Anthropic PBC wants users to give Claude the keys to their computer, announcing new “computer use” functionality for its artificial intelligence assistant so it can now click, scroll and navigate through web pages and applications to complete tasks on behalf of users.

The feature was launched today as a research preview for Claude Pro and Max subscribers, and it notably pairs with Dispatch, which is a mobile tool that debuted last week. With Dispatch, users can assign tasks to Claude from their smartphone, and have it complete those assignments using their computer.

In a blog post, Anthropic explained that when Claude is given a task to complete, it will look to see if it has the right integrations to perform it using tools such as Google Calendar and Slack. But if it doesn’t have the right connector, it will fall back onto controlling the computer like a human does, using the screen to navigate. It can open files and documents, operate a web browser and run development tools autonomously.

The update should be especially useful for developers. Anthropic said Claude can make changes within an integrated development environment, submit pull requests, run tests and more besides, leaving the human user to focus on other tasks. The system is based on a permission-first approach for safety reasons, Anthropic said. What that means is that Claude will request access before it touches a new application, and users have the ability to stop it at any time.

Not yet ready for prime time

Anthropic isn’t trying to oversell the feature, admitting it remains a work in progress. “Computer use is still early compared to Claude’s ability to code or interact with text,” the company explained. “Claude can make mistakes, and while we continue to improve our safeguards, threats are constantly evolving.”

One obvious limitation is that it currently only works with Mac computers, so anyone with a Windows or Linux PC will have to sit this one out for now. The idea is to gain insights from users on where it needs more attention, so the company can create a more powerful tool. As a result, some complex tasks might require multiple attempts before Claude gets them right, and its screen-based operations are a lot slower than direct application programming interface integrations.

There are some security concerns too. Giving Claude the keys to a computer may be very convenient, but it can also leave Macs vulnerable to some types of cyberattacks. Anthropic hasn’t said so, but it’s likely Claude’s computer use capabilities are enabled via OpenClaw, which has built an ecosystem of “Claws” that enable AI models to use third-party software.

But these Claws can be vulnerable to hijacking, potentially enabling malicious actors to take control of whatever data and systems the chatbot has access to. Anthropic said it has implemented safeguards to protect against risks such as prompt injection attacks, which is a popular attack vector for AI tools that have computer access. It will automatically scan for other vulnerabilities too, but even so, Anthropic says to avoid letting Claude have access to sensitive data during the research preview.

The release comes at a busy time for Anthropic, which has unveiled a flurry of updates in recent weeks while simultaneously squaring off against the Trump administration over Claude’s restrictions. Last month it shipped major updates in the shape of Claude Opus 4.6 and Claude Sonnet 4.6, with both models aimed at complex agentic workflows – just the kind of tasks that would benefit from using a computer.

Computer control is a key goal for AI developers, because it’s essential for the transition from intelligent conversation partners to autonomous assistants that do more than just provide information. Rivals including OpenAI Group PBC and Google LLC are also racing to deliver systems that can safely operate computers.

Image: Anthropic