OpenAI Expands Desktop Automation Capabilities as Codex Gains Windows Support
On May 29, 2026, OpenAI released version 26.527 of its AI programming assistant, the Codex desktop client. The core of this update is the official introduction of the “Computer Use” feature to the Windows operating system. This feature, which first launched on macOS in April 2026, allows the AI to directly control graphical user interface (GUI) applications by analyzing screen content and simulating user input. This marks an expansion of Codex from pure code generation and execution to an AI Agent with cross-application desktop automation capabilities.
What is the “Computer Use” Feature?
“Computer Use” is an advanced feature of the Codex desktop client that endows the AI with visual perception and simulated physical interaction abilities. Technically, the AI Agent works by capturing an image of the current screen, using a multimodal model to understand the elements, text, and layout of the interface. It then autonomously plans operational steps and simulates actions like mouse clicks, scrolling, and keyboard input through system APIs. This capability bridges the gap where traditional command-line tools cannot interact with graphical interfaces.
This feature extends the application of AI from coding environments to everyday desktop workflows, enabling it to perform a series of tasks that previously required manual operation, such as:
- Automated UI Testing: Simulating a user’s complete operational path within a desktop or mobile application and automatically verifying the correctness of interface elements through screenshot comparisons.
- Complex Bug Reproduction: Precisely recreating software defects that require a specific sequence of UI interactions to trigger, simplifying the debugging process.
- Configuration of No-API Applications: For legacy software that does not provide a command-line interface or API, batch settings and adjustments can be completed by simulating interface clicks.
- Cross-Application Data Coordination: Automating repetitive tasks like copying data from one application (e.g., a CRM system) and pasting it into another (e.g., a spreadsheet), serving as an intelligent upgrade to traditional Robotic Process Automation (RPA).
Three Key Updates in Version 26.527
This update includes three main aspects:
Windows Platform Support: Codex can now operate desktop applications in a Windows environment by taking control of the foreground window. During task execution, the target application must remain visible, and the mouse and keyboard will be controlled by Codex.
Mobile Remote Control: Users can remotely connect to and manage Codex instances running on Windows or macOS via the ChatGPT client on their mobile phones (supporting iOS and Android). This allows users to start tasks, monitor progress, and intervene from anywhere, without being tied to their computer.
Profile Page Optimization: The functionality of the Profile area has been enhanced. It now clearly displays user account details, resource usage statistics, and a detailed breakdown of token consumption, making it easier for users to manage costs and analyze usage.
Implementation Differences Between Windows and macOS Versions
Although the core functionality is the same, there are significant differences in how Codex is implemented on Windows and macOS, which directly affects the user experience.
macOS Version: Supports a mode called “locked use.” After starting a task, the user can lock the screen, and Codex can continue executing the task in the background with the help of an authorized plugin. This mode is ideal for long-running automation jobs.
Windows Version: Currently described as an “early access” feature, its implementation is more rudimentary. It requires tasks to be executed in the system foreground, meaning the target application window must be active and visible. Users cannot use the mouse and keyboard while a task is running. Background or locked-screen operation is not yet supported, but OpenAI has stated that it will continue to make improvements in the future.
In terms of permissions, macOS requires users to grant “Screen Recording” and “Accessibility” permissions. The main limitation on Windows is keeping the application in the foreground. Both platforms require the Computer Use plugin to be pre-installed.
Security Boundaries and Usage Recommendations
Granting AI the ability to directly control a local computer has raised significant security concerns. In response, OpenAI has designed multiple layers of security mechanisms to ensure user control:
- Permission Approval: The first time Codex attempts to access any application, a dialog box will appear requesting authorization. Users can choose to allow it once or always.
- Interrupt at Any Time: Users can pause or terminate an ongoing task at any moment to immediately regain control of their computer.
- Sandboxing Mechanism: High-risk operations such as file modification and Shell command execution still follow Codex’s original sandboxing environment and approval process. The “Computer Use” feature cannot bypass this restriction.
- Explicit Restrictions: To prevent misuse of permissions, Codex is explicitly forbidden from operating terminal applications, the Codex interface itself, running programs as an administrator, or automatically handling operating system security prompts (such as UAC pop-ups).
The official recommendation is for users to limit the scope of tasks to specific, singular goals and to close irrelevant, sensitive applications to ensure stability and security during use.
