ChatGPT's Codex Can Now Drive Your Windows Apps and You Can Check On It From Your Phone

What You Will Learn

• What Computer Use on Windows lets Codex actually do
• The remote continuation feature and why supervising from mobile matters
• Who this is for and who should wait
• The regional and safety limits you need to know
• How it compares to Copilot Studio agents and Claude's computer use

There is a category of work that no chatbot has been able to take off your plate: the tedious, click-heavy tasks that live inside a desktop app. Reconciling two systems that do not talk to each other. Running the same export, reformat, and upload sequence every week. Testing a build by clicking through it like a user would. Until now, an AI could tell you how to do those things but could not do them.

OpenAI's late-May update changes that for Windows users. Codex, the agent inside ChatGPT, now has Computer Use on Windows, which lets it see the screen, click, and type inside Windows applications, not just inside a browser. Paired with that is remote continuation: you can start a task on your Windows machine and then check progress, respond to prompts, and steer the work from ChatGPT on iOS or Android, or from Codex on a Mac.

The pitch for an executive is supervision without being chained to the desk. You set a job running, walk into a meeting, and approve or redirect it from your phone when it hits a decision point. That is a different relationship with automation than copy-pasting prompts all day.

This is a PromptHacker Premium deep dive.

The setup, the safe first tasks, and the limits that will trip up European teams are below.

What It Can Do That a Browser Agent Cannot

Earlier computer-use agents were mostly confined to the browser. They could navigate websites and fill web forms, but a native Windows application was a wall. Computer Use on Windows removes that wall. Codex can now operate desktop software the way a person does: reading what is on screen, moving the cursor, clicking buttons, and typing into fields inside the app itself.

For most executives the immediate value is not coding, even though Codex is a developer tool at heart. It is the class of repetitive, cross-application chores that have resisted automation because they required a human to drive a desktop interface. The update also adds Codex Profiles, which show your activity, usage stats, and token activity over time, plus general speed and stability improvements to the in-app browser.

Remote Continuation: The Feature That Earns Its Keep

The standout is not the clicking. It is the handoff. You can begin work on a Windows machine and continue the same thread from your phone or a Mac, responding to the agent's questions and steering it while you are away from the desk. An agent that needs constant attention is a second job. An agent you can supervise in glances between meetings is leverage.

Who Should Use It Now, and Who Should Wait

Use it now if you or your team run repetitive desktop workflows, build or test software, or manage processes that bounce between two applications that do not integrate. Wait if your work is mostly judgment and communication that already lives in the cloud, where a browser-based agent or a model like Opus 4.8 will serve you better with less setup.

Reality check: this is a prosumer and developer feature first. If the phrase build and test does not describe part of your week, the bigger story for you this week is Opus 4.8 or the Copilot redesign, not Codex.

The Limits You Need to Know

Two constraints matter. First, geography: Computer Use on Windows is not available in the European Economic Area, the United Kingdom, or Switzerland at launch, so distributed teams will have an uneven rollout. Second, supervision is not optional. An agent that can click and type inside your desktop apps can also click the wrong thing, so keep it on reversible, low-stakes tasks until you trust its behavior, and never leave a destructive action unattended.

How It Stacks Up

Microsoft made computer-using agents generally available in Copilot Studio in the same window, aimed at organizations building governed automations. Claude offers computer use through its own tools as well. The OpenAI version's edge is the consumer-grade mobile supervision: most people already have ChatGPT on their phone, so the handoff requires nothing new to install. The trade-off is that it is less governed than an enterprise Copilot Studio build, which is exactly why you keep early tasks small and reversible.

Action Steps Summary

• Confirm eligibility and region. Check that Computer Use on Windows is available for your account and that your team is not in the EEA, UK, or Switzerland.
• Pick one reversible chore. Choose a repetitive desktop task with no destructive step for the first run, such as a weekly export and reformat.
• Stay on the leash. Keep the ChatGPT app open on your phone so you can approve or redirect at each decision point.
• Use Codex Profiles. Review usage and token activity so a long-running agent does not surprise you on cost.
• Escalate only what works. Once a task runs cleanly twice, let it run longer between check-ins. Never automate a destructive action unattended.