Anthropic launches Claude Duet cloud agent for Pro and Max users

A computer in the cloud

Anthropic recently announced Claude Duet, a new agent application designed to handle complex tasks across various connected apps. Unlike standard AI chat interfaces that require a user to be actively typing or viewing the screen, this tool operates on a virtual machine. This means the agent has its own dedicated environment in the cloud to execute commands.

Because it lives on a virtual machine with its own browser, Claude Duet does not require your physical laptop or smartphone to remain powered on or connected to the internet. It can process instructions and navigate websites independently. This architecture allows the AI to function as a background worker rather than just a reactive assistant.

The product is currently available as a research preview for users in the United States who subscribe to Claude Pro or Max. It is accessible as a standalone app for iOS and Android devices, as well as a web version at claude.ai. The setup process automatically carries over the memory, skills, and connected apps from a user's existing Claude account.

How the agent executes tasks

The system is built around the Claude Opus 5.5 model and is designed to interpret short messages from users to determine which apps to use. Once a user describes a goal, the agent works out the necessary steps and performs them in the background. It then sends a notification to the user when a task is finished or when it hits a point where it requires human approval.

To manage resources, the agent checks connected accounts every few hours rather than constantly. Anthropic explained that checking nonstop would use up a user's plan limits much faster. During the research preview, the background work and checks performed by Claude Duet count toward the same usage limits as the rest of the Claude service.

The tool is currently optimized for work-related tasks rather than personal errands. For example, the company noted that booking tickets is currently a weaker capability for the agent than professional tasks. Additionally, the agent cannot yet show the user its browser window as it navigates through different applications or websites.

Privacy and safety guardrails

Security is a primary focus for the new agent, as it handles logins and approvals within the user's own cloud computer. Anthropic states that this setup keeps sensitive details private. Credentials are managed in a way that ensures the agent never directly sees passwords or sign-in tokens, similar to how Claude Managed Agents function.

The company has also introduced a new classifier and safety system prompt to monitor every message. This system is specifically designed to reject requests to share high-tier sensitive information, such as government identification or credit card numbers. Other types of sensitive data, like personal finances or health information, are only permitted if the user explicitly tells the agent to send them.

Before the agent performs high-stakes actions, it must stop and wait for user permission. This includes sending messages, making payments, signing documents, or deleting any information. However, certain actions remain strictly off-limits even with user approval, such as trading securities or moving money between different accounts. Users also maintain the ability to disconnect any app they no longer want the agent to access.

Testing and limitations

Anthropic has put the tool through significant testing before the research preview. The company utilized both internal red teams and an external firm to perform penetration testing on Claude Duet. To ensure ongoing security, the product is also covered by Anthropic's public bug bounty program, which rewards researchers for finding vulnerabilities.

Despite these protections, the company provides a caveat regarding the agent's current capabilities. It may occasionally misread a specific thread or miscalibrate what a user considers to be urgent. Furthermore, because it checks in periodically, it might miss information that arrives in the interim between its check-ins with the user.

Larger, more complex projects like building a financial model or a presentation deck require more model turns and background work. These intensive tasks are expected to consume more of the user's plan limits. For now, the tool is restricted to Pro and Max users, and Team or Enterprise plans are not yet included in the rollout.