Anthropic just unveiled a significant upgrade to its Claude AI assistant by introducing a feature called “Computer Use.” The announcement came on March 23, when the company opened a research preview for Claude Pro and Max subscribers.
According to Anthropic’s official release, the feature allows Claude to directly control a Mac. It can move the cursor, click buttons, type text, open applications and navigate the operating system just as a human would.
This transforms Claude from a passive conversational tool into an active agent capable of executing complex tasks across different software.
How Claude’s Computer Use Works
To understand how this works, it helps to know that Claude interacts visually with the computer.
First, it captures screenshots of the user’s display and analyzes the interface to identify windows, buttons, menus, and text fields. Then, after determining the necessary steps, Claude generates mouse movements and keyboard inputs to carry them out.
Moreover, when Claude lacks a direct API connection for an application like Slack or Google Calendar, it falls back to controlling the computer like a human by using the screen to navigate.
For example, it can open files, use the browser, and run development tools automatically without requiring manual integration. Once the task finishes, it reports the outcome and adapts to unexpected pop-ups or errors in real time.
What This Means for Productivity
This capability turns Claude into a proper digital assistant and it significantly improves productivity.
For instance, a user can ask it to extract data from a spreadsheet, build a presentation with that information, compose an email summary and send it without touching the mouse or keyboard.
Consequently, professionals can delegate repetitive workflows to Claude and focus on higher-level decision-making. In addition, Anthropic launched a companion feature called Dispatch around the same time.
According to reports, Dispatch allows users to assign tasks from their iPhone and then return to finished work on their desktop. Users can send voice or text instructions from their phone and Claude executes the work on their Mac while they are away.
Security and Privacy Considerations with Claude
Despite the impressive capabilities, security remains a significant concern. When Claude controls a Mac, it must view screenshots of the screen which may contain sensitive information.
However, Anthropic has implemented multiple safeguards. For example, the system uses a permission-first approach, requesting user approval before accessing new applications.
Claude prioritizes native integrations with services like Slack and Google Calendar before resorting to screen control. The company has also disabled access to sensitive apps such as investment and cryptocurrency platforms by default.
Nevertheless, researchers have identified vulnerabilities. A Security firm reported in mid-March that Claude can be manipulated through “indirect prompt injection,” where malicious instructions hidden in files could potentially cause the AI to upload sensitive data. This vulnerability was known before the feature launched but Anthropic still proceeded with the release.
Additionally, early testing shows that for non-trivial tasks, the success rate is around 50% and complex multi-step workflows often stall or produce incomplete results.
For these reasons, Anthropic explicitly warns users not to use Computer Use with applications that handle sensitive data.
How to Use Computer Use Safely
Given these limitations, users should take specific Safety precautions.
First, start by testing Computer Use in isolated environments or virtual machines before deploying it on primary workstations. Make sure to enable action-confirmation prompts and usage logging to maintain visibility into Claude’s activities.
In addition, keep macOS and all applications updated to receive the latest security fixes, and consider using a dedicated user account without administrative privileges.
Make sure to regularly review activity logs to spot unusual patterns and remember that Claude requires the desktop app to remain open and the computer to stay awake during operation.
Finally, because Anthropic acknowledges that “threats are constantly evolving” and the feature remains in research preview, users should approach Computer Use as a promising but imperfect tool that still requires human oversight.
