Codex Computer Use — Let ChatGPT See and Operate the Interface on Your Mac or Windows
As of 2026-10-19, the docs say Computer Use in the ChatGPT desktop app covers macOS and Windows in supported regions; install the plugin inside the app and approve each app separately. macOS tasks can run in the background without taking over your machine; Windows tasks run in the foreground on the active desktop. At the 2026-04-16 launch it was Mac-only: see, click, and type across any Mac app, multiple agents in parallel, 90+ plugins, the Codex Super App vision landing. The EEA, the UK and Switzerland opened on 2026-06-16.
What Computer Use Is — From macOS to Windows
When it shipped on 2026-04-16, Computer Use drove macOS apps only: OpenAI let Codex see the screen, move a cursor, click, and type across browsers, desktop apps, data sources, and settings — any GUI work. The key difference from Anthropic's Claude Computer Use at the time: Codex ran in the background with no machine takeover — you kept using your Mac while Codex worked in its own window/virtual desktop, while the Claude version takes over the screen. That design enabled multiple agents to work on different things in parallel. OpenAI added Windows on 2026-05-29. The current wording is that, in supported regions, Computer Use in the ChatGPT desktop app is available on macOS and Windows, with ChatGPT Work and Codex. It can see and operate graphical user interfaces directly — work that command-line tools and ready-made integrations cannot cover. The difference: on macOS the task can run in the background while you keep using your computer, while on Windows it runs in the foreground on the active desktop.
When Computer Use Is the Right Tool
OpenAI's test: choose Computer Use when the task depends on a graphical interface that files and command output alone cannot verify. Fitting cases include testing a macOS or Windows app or a simulator flow, tasks that need your browser, reproducing a bug that only appears in a graphical interface, changing settings you have to click through, and reading a data source that has no plugin. Start a task by mentioning @Computer or @AppName and naming the exact app, window or flow. If the target app already exposes a dedicated plugin or MCP server, the docs say to prefer that structured integration and keep Computer Use for the parts that need to be seen or operated. Broken down into concrete actions: 1) GUI-only actions: change desktop-app settings, click through menus; 2) data sources without a plugin/MCP: Codex logs into a SaaS dashboard via the browser and pulls data; 3) cross-app workflows: read a dashboard in the browser and post results to Slack (the feature can't automate terminal apps — a limit the docs state); 4) reproducing GUI bugs: have Codex drive the app to trigger a bug and analyze the screenshot; 5) verifying setting changes: edit prefs, launch the app, observe the effect, full loop.
How to Turn It On + Parallel Agents and the 90+ Plugin Ecosystem
Open Computer Use under Plugins in the ChatGPT desktop app, install the plugin when prompted, then turn on the server and skill toggles. macOS additionally asks for Screen Recording and Accessibility permissions; on Windows keep the target app visible on the active desktop. After that, review app access in the Computer Use section of settings — approved apps appear in the Always allow list and can be removed there. Workspace admins can restrict which apps Computer Use may access and can switch the feature off entirely in requirements.toml. That plugin-based setup is how it ships today; what the 2026-04-16 launch added around it was: the new Codex Super App let multiple agents run on your Mac at once, each with its own Computer Use session, without disturbing you, and alongside Computer Use OpenAI shipped 90+ plugins (a tri-pack of skills + app integrations + MCP servers) covering GitHub PR review, multiple terminal tabs, SSH to a remote devbox (alpha), rich previews (PDF/spreadsheet/slides/docs), in-app browser, and more. It was the milestone of evolving Codex from a coding assistant into a general desktop work agent.
What Is Available Now — Regions and Model Choice
At launch on 2026-04-16 the feature excluded the EEA, the UK and Switzerland; the official update dated 2026-06-16 states Computer Use is available on macOS and Windows in those regions. As of 2026-10-19 the docs still describe it as available in supported regions, and administrators can restrict or disable it. The launch set was paid ChatGPT Plus/Pro/Business/Enterprise/Edu users, with limited Free/Go access at the 2026-04-16 launch; it was Mac-only until OpenAI added Windows on 2026-05-29. The Codex Super App vs Claude Code Desktop guidance of that round: long-horizon pure coding works on both; for desktop GUI actions (changing app settings / scraping a SaaS dashboard), Codex is your pick (Anthropic had no desktop-level background use yet); for raw SWE-Bench Pro performance, the GPT-5.3-Codex backend won. For hard tasks that lean on screenshots or visual judgment, the docs now point to GPT-6 Astra when it appears in your model selector.
FAQ
Is Computer Use safe? What permission controls exist?
When it shipped on 2026-04-16, Codex ran in an isolated window / virtual desktop, didn't take direct control of your full machine, and required permission confirmation; OpenAI has been iterating fast on the permission model since (similar to Claude's permission-mode). For production, isolate behind a sandbox user. Per the docs as of 2026-10-19: ChatGPT asks for your permission before it can use an app, and you can stop the task or take over your computer at any time. System permissions and app approvals are separate: Screen Recording and Accessibility on macOS decide whether it can see and operate the interface, while app approvals decide which apps it may use. The feature cannot automate terminal apps or ChatGPT itself, and it cannot authenticate as an administrator or approve security and privacy prompts for you. Your ChatGPT data controls apply to content processed through ChatGPT, including the screenshots Computer Use takes. Locked use on macOS has to be enabled in settings first, and then Computer Use can keep working after your Mac locks; Windows only runs in the foreground on the active desktop.
Can people in the EEA, the UK and Switzerland use it now?
Yes. The official update dated 2026-06-16 states that Computer Use is available on macOS and Windows in these regions, while the launch notes had excluded them. The docs now phrase availability as supported regions, and workspace admins can still turn the feature off. OpenAI earlier said 'soon' without a date; the EU usually needs additional compliance review (GDPR / DSA / AI Act). Watch the official OpenAI blog.
What part of this does QCode cover?
Codex can reach GPT models through a QCode API key, with no separate OpenAI account to register or manage (see /models for the callable ids). Computer Use itself is installed, authorized and used inside the ChatGPT desktop app; at the 2026-04-16 launch you installed the Mac desktop app yourself and routed model inference through QCode, while today the Computer Use plugin lives inside the ChatGPT desktop app and covers macOS and Windows. QCode covers the Codex CLI side: in config.toml set base_url to https://api.qcode.cc/openai and wire_api to responses, and Codex uses the OpenAI Responses protocol for GPT models — Claude does not go through it. Check /models for what is callable; billing is per token with per-model rates on qcode.cc/models. Not ready for a monthly plan? Pay-as-you-go Packs start at $4 for a one-time quota you use within 7 days.
Access the Codex Family via QCode
One QCode API key serves Claude Code, Codex and other tools; billed per token, with per-model rates on qcode.cc/models.
Start Your QCode Plan