Cursor can now control your desktop apps. How it compares with Codex, Claude and Copilot

What Cursor announced

On October 23, 2026, Cursor posted: “Cursor can now control your computer. Agents can drive desktop apps like the iOS Simulator to test the features they build.” Its follow-up is most of what it has said so far:

• “Computer Use lets agents click and type across any app on your computer, not just the browser.”
• “Cursor uses it on its own when a task needs it.”
• “You can also ask for it by @-mentioning Computer Use in your prompt.”

What’s new is where it runs. Cursor’s cloud agents have had computers of their own since February 24: each runs in “its own isolated VM with a full desktop environment,” as Cursor’s docs put it, apart from yours. Now the agent in Cursor’s app can use the apps on your computer, so the agent that wrote a feature can open the app it built and check it.

How to turn it on

1. Update Cursor. Cursor’s follow-up links to its download page and names no version.
2. Give the agent a task that needs an app, or @-mention Computer Use in the prompt to ask for it.
3. On a Mac, grant two permissions. Cursor’s docs for self-hosted workers say computer use runs through a helper app called Cursor Computer Use, which needs Accessibility for the mouse and keyboard and Screen Recording for screenshots. A guide by developer Flavio Copes says the desktop app uses the same helper: grant both to Cursor Computer Use in System Settings > Privacy & Security, not to your terminal.

The iOS Simulator comes with Xcode, so install Xcode and its simulators before asking an agent to test an iPhone app.

The same docs cover computer use for self-hosted workers, Cursor’s cloud agents running on machines you manage: start one with agent worker --computer-use start, on macOS or Linux. On Enterprise teams, a Computer Use switch under MCP Configuration in team settings turns it off for those workers. It’s on by default.

What Cursor has said since

On October 24, a user reported on Cursor’s forum that in the Windows app, version 3.27.12, the agent clicked and typed in Chrome without asking, though the always-allowed apps list under Settings > Agents > Computer Use was empty. A Cursor staff member replied the same day:

• It runs on Windows. The report and the reply are both about Cursor’s Windows app.
• On Windows, it doesn’t ask app by app. “On Windows, there just isn’t a separate prompt for each app right now. Each Computer Use action is approved the same way as any other MCP tool call, based on your run mode.” In Auto-review, Computer Use calls “are checked by a classifier, and anything it allows runs without a prompt.”
• To be asked every time, open Settings > Agents > Approvals & Execution and switch the run mode to Ask Every Time.
• To turn it off, disable the Computer Use entry in the MCP servers list. “There isn’t a separate Computer Use toggle in Agents settings yet.”

The staff member agreed that the Computer Use settings, which say Cursor will ask, are “confusing” on Windows, and passed on the request for a separate switch. The user replied that the Computer Use server had arrived switched on.

What Cursor still hasn’t said

• Which plans include it, and what it costs. No price was announced, and nothing says how the screenshots an agent takes count toward your usage.
• How it asks on a Mac. Its settings promise a prompt when Cursor wants to use an app, but Cursor hasn’t documented it, or a key that stops an agent in the middle of an action.
• Whether it works in the Linux app. Cursor’s docs cover Linux only for self-hosted workers.

As of October 27, Cursor’s docs still have no page for the desktop feature, and the newest entry in its changelog, the Agent SDK, dated October 20, doesn’t mention it.

How Codex, Claude Code and Copilot compare

Cursor joins three agents that already document their own versions:

	Cursor	Codex	Claude Code	GitHub Copilot
Where	Cursor’s app, cloud agents, self-hosted workers	ChatGPT desktop app, in Work and Codex	Claude Desktop, and the CLI	Copilot app and Copilot CLI
Systems	Windows, by Cursor staff; macOS, by a developer’s guide	macOS, Windows	Desktop: macOS 14 or later, Windows. CLI: macOS	macOS, Windows
Plans	Not stated	Plus, Pro, Business, Enterprise, Edu and API key, in some regions	Pro and Max, as a research preview	Public preview, plans not listed
Turn it on	Update Cursor	Install the Computer Use plugin	Settings > This computer > System, or /mcp in the CLI	Settings > Computer Use, or /computer on

The sources are OpenAI’s Computer Use guide and Codex pricing, Anthropic’s docs for the CLI and the desktop app, and GitHub’s changelog and docs. A few differences matter in practice:

• Working while you work. On a Mac, Codex and Claude’s desktop app can use the apps you’ve approved in the background while you keep working. On Windows, OpenAI says Codex “can’t operate in the background,” so it takes over the pointer and keyboard, and it suggests a Windows virtual machine for long tasks.
• What’s off-limits. Codex can’t automate terminal apps or ChatGPT itself, or approve security prompts. Claude gets view-only access to browsers and trading platforms, click-only access to terminals and IDEs, and full control of everything else.
• iOS apps without screen control. On a Mac, Claude Code Desktop opens an iOS Simulator pane that drives the simulator directly, so it “doesn’t need computer use and never takes over your screen.” It’s in public beta on Pro, Max, Team and Enterprise, except Enterprise organizations with a HIPAA configuration.
• Not everywhere the agent runs. Claude Code’s CLI version is macOS-only, needs an interactive session, and isn’t available through Amazon Bedrock, Google Cloud or Microsoft Foundry. OpenAI documents Codex’s version only in the ChatGPT desktop app.

Keeping it safe

Controlling your desktop is the broadest thing an agent can do. Anthropic’s docs put it plainly: “Unlike the sandboxed Bash tool, computer use runs on your actual desktop with access to the apps you approve.” OpenAI warns that when Codex uses your browser, “it can interact with pages where you’re already signed in.”

• Test a development build with test data, not the copy of the app you use for real work. Anthropic says not to sign in to real accounts on a simulator Claude uses.
• Approve apps one at a time. Claude’s approvals last for the session. Codex and Copilot can remember an app with “Always allow,” which you can remove in their settings. In Cursor on Windows, there’s no prompt per app yet, so choose the run mode with that in mind.
• Know how to stop it. Esc stops it in Claude Code’s CLI. In Copilot, press Esc twice in the CLI, or click Stop or press Esc in the app. Cursor hasn’t documented a stop key; turning off its Computer Use entry in the MCP servers list keeps the agent from using it at all.
• Use the precise tool first. Anthropic calls computer use “the broadest and slowest” way to reach an app, so Claude tries an MCP server, the shell or Claude in Chrome before it. OpenAI’s advice for web apps you’re building is the same: “use the built-in browser first.”

When a browser is enough

If you’re building a website, you may not need desktop control at all. Enjoy puts a browser beside each conversation, and the agent you’re running, Claude Code, Codex or another lab’s, can start your site there, read the page, and click and fill in forms to check its own work. On a Mac it can also take screenshots and type. It’s the same page you see, so you can watch it work, and it stays inside that browser.

Enjoy runs each agent on the plan you already have, and Enjoy itself is a monthly subscription, with the first seven days free (pricing). It doesn’t drive desktop apps or the iOS Simulator; for those, use the agents’ own computer use above.

We’ll update this memo when Cursor documents the desktop feature.