Anthropic moved Claude's computer control to hidden windows on macOS: the agent clicks and enters text while the user continues to work.
The computer control had so far suffered from a rectibitory defect: to act, the agent took control of the cursor and keyboard, condemning the user to watch the mouse move on its own. The update published by Anthropic on September 2 removes this obstacle on macOS by having Claude work in masked background windows. The agent clicks, grabs text and opens applications without confiscating the devices, which turns a demonstration into a really usable tool while doing something else.
The capacity is available in Claude Code and Claude Cowork, reserved for Pro and Max subscriptions, and disabled by default: it must be explicitly allowed in the desktop application settings. Claude also favours application integrations when they exist — Slack, calendar — and only descends to the level of interface control when no other path is available.
Anthropic accompanies the functionality of a warning that deserves to be read: the safeguards are not absolute, the execution has no sandbox, and the company advises against pointing the agent to financial, legal or medical applications. An agent who manipulates a graphical interface without direct visual monitoring remains an agent whose errors are no longer seen in real time.
Read the full article on the original source
Go to the source →