Anthropic has updated Claude so you can autonomously perform tasks in your AI tools code and coworkusing your computer for you. According to the announcement, the new feature can be used to ‘open files, use the browser, and run development tools automatically, with no configuration required,’ even when you are away from your computer.
This is reflected in a video published by Anthropic that shows how, In a conversation with Claude on the mobile, he is asked to execute a series of tasks on the computer. The feature is based on autonomous capabilities that were introduced in the AI model Claude 3.5 Sonnet in 2024 and now more work is coming to Claude Code and Cowork.
This agentic capacity works by giving priority to supported service connectorsas Slack and applications of Google Workspacebut will still run tasks if there is no connector available, directly controlling your browser, mouse, keyboard and screen. Anthropic states that Claude ‘will always ask for your explicit permission’when necessary, before completing a task.
The new feature is now available for subscribers of Claude Pro and Max that they use macOSin what Anthropic calls a ‘research preview.’ This means that the system ‘it won’t always work perfectly’ and that, in complex tasks, will sometimes require ‘a second try’warns the company. Complete tasks by using the computer too ‘takes much longer and is more error prone’ than performing the same task using Connectors, he points out.
This type of direct computer control can also be started and managed remotely using the new function Dispatch by Claude Cowork, as long as the target computer remains on. You have to open Cowork on your mobile or desktop, enter the Dispatch section and when delegating a task Claude decides whether to execute it in Claude Code or in Cowork, depending on the type of work.
Bugs and security
Giving an admittedly imperfect and ‘error-prone’ AI tool the ability to scan your computer’s desktop ‘as needed’ can trigger security alarms for many users. Anthropic assures that it has implemented safeguards to prevent common riskssuch as prompt injection, and that will limit access to certain ‘banned’ applications by defaultfor example, ‘investment and trading platforms’ or cryptocurrency services.
The company also notes on a support page that the model is trained to avoid ‘risky operations’such as moving or investing money, modifying files, collecting facial images or entering ‘sensitive data’. But Anthropic also warns that these training-based safeguards ‘they are not perfect’ and ‘they are not absolute’which means that ‘Claude may occasionally act outside those limits’.
Anthropic also highlights that, when computer use is activated, Claude will be able to see anything that is visible on the screen, including ‘personal data, sensitive documents or private information’. For all these reasons, the company recommends ‘starting with the apps you trust and do not work with sensitive data‘.
Anthropic’s announcement comes just weeks after the release of Personal Computer of Perplexity, My Computer of Manus and NemoClaw of Nvidiawhich also allow your AI agents to take direct control of the desktop. All these movements come after the success, at the beginning of the year, of OpenClawwhich led to OpenAI to hire its creator, Peter Steinberger‘to power the next generation of personal agents’.