OpenAI brings Codex computer use to Windows PCs with ChatGPT mobile app steering.
Bringing native agentic computer use to Windows OS is a massive step forward for AI-driven RPA and developer workflows. The addition of mobile app steering introduces asynchronous human-in-the-loop control, which is critical for safely managing long-running agent tasks. This signals OpenAI's aggressive push into OS-level execution, directly challenging Anthropic's recent computer use capabilities.
OpenAI has officially rolled out Codex-powered "computer use" capabilities for Windows PCs, accompanied by a novel control mechanism via the ChatGPT mobile app. This update allows the AI to autonomously navigate, click, and type across the Windows operating system, effectively acting as an OS-level agent. Notably, users can now start, review, and steer these automated tasks remotely using their mobile devices.
Technical Implications By integrating at the OS level, OpenAI is moving beyond isolated browser environments or sandboxed code execution. While the exact underlying architecture wasn't fully detailed in the brief announcement, the capability implies a robust multimodal understanding of the Windows GUI. The most architecturally significant feature is the mobile app integration. By decoupling the execution environment (the PC) from the control plane (the mobile app), OpenAI has implemented an elegant asynchronous human-in-the-loop (HITL) system. Users no longer need to babysit long-running tasks on their desktops; they can approve actions, correct course, or halt execution on the go.
Why It Matters This release marks a major escalation in the AI agent wars, serving as a direct response to Anthropic's recent "computer use" API for Claude. For engineers and enterprise IT, this democratizes Robotic Process Automation (RPA), replacing brittle, selector-based automation scripts with resilient, vision-language-action models. The mobile steering feature specifically solves a massive UX hurdle in agentic workflows: trust and oversight.
What to Watch Next Security and permissions will be the immediate focal point. Granting an AI model autonomous control over a local Windows environment introduces significant risks regarding data exfiltration, unintended destructive actions, and prompt injection attacks. We will be watching closely to see how OpenAI sandboxes these capabilities and whether an API version will be released to let developers build custom OS-level agents. Additionally, watch for equivalent support rolling out to macOS and Linux environments in the near future.