Anthropic announced on Monday, March 23, that Claude can now take control of a user’s computer to complete tasks, marking a significant step toward autonomous AI agents that navigate apps and operate software like humans.

Claude’s new computer-use capabilities allow it to handle tasks such as locating and sending files directly from a user’s hard drive. By sending a prompt in the Claude app on their phone, users can set the process in motion, starting with Claude looking for the right tools to complete the task via connectors with apps like Google Calendar or Slack.

If it can’t find the right tool or connector, Claude can manually perform the task by scrolling, clicking, typing or moving the cursor around to open files, access the web browser, and launch developer tools – as if it were using the keyboard and mouse like a human normally would.

Users will have to approve each app that Claude wants to use (if there is no connector available) but not every action that it has to perform. In the demo clip, Claude outlines its actions upfront and replies with a confirmation once the task is done. Claude’s computer-use update has been released in research preview to those who use MacOS devices, and have subscribed to its Claude Pro and Claude Max plans.

Users can delegate tasks to Claude using their phones by enabling the new Dispatch feature in Cowork. “Assign a task from your phone, turn your attention to something else, and come back to finished work on your computer,” Anthropic said in a post on X. In addition to routine computer tasks such as checking emails and exporting files, users can also start a live Claude Cowork or Claude Code session to generate code or run tests. More complex tasks may not work the first time as both the Dispatch and computer-use features are still in research preview, as per the company.

Anthropic’s latest Claude update comes as LLM-powered AI agents gain momentum, fueled in part by the popularity of an open-source, agentic AI framework called OpenClaw that lets developers orchestrate AI agents in structured flows.

Anthropic’s launch blitz of AI products and features is also in sharp focus lately as fears mount over automation-driven job losses. Earlier this year, the company’s release of a new suite of workplace automation tools and plugins spooked global technology markets and led to a sharp decline in Indian IT stocks as it threatened to disrupt the SaaS business model by automating several tasks that SaaS companies perform.

Anthropic first previewed computer-use by AI systems nearly two years ago. In 2024, the company announced an upgraded version of Claude 3.5 Sonnet model that could understand and interact with any desktop app by imitating keystrokes, button clicks, and mouse gestures.

While the ‘Computer Use’ API was released in open beta, it was reportedly prone to errors and consumed a lot of tokens because the capability relied heavily on LLM reasoning. The model took screenshots of the computer screen, inferred what different UI elements meant, and then decided what it should do.

To deliver computer-use by AI agents at enterprise scale, Anthropic acquired Vercept in February 2026. The AI startup specialised in creating tools for more complex agentic tasks, such as its product Vy, a computer-use agent in the cloud that could operate Apple MacBook remotely.

The price Anthropic paid to acquire the startup is not known. However, Vercept’s founders and some of its employees joined the Claude-maker as part of the arrangement. “It’s been less than four weeks since we joined, and with the team here behind us and joining forces, we just shipped our first product launch,” Kiana Ehsani, one of the co-founders, said in a post on X. “Four weeks down, many more to come. We’re just getting warmed up,” she added.

While Anthropic’s LLMs are among the highly advanced models out there, internal reasoning alone proved not to be enough to make AI agents more active. They also require an understanding of dynamic environments through grounded interactions. Vercept’s focus on bridging this gap made it an attractive acquisition for Anthropic.

It followed Anthropic’s earlier purchase of Bun, a developer-focused startup that built tools for running and orchestrating AI agents in production environments. The consolidation of these layers is key to Claude’s evolution from a conversational model into an execution platform.

Anthropic’s announcement on Monday could be seen as a response to OpenClaw’s rising popularity. The open-source platform has spawned an ecosystem of autonomous AI agents or ‘claws’, which can further spin up their own sub-agents to autonomously execute specialised tasks with access to local file systems and data.

Early setups of OpenClaw hosted on low-cost Mac Minis were focused on running Claude Code through WhatsApp on your phone. Thanks to its open-source nature, OpenClaw developer projects have since evolved into orchestrating AI agents to book flights, control smart home devices, and manage social media campaigns across multiple platforms and devices in a structured manner.

While setting up OpenClaw can feel more makeshift and experimental, Anthropic’s computer-use update could make for a more reliable, integrated solution that requires little technical expertise. The update, which is expected to support Windows devices soon, could eliminate the need for self-hosting on Mac Minis which have reportedly sold out in China amid the OpenClaw frenzy.

The requisite of giving OpenClaw access to all of a user’s data and systems in order for it to work as a true personal assistant, has further sparked concerns that these agents can go rogue and tamper with or delete valuable files. These security tradeoffs have resulted in the need for safety-focused forks and alternatives. Last week, Nvidia debuted NemoClaw, its software toolkit for easily setting up and installing OpenClaw with some security controls.

To be sure, Claude’s computer-use update poses security risks as well. Anthropic has said that it has implemented safeguards against threats such as prompt injection attacks. But it has also warned that the update is new, and Claude should not be given access to apps that handle sensitive data.