Anthropic Allows Claude to Control Computers in Background Windows
Anthropic has updated the Claude desktop application, enabling Computer Use to run in the background within Claude Cowork and Claude Code: it can click, type, and open applications to complete tasks while users can do other things simultaneously. The company states that it no longer takes over the pointer and keyboard, typically waiting during input.
The background window is the default mode for macOS 15 and above. The first time a task requires full screen, each session will need permission once; if users want it to take over the screen directly, they can change the application access to full control in the desktop application settings. Computer Use is currently in open testing for Pro and Max plans on macOS and Windows desktop, but must be manually enabled in settings, with an official prompt to read the safety instructions first.
The execution order starts with connectors like Gmail, Google Drive, Microsoft 365, and Slack, followed by built-in or Chrome browsers, with screen-level operations only falling to internal backends or specialized tools without interfaces. The desktop must remain awake and Claude Desktop open for the computer to access local files and applications; pure cloud sessions can continue after closing the lid, but applications cannot be closed when accessing local resources.
Dispatch allows tasks to be sent from a mobile phone to a desktop computer, using the same set of agents to access local files, plugins, and applications. Complex multi-step processes may need to be tried a second time, as screen operations are slower than direct integration. Linux desktop testing does not include Computer Use. The functionality carries risks such as prompt injection, and the system will scan activations for anomalies.
Earlier, OpenAI launched background computer control on Mac. Anthropic has shifted the difference from "can it click the mouse" to "will it compete with you for the same mouse and keyboard." Computer Use is not a brand new capability; it appeared with the model research preview in 2024 and is now integrated into the knowledge work-focused Cowork and programming-focused Claude Code.
In market mechanics, users are buying time slices for agents to take over their desktops, while the old interaction that requires watching the screen as the model clicks around is being sold. Funds remain in Pro/Max subscriptions rather than selling remote desktop access per use. Beneficiaries are Anthropic's desktop stickiness and agent duration; those under pressure are OpenAI, which first created the background mode, and assistants that only provide chat boxes without local application control. The connectors prioritize explaining that what they really want to collect is workflow permissions, with screen clicks being just a patch.
Note: The latest Claude Desktop is required, with the path being Settings—General—Enable Computer Use. Applications without connectors rely on visible interfaces, and any change in the interface may lead the agent to go wrong.
Source: Public Information
ABAB AI Insight
The background window transforms the agent from a "demonstration feature" to a "second operator." The Computer Use that competes for the mouse can only serve as a performance; not competing for the mouse allows it to enter the workday. Anthropic places the default in a standalone window on macOS 15, acknowledging that the failure point of the previous version was not in recognizing buttons, but in competing with users for the same input device. The three-tier fallback from connectors—browsers—screen turns the desktop back into a legacy system adapter without APIs.
The capital path is to cover the time agents stay locally with subscriptions. Cowork opens the agent core of Claude Code to non-programmers, while Dispatch places triggers on mobile phones, turning desktop machines into always-on workers. The Pro/Max threshold locks high-risk permissions within paying users, also limiting the first wave of security incidents to those willing to participate in testing. Competing against OpenAI's background mode, the battle is over who becomes the default second cursor in the operating system first.
Analogous objects are early RPA and remote desktop scripts, as well as Microsoft's permission battle embedding Copilot into Windows. The industry phase shifts from chat completing answers to agents completing clicks. Whoever can finish internal systems without APIs without freezing the foreground will capture that endless integration work in enterprises.
The structural change is technological substitution. The mechanism is that keyboard and mouse interfaces are more universal and also more dangerous than application interfaces; the background window reduces friction, turning "allowing a model to operate my computer" from a performance permission into a parallel habit. Permission pop-ups still exist, but the default has shifted from taking over the screen to opening a new window; once the habit is formed, connector permissions will be more valuable than chat records.
ABAB News · Cognitive Laws
Not competing for keyboard and mouse, agents are worthy of entering the workday
Software without interfaces will ultimately be swallowed by screen-level operations
Once the second window becomes the default, permissions are more valuable than conversations