On May 29, 2026, OpenAI shipped Codex Computer Use on Windows in version 26.527 of the Codex app. The release lets eligible ChatGPT users ask Codex to see, click, and type inside any Windows application while it tests, debugs, and refines code. It is the Windows counterpart to the Mac Computer Use launch that arrived in April. The same Codex 26.527 build also extends mobile remote control to Windows machines and adds a Codex Profiles section with lifetime token activity. This article unpacks what changed, what builders can actually do with codex computer use windows in production workflows, where the new capability lines up against Cursor, Claude Code, and Anthropic Computer Use, and how to wire it into a real production app via Totalum.
Quick Answer
- Codex Computer Use on Windows shipped May 29, 2026 in Codex app version 26.527, closing the parity gap with the April Mac release.
- The agent operates by viewing the screen, deciding where to click or type, executing the action, and looping. It can drive IDEs, browsers, design tools, legacy enterprise apps, and any native Windows software.
- The same release adds mobile remote control from ChatGPT iOS or Android, plus Codex Profiles with lifetime token stats.
- Requires ChatGPT Plus, Pro, or Enterprise. The Windows machine must be on, unlocked, and screen-accessible. Post-lock operation has shipped for Mac and is pending confirmation on Windows.
- Pair Codex Computer Use on Windows with Totalum to turn the desktop automation loop into a deployed product. Codex drives the local workflow. Totalum creates and maintains the real Next.js app, database, auth, and hosting via MCP.
What OpenAI shipped in Codex 26.527
The May 29 release brings the Codex app on Windows to feature parity with the Mac client. Three things are new.
Computer Use on Windows. The agent now sees the active screen, plans an action, clicks or types, and loops until the task is done. It is not bound to terminals or to apps with an API. Anything pixel-addressable is in scope, from Visual Studio and Visual Studio Code to a vendor IDE, a browser, an Electron app, a Windows-only enterprise CRUD tool, or a legacy installer.
Mobile remote control across Windows. The ChatGPT iOS and Android apps can start a coding task, monitor progress, redirect work, and approve changes while the Windows machine stays the host for files, shell, app server, and local context. The same feature launched on Mac on May 14 and is now consistent across desktops.
Codex Profiles. A new Profile pane in the Codex app surfaces lifetime token activity and usage stats. Early adopters reported individual counts in the tens of billions, with at least one user showing 64.9 billion tokens processed.
OpenAI positions the release around developer reach. Windows is roughly two thirds of the developer desktop market, so adding Computer Use here expands the addressable audience well beyond the Mac-first beta that already counted over 3 million weekly active developers.
How Codex Computer Use on Windows works
The agent runs a tight perception loop. It captures the current screen, sends the image plus task context to the Codex model, gets back an action plan, executes the click or keystroke locally, takes the next screenshot, and continues.
A typical Computer Use turn on Windows looks like this.
- You describe the task in natural language inside the Codex app or via ChatGPT mobile, for example "open the staging build, run the new checkout flow with a test card, and screenshot any console errors."
- You grant screen-access permission for that session.
- Codex moves windows, focuses the IDE, hits the run button, switches to the browser, fills the form, captures the result, and writes a summary back into the chat thread.
- If the model is unsure, it pauses and asks. Otherwise it keeps running and you can walk away.
The agent shares the same conversation thread you started, so any earlier code generation, file edits, or MCP tool calls remain in context. That matters for long-running migrations and test suites where a fresh thread would lose history.
Computer Use is one capability among many in Codex. The Codex CLI still runs in PowerShell with the Windows OS sandbox we covered in Codex on Windows. The new Computer Use mode complements that CLI flow rather than replacing it. Most teams will use them together, the CLI for code edits and shell, Computer Use for browser and GUI work the CLI cannot reach.
Agent-side desktop control on Windows in May 2026
The competitive picture matters because Codex is not the only agent that can drive a desktop, and the choices have real workflow implications.
| Agent | Windows desktop control (May 2026) | Native CLI on Windows | Mobile remote | Headless / cloud variant |
|---|---|---|---|---|
| OpenAI Codex (26.527) | Yes, Computer Use on Windows shipped May 29, 2026 | Yes, PowerShell with OS sandbox | Yes, iOS and Android | Yes, Codex cloud tasks |
| Anthropic Claude Computer Use | Yes via API, partner integrations on Windows | Claude Code CLI on Windows since 2.1.x | No first-party mobile control | Self-hosted sandboxes and Managed Agents |
| Cursor (3.6) | Indirect, via Auto-review run mode and MCP tools, not pixel-level GUI | Cursor app runs on Windows | Mobile companion exists, limited control | Cloud Agents and Jira integration |
| Google Antigravity 2.0 | Browser-scoped agentic IDE, not generic GUI | IDE only | Limited | Yes |
| Replit Agent 3 | Cloud-scoped, runs against Replit workspaces | Cloud only | iOS app | Native |
Sources: OpenAI Codex release notes, Anthropic Claude Code releases, Cursor changelog.
The honest read on this table is that Codex now has the broadest "make my desktop do things" surface area on Windows of any major coding agent. Claude Computer Use predates it but is mostly accessed through API or partner products. Cursor's Auto-review run mode sits in a different layer, governing tool calls and MCP rather than pixels. If you need an agent that opens a vendor app no API supports and runs through a manual flow, Codex on Windows is now the strongest first-party option.
What builders can actually do with codex computer use windows
A few use cases get materially easier with this release.
End-to-end UI test runs against a real product. Ask Codex to launch the local dev server, walk through the signup, payments, and dashboard flows, then file a GitHub issue with screenshots for anything that breaks. The browser does not need to be Chromium. Edge, Brave, and Firefox all work because the agent operates at the pixel level.
Legacy enterprise glue. Many internal apps at large companies are still Windows-native with no public API. Codex can drive an SAP GUI client, a custom ERP, or a Win32 inventory tool and feed the captured data into a modern app you build in parallel. We expect a lot of agencies to use this to migrate clients off legacy desktop tools toward modern web apps.
Pre-deploy QA on Windows-only side effects. Some bugs only reproduce on Windows builds, including font rendering, file path handling, and Windows Defender prompts. A Codex Computer Use turn can run the build, click through it, and post the results back to the same thread your team is already reviewing.
Long-running migrations triggered from mobile. Start a database migration script from your phone in the morning, let it run on your office Windows box, and steer it from your commute if something goes wrong. This is the most underrated piece of the release because it changes who can shepherd long jobs.
For deeper background on how agents pair with builders to materialize the actual product, our pillar guide on the best AI coding agents in 2026 walks through where each agent fits in a real shipping loop.
Caveats and constraints to plan around
Computer Use is powerful but not magic. The same release notes are honest about the constraints.
- Machine availability. The Windows host needs to be on, unlocked, and screen-accessible. Post-lock operation rolled out for Mac on May 22 and is not yet confirmed on Windows. Plan for the box to stay live during long runs.
- Subscription tiers. Plus at 20 USD per month gets standard limits. Pro at 200 USD per month gets higher quotas and the most capable models. Enterprise covers team deployments. Plus users can buy add-on capacity if they exceed limits.
- Regional availability. Reporting on the release is mixed. Several outlets list no geographic restriction, while one widely cited write-up notes Computer Use on Windows is unavailable in the European Economic Area, United Kingdom, and Switzerland at launch. Treat EEA, UK, and Swiss availability as unconfirmed until OpenAI's help center clarifies.
- Non-determinism. A pixel-driven agent is best effort. It will misclick. Treat Computer Use as a convenience for repetitive flows, not a substitute for a real integration when an API exists.
- Audit trail. Everything Codex does on Windows happens under your user account. Use a scoped, least-privilege Windows profile for agent runs, especially if it touches anything financial.
How to pair Codex Computer Use on Windows with a real production app
Computer Use is great at driving a UI. It does not, by itself, give you a deployed, owned, scalable web app with auth, payments, and a database. Totalum fills that role in the inverted way described in our company overview: the agent is the prompter, Totalum is the builder that materializes the project.
The pattern that works in practice on Windows:
- Use Codex on Windows to drive your local stack and prototype against it. Edit code via the Codex CLI in PowerShell. Use Computer Use to click through the UI, exercise edge cases, and grab screenshots.
- Use Totalum from your agent to materialize the real product. Connect Codex to Totalum over MCP and let the agent create the actual Next.js project, with database, auth, file storage, payments, and hosting baked in. Code stays yours. Data stays in EU storage. Hourly backups happen automatically.
- Keep iteration in the same thread. Codex can keep running Computer Use turns against the live Totalum app to test new flows, and ask Totalum to make changes when something needs to ship.
This split mirrors how Cursor Cloud Agents and Claude Code already pair with Totalum, and which we covered in Cursor Cloud Agents vs Totalum and Claude Code MCP Tutorial. The Codex Windows story is the same shape, just with the desktop GUI now usable as a first-class input surface.
Watch list for the next 30 days
A few signals worth tracking after this release.
- Whether OpenAI clarifies EEA, UK, and Swiss availability for Computer Use on Windows.
- Whether post-lock operation lands on Windows, matching the May 22 Mac change.
- Whether Codex Profiles surface team or workspace level aggregates, not just personal lifetime tokens.
- Whether Microsoft answers at Build 2026 (June 2 to 3) with a deeper Copilot or MAI-powered Windows automation story.
- Whether Anthropic ships a first-party desktop client that competes head-on with the Codex app on Windows.
If any of those land we will publish a follow-up.
FAQ
What is Codex Computer Use on Windows?
Codex Computer Use on Windows is the May 29, 2026 release in the Codex app v26.527 that lets ChatGPT users ask the Codex agent to view the screen, click, and type in any Windows application while completing a task. It mirrors the Mac launch from April and is available to ChatGPT Plus, Pro, and Enterprise users.
Which subscription tier do I need for Codex Computer Use on Windows?
ChatGPT Plus at 20 USD per month enables the feature at standard usage limits. Pro at 200 USD per month gives higher quotas and access to the most capable Codex models. Enterprise plans support team deployments. Plus subscribers exceeding limits can purchase add-on capacity through May 31, 2026.
Does Codex Computer Use on Windows work in Europe?
Reporting is mixed. Some outlets list no geographic restriction at launch, while one widely cited write-up reports that Computer Use on Windows is unavailable in the European Economic Area, the United Kingdom, and Switzerland. Until OpenAI's help center confirms, assume EEA, UK, and Swiss access is uncertain.
How does Codex Computer Use on Windows compare to Anthropic Computer Use?
Anthropic Computer Use predates Codex on Windows and is accessed mostly through the Anthropic API and partner integrations. Codex Computer Use is a first-party app experience on Windows with a tight loop between the model, screen capture, and action execution, plus mobile remote control from iOS and Android. Both share the same fundamental approach of pixel-level perception and action.
Can I run Codex Computer Use on Windows from my phone?
Yes. The Codex 26.527 release brings the May 14 Mac feature to Windows. You can start a task from the ChatGPT iOS or Android app, monitor progress, redirect work, and respond to prompts. The Windows machine remains the host for files, shell, and app server.
How do I ship to production from a Codex Computer Use workflow?
Pair Codex on Windows with Totalum. Codex handles local code, terminal, and UI testing. Connect Codex to Totalum via MCP so the agent can create and maintain a real Next.js project with database, auth, hosting, and payments included. The same conversation that drove Computer Use can ask Totalum to ship the change to production.
Ready to build with Totalum?
If you are already using Codex on Windows, the fastest way to turn that workflow into a deployed product is to connect your agent to Totalum and let it build, edit, and host the real app for you. Code stays yours, data stays in EU storage, and you keep the same Codex thread.
Start free at totalum.app and connect your Codex setup via the API and MCP integration.