs commands via native PowerShell, it wraps the agent within an isolated Windows sandbox. This applies ACL-based filesystem boundaries that block the AI from writing files outside the explicitly designated working project folder.
WSL2 Compatibility: For developers working heavily within Linux-centric pipelines, Codex can be toggled in settings to run inside the Windows Subsystem for Linux 2 (WSL2) environment, which instantly shifts the agent to a highly secure Linux-native container sandbox.
Bounded Privileges: By default, the agent operates with restricted user tokens and offline-level environment controls. If a specialized development task requires administrative execution, the user must explicitly launch the master Codex application using the Run as Administrator prompt.

3. Turning Smartphones into Remote Engineering Controllers

Simultaneously with the desktop patch, OpenAI updated the official ChatGPT app for iOS and Android to support full remote control over Windows host machines.

               [ CODEX REMOTE CONTROL PIPELINE ]
                              │
       ┌──────────────────────┴──────────────────────┐
       ▼                                             ▼
 [ ChatGPT Mobile App ]                        [ Windows Host PC ]
 • Trigger threads remotely                    • Runs tasks in active foreground
 • Review end-of-turn diffs                    • Confined to secure local sandbox
 • One-tap approval/rejection                  • Operates IDE, terminal, & browser

The cross-platform link allows developers to treat their mobile phones or tablets as portable command consoles. While away from their desks, engineers can open the ChatGPT app to initialize multi-step agent threads, review real-time progress trackers, read automated end-of-turn code diff summaries, and issue one-tap approvals or rejections for code changes running locally on their main Windows workstation.

Because foreground screen control completely dominates a user’s desktop session, OpenAI explicitly recommends using mobile remote control as an ideal way to monitor and guide massive, long-running diagnostics, migrations, or test suites while leaving the host machine completely dedicated to running the physical GUI tasks.

Search for an article