OpenAI Codex Computer Use Comes to Windows — Reading "Windows users, this one's for you." from the Primary Sources and What It Means for Japanese Enterprises (May 2026)

OpenAI's three-line message — "Windows users, this one's for you. Computer use now works on Windows, so Codex can take action on your Windows computer." — maps to Codex desktop app v26.527, released May 29, 2026, per the official Codex Changelog entry "Computer use and mobile access on Windows."

Four headline points:

1. Windows support for Computer Use lands May 29, 2026. Previously macOS-only since "Codex for (almost) everything" on April 16, 2026 2. Windows runs foreground-only — unlike macOS's parallel background mode, Codex physically takes over the pointer and keyboard while it works 3. EEA, UK, and Switzerland are excluded at launch. Japan is in scope 4. Install via Microsoft Store or `winget install Codex -s msstore` — no WSL required, OS-level PowerShell sandbox

Most Japanese office endpoints run Windows 11 / Windows 10. With the April release being macOS-only, this Windows launch is effectively Computer Use's real debut for the Japanese enterprise market.

"Codex" today is not a single product. The form factors:

What's new on Windows this week is Computer Use inside the desktop app. CLI and IDE extensions have been Windows-capable for some time.

Codex with Computer Use sees the screen, clicks, and types as if it had its own cursor, controlling any desktop app — including those without an AI API. Conceptually it is OpenAI's answer to Claude Computer Use and the evolution of OpenAI's prior Operator. Docs: Computer Use – Codex app.

Expected uses:

- Visual review of frontends, screenshot QA - Automation of legacy business apps that have no AI API - E2E and regression testing - Browser-driven internal-tool operation - Commenting on a page in the in-app browser and pointing the agent at it

The single most important spec detail in the docs:

> On Windows, computer use runs on the active desktop. It can't operate in the background while you keep using the same Windows session, so expect Codex to move the pointer, type, and take over the foreground while the task runs.

Translation:

- You cannot use the same Windows session for your own work while Codex is operating - Codex physically drives the mouse and keyboard - Target apps must be visible on the active desktop

Realistic deployment patterns on Windows:

- A dedicated secondary PC separate from your primary workstation - A VM / Remote Desktop sandbox with RDP from the host - After-hours runs while you're away - Separate Windows user account so you can switch contexts

OpenAI describes the Windows sandbox in Building a safe, effective sandbox to enable Codex on Windows.

Technical highlights:

- OS-level sandbox in PowerShell (restricted token + filesystem ACLs) - WSL not required — pure Windows-native - File edits and shell commands go through Codex's standard approval / sandbox flow - Per-app access prompts on first use

Items not explicitly documented:

- Whether screen captures and action data are uploaded to OpenAI's cloud — Computer Use doc does not say explicitly. Realistically expected to be uploaded - Retention period and whether the data is used for training — verify via your enterprise DPA

- Install: Microsoft Store or `winget install Codex -s msstore` - Architecture: x64 confirmed via the MSIX package name (`OpenAI.Codex_26.506.2212.0_x64...`) - Windows version: not explicitly documented. Third-party reporting treats Windows 11 as the primary target (PCWorld) - ARM64 (Snapdragon X, etc.): not documented. Copilot+ PC users should verify - Windows Server: not documented - Windows 10: not documented (Windows 11 is the implied recommendation)

The docs spell it out:

> In the Codex app, computer use is available on macOS and Windows, except in the European Economic Area, the United Kingdom, and Switzerland at launch.

- In scope: US, Canada, Japan, Korea, Australia, Southeast Asia, etc. - Excluded at launch: EEA (EU + Norway / Iceland / Liechtenstein), UK, Switzerland - Reason isn't stated — likely tied to GDPR and the EU AI Act

Japan is in scope from day one.

Codex isn't a standalone purchase — it's bundled across the ChatGPT plans.

Codex switched to token-based pricing on April 2, 2026 (previously message-count). Whether Computer Use specifically requires a higher tier is not explicitly documented — third-party reporting treats it as broadly available.

Refs: Codex Pricing / Codex rate card

Codex Computer Use's positioning: ChatGPT-bundled, natural-language-first, coding-and-OS unified. Deterministic RPA still wins where reliability is the primary metric.

Before rolling this out in production on Windows:

1. Foreground-only → dedicated PC or VM is the practical approach 2. Frequent approval prompts during the first run 3. Runs as your OS user → least-privilege policies need explicit design 4. Screen content sent off-device → not documented in detail. For finance / healthcare / public-sector use, confirm via DPA 5. EEA / UK / Switzerland blocked at launch → globally distributed Japanese firms need region-specific routing 6. ARM64 / Win 10 / Server unconfirmed → verify before deploying on Copilot+ PCs or legacy fleets

Japanese office endpoints overwhelmingly run Windows 11 / 10. This launch matters because:

- April's macOS-only Computer Use was effectively unreachable for 99% of Japanese enterprises. May 29 is the actual market debut for Japan - Coexistence with existing RPA (WinActor / UiPath / Power Automate Desktop) — RPA is deterministic, Codex is natural-language-flexible. They are complementary, not competitive - Broad app coverage — Excel, Outlook, Edge, Teams, Cybozu, kintone, internal web tools where the AI API surface is thin - After-hours automation — given the foreground-only constraint, off-hours batch runs are the realistic deployment pattern

Our AI consulting practice typically deploys Codex Computer Use alongside Forward Deployed Engineer–style on-site enablement. "RPA + AI agent hybrid" is going to be the dominant office-automation theme for mid-market Japanese firms through late 2026.

Looking at the broader market, May 2026 stacks up as the moment AI agents formally leave the codebase and start operating the OS and apps directly.

- [Claude Code Agent View](../columns/claude-code-agent-view-parallel-orchestration-2026) (May 2026) — multi-session orchestration with worktree isolation - [Cursor Automations](../columns/cursor-automations-agents-window-may-2026) (May 2026) — Agents Window + schedule / webhook triggers - [Google Antigravity 2.0](../columns/google-antigravity-2-agent-platform-2026) (May 2026) — from IDE to agent platform - OpenAI Codex Computer Use on Windows (May 2026) ← this column - [Argent × Gemma 4](../columns/argent-software-mansion-on-device-agent-ios-2026) (May 2026) — on-device agent driving an iOS simulator

Q1. When can I start using this? A. May 29, 2026 onward via Codex desktop app v26.527 (Microsoft Store or `winget install Codex -s msstore`). Q2. Does Plus include it? A. Not explicitly spelled out, but reporting suggests broad availability. Plus / Pro / Business / Enterprise all include Codex; usage caps differ. Q3. Will it run on Windows 10 or ARM64? A. Not documented. Confirmed x64 via the MSIX package name; Windows 11 is the implied target. Test before deployment if you're on Copilot+ PC or Windows 10. Q4. macOS vs Windows differences? A. The big one is foreground-only on Windows. macOS lets you run multiple agents in parallel in the background; Windows physically takes over your cursor and keyboard. Q5. Do I need WSL? A. No. The sandbox runs natively in PowerShell with OS-level restrictions. Q6. Is Japan in scope? A. Yes. Only EEA, UK, and Switzerland are excluded at launch. Q7. Should I replace WinActor / UiPath? A. Run them side by side. RPA is reliable and deterministic; Codex Computer Use is flexible on semi-structured work and exceptions. Map structured templated work to RPA, semi-structured edge-case work to Codex. Q8. Is screen content sent to OpenAI? A. The doc doesn't say explicitly. Realistic assumption: yes. For finance / healthcare / public sector, confirm via your DPA. Enterprise contracts have different data handling clauses.

May 29, 2026 is when AI agents directly drive Japanese office endpoints became a real possibility. The macOS-only April release didn't reach the Japanese enterprise majority; this one does.

The constraints are real — foreground-only on Windows, frequent approval prompts, unclear data flow for sensitive screen content, EEA / UK / Switzerland blocked. Build production rollouts on a dedicated machine, after-hours batch model, RPA coexistence, DPA verification baseline and this becomes one of the most useful new tools in your H2 2026 stack.

Primary: - Codex Changelog - Computer Use – Codex app - Windows – Codex - Codex for (almost) everything (Apr 16, 2026) - Building a safe, effective sandbox to enable Codex on Windows - Codex Pricing - Codex rate card Third-party: - PCWorld — Codex on Windows at last - VentureBeat — Codex Desktop Update - VentureBeat — ChatGPT Pro $100 tier - BuildFastWithAI — Codex mobile - Releasebot — Codex Updates - GitHub openai/codex Issue #21692 Related columns: - OpenAI Symphony — ticket-driven Codex orchestration - Claude Code Agent View deep dive - Cursor Automations in the Agents Window - Google Antigravity 2.0 deep dive - Argent × Gemma 4 — on-device AI agent - Forward Deployed Engineer (FDE) Note: An exact X post URL matching "Windows users, this one's for you." wasn't located in this research, but the content matches the official Changelog entry of the same date. Specific Windows version support (10 / 11 / Server), ARM64 status, and Computer Use data retention are not officially documented as of May 30, 2026 — re-verify in current docs and your DPA before deployment.