Waitlist only, and usage still hits the plan

Claude's product page calls Claude Duet a research preview now. The judgment first: it is powered by Opus 5.5, and Pro and Max plans in the United States can join a waitlist. Team and Enterprise plans are not included yet, the page says. During the preview it shares the plan's usage limits with the rest of Claude, and background checks count toward those limits. Larger tasks, such as building a financial model or a deck, take more model turns and more background work. Shared usage is the rule now, not a bill that has already been spent.

The page describes a proactive agent that comes to you. The Claude app waits until you open a chat. Duet tells you each morning what needs attention, takes tasks you hand it by message, and reaches out when one is done or needs approval. It uses your existing Claude account, including memory, skills, and connected apps. More than 6,000 connectors and plugins is the number on the page. This page does not list them, so the number cannot be read as a catalog that has already been checked item by item.

It works on its own computer in the cloud, using apps you connect, and messages you when it is done or needs a decision. Text or voice both work. The standalone app is for iOS and Android. On a computer you open the web at claude.ai. For sensitive operations, such as sending messages or paying, the page says it asks first. Approvals, logins, and safety checks run in your own cloud computer. You can see what it has done and what it is doing, and you can steer it. Those are the product page's description of control, not a published incident count.

The capabilities listed include a morning brief, flagging what needs you in the inbox and calendar, drafting replies for you to approve, finding unpaid invoices and drafting reminders, preparing payroll for review, getting contracts ready to send, and putting together a brief, a deck, or a dashboard. The list is a product description. The page does not give a success rate for these tasks, and it does not give a count of messages sent by mistake or payments made by mistake.

What is not ready yet is written on the page itself. The research preview may misread a thread, miscalibrate what is urgent, and miss something that arrives between check-ins. Those three sentences are the limits. They are not a vague promise of later improvement. Higher on the same page is an interface story about a shop, hiring, and a fundraise, with dollar figures, cities, and names. Those figures sit in the sample interface. This page does not present them as a real customer's results, and this piece does not treat them as business that has already happened.