Anthropic takes on Grok Bot, Meta's Muse, and OpenAI's Dots with its always-on agent "Claude Duet"
Key Points
- Anthropic has released an early preview of "Claude Duet," an app where an AI agent works on tasks for users in the background.
- The app runs on cloud computers and connects to outside services such as Gmail, Slack, and Stripe. It can schedule meetings and prepare bills, among other tasks.
- The app always asks users for permission before making payments or sending messages. Bank transfers are not allowed at all.
Anthropic has responded to Grok Bot, Meta's Muse, and OpenAI's Dots by releasing an early research preview of Claude Duet. This separate app uses Claude as an AI agent that keeps working on tasks while users are away. It asks for permission before taking sensitive actions.
Anthropic is offering Claude Duet to professionals and small business owners who have "more on their plates than they can track." The app works on iOS, Android, and the web at claude.ai. It is built using Claude Opus 5.5. Pro and Max subscribers in the US can apply for access through a waitlist. Team and Enterprise subscribers cannot use it yet.
Anthropic says the main difference between Duet and the regular Claude app is how they work with users. In the regular Claude app, users start a chat. Duet reaches out to them instead. Every morning, the agent sends a short summary of the most important updates from connected apps. After that, it only contacts users when something needs their attention. It uses the saved memories, skills, and app connections from their existing Claude account.
Users give Duet tasks through text or voice messages in one ongoing conversation. Anthropic says it can handle many kinds of work, from scheduling meetings to creating dashboards, financial forecasts, prototypes, and web apps. The agent works on a virtual machine, which is its own computer in the cloud with a web browser. This means the user's laptop or phone does not need to stay switched on.
Duet connects to services such as Gmail, Slack, Stripe, Docusign, and Shopify through more than 6,000 connectors and plugins available in the Claude Marketplace. For example, it can find unpaid bills and write payment reminders, or prepare payroll. When making purchases, it uses Stripe Link with a virtual card. This keeps the user's real card number private.
Several competitors have released similar products in recent months. SpaceXAI opened the Grok Bot beta in August. In September, Meta launched Muse, and OpenAI launched its Dots. All these companies are betting on personal AI agents that can handle tasks independently and take action without waiting for users to start every step.
Users keep control of anything risky
Anthropic says Duet can read information, write drafts, and prepare tasks on its own. But before it sends a message, makes a payment, signs a document, or deletes anything, it shows users exactly what it plans to do and waits for their approval. It cannot trade stocks or other securities, or move money between accounts, even if users approve.
Technical safety measures provide extra protection. Anthropic says the agent never directly sees passwords or login tokens. Instead, it handles login details using the design of Claude Managed Agents. To protect against hidden instructions, called prompt injections, Anthropic uses safety methods developed through its research on browser use. Internal security teams and an outside company also tested Duet with attacks to look for weaknesses.
Anthropic also created a new system to check Duet's messages for sensitive information that might be shared by mistake. This system works with special safety instructions built into the agent. Together, they are designed to stop requests that would reveal credit card numbers or identity details. Health information and personal financial information can only be shared when the user clearly gives permission.
Anthropic admits Duet still makes mistakes
Anthropic admits that this early version can misunderstand conversations, get the importance of a message wrong, or miss new information that arrives between checks. It is also less effective at personal tasks, such as booking tickets, because its main features are designed for work.
Duet checks connected accounts only every few hours. Anthropic says this is because of usage limits. Checking constantly would use up subscription allowances much faster. Duet shares these limits with other Claude products, and its background checks count toward them too.
Anthropic plans to use feedback from testers to improve when the agent contacts users. It is still unclear how Duet will work alongside the regular Claude app. For now, the two run separately. The experiences of early users will help Anthropic decide how to bring them together.