Work for many people now involves tracking a steady stream of messages, invoices, calendar items, and small administrative steps that accumulate across different applications.
Keeping up with them often means shifting attention repeatedly throughout the day, and tools that only respond when asked can leave gaps where something sits unaddressed until someone notices it.
To address this issue, Anthropic opened a research preview of 'Claude Duet,' an application in which Claude operates as a proactive agent rather than waiting for a user to open a conversation.
Claude Duet is built around Claude Opus 5.5.
It draws on the memory, skills, and connected applications already set up in a user’s Claude account, so the initial setup involves relatively little additional configuration. Users interact with it by typing or sending voice messages in a persistent thread, and they can open separate channels for different topics that still share the same underlying memory.
The agent runs on its own computer in the cloud, which means it can continue working on tasks after the user closes the app on a phone or computer.
Each morning it prepares a short brief summarizing the most important items across connected applications such as email and calendars. During the day it checks those accounts at intervals and sends a message only when it judges that something requires attention.
Users can also assign it work directly, ranging from scheduling a meeting to assembling a dashboard, preparing a financial forecast, or drafting materials such as a brief or slide deck.
When it finishes, it returns the result for review. Background monitoring and larger tasks draw on the same usage limits as the rest of Claude under Pro and Max plans, though they consume capacity differently from ordinary conversations.
For actions that affect other people or involve money or documents, Claude Duet is designed to pause and request approval before sending a message, making a payment, signing a document, or deleting anything.
Certain categories of action remain blocked even with approval, such as trading securities or transferring money between accounts. It does not directly see passwords or sign-in tokens.
A dedicated classifier and safety prompt review messages to limit the sharing of sensitive information, rejecting the strictest categories such as credit card numbers or government identification numbers unless the user has given explicit instructions for other types of sensitive material. The system also incorporates earlier work on agent safeguards, including protections against hidden instructions.
Access during the research preview is limited to subscribers on Claude Pro and Max plans in the U.S. who join a waitlist. Invitations are sent in waves.
The application is available as standalone apps for iOS and Android and can also be opened through claude.ai. Team and Enterprise plans are not included yet.
Anthropic has indicated that features may change as feedback comes in from early users, and the agent can still misread a thread or misjudge what is urgent. Items that arrive between its periodic checks can also be missed.
It's worth noting that the agent can work with more than 6,000 connectors and plugins available through the Claude Marketplace.
Examples include checking QuickBooks or Stripe records for unpaid invoices and preparing reminders, assisting with payroll preparation in Gusto, drafting follow-ups that draw on HubSpot history, preparing agreements for review in Docusign, or setting up elements of a Shopify store. Purchases can be handled through Stripe's Link or Shopify's Shop Pay after the user approves, with one-time cards used so that full payment details are not shared.
Users can see a record of what the agent has done and is currently doing, and they can steer it or disconnect any connected application.
Anthropic has described the current version as early and has noted that it is stronger on work-related tasks than on purely personal ones such as booking tickets. Users cannot watch or take over the cloud browser it uses, and it cannot place phone calls or send SMS messages. Microsoft 365 is not supported at this stage.
Since Claude Duet is still in preview, it is separate from the main Claude application, with any future integration depending on how people use it.
Early internal users at Anthropic have mentioned using it to discuss task lists while walking or to clear administrative items before time with family, and the company has invited feedback from small business owners and entrepreneurs who receive access.
The release comes amid similar efforts elsewhere to build agents that monitor applications and act with less constant prompting.
