Anthropic opens background computer use in beta to Pro and Max, with its own warnings

Anthropic said on 2 September 2026 that Claude can now use a person's computer in the background inside Claude Cowork and Claude Code: the user gives it something to do and it "clicks, types, and opens apps" while the user works on something else. The post passed 1.3 million views by 3 September.

Where it runs, and what it asks first

The company's two sources disagree: the post names only macOS, the support document lists macOS and Windows. On macOS 15 and later, that document says, Claude runs in background windows, does not take the cursor or keyboard, and usually waits while the user types. Background running is the default there; Settings > General also offers "Full control", handing Claude the screen directly.

According to the same document, Claude asks before reaching any application, and investment and trading platforms and cryptocurrency apps are blocked by default; users can extend that block list, and requests for apps on it are refused automatically. It says the system scans for prompt injection while Claude works and the user can stop it at any time. The computer must stay on with the desktop app open, complex multi-step tasks sometimes need a second try, screen work is slower than the connectors the company prefers, and Team and Enterprise plans lack it at this time.

The warnings come from the company itself

The support page states there is no sandbox: Claude interacts directly with the desktop, apps and browser. It calls the capability early-stage, says attacks keep developing and no safeguards are perfect, and tells users to be careful. The protections in Claude's training and instructions, it adds, are not absolute and no substitute for blocking sensitive apps; it strongly advises against financial accounts, legal contracts, medical information and apps holding other people's data, and against permitting banking, health or government apps. One app can also reach another: an emailed link can open Chrome even when Chrome was never permitted, which the company cannot prevent. It recommends simple tasks and narrow prompts.

Open note: GlobalFeed AI's publishing engine runs on Anthropic's Claude model. It uses only the company's announcement and support document. Cowork gained a built-in browser on 27 August 2026; this reaches the whole desktop.