Alerts before being asked, works while the user is away
Anthropic said on Claude's official X account on Oct. 21 (Oct. 22 Japan time) that it had begun a research preview of a new app, "Claude Duet." It is an AI agent that alerts users to things they should pay attention to before they ask and keeps working while they are away.
Duet regularly checks connected email and work services and creates a morning summary. Tasks it is given are carried out on a computer in the cloud with its own browser. The company cited examples such as scheduling meetings, data dashboards and financial forecasts, and building prototypes and web apps. It carries over Claude's memory, skills and connection settings, and can use more than 6,000 connections, including Gmail, Slack, HubSpot and Docusign.
Always asks before sending or paying
Duet carries out tasks such as reading, drafting and preparing on its own. However, before sending messages, making payments, signing documents or deleting, it shows the details and waits for the user's approval. Purchases go through payments from U.S.-based Stripe or Canada's Shopify, require approval each time and use a single-use card number. It cannot trade stocks or other securities or transfer money between accounts.
It uses the company's top model, "Claude Opus 5.5," to make judgments. Anthropic's Mike Krieger said Opus 5.5 is the first model with the judgment to be proactive in the right way.
For Pro and Max in the U.S., usage limits shared
It is available to users of the paid "Pro" and "Max" plans in the United States. Users sign up for a waitlist on the product page, and invitations are sent out in stages. It can be used in apps for iPhone and Android and on the web, and during the preview it is separate from the regular Claude app. It will later be expanded to team and enterprise plans.
There is no extra charge, but it shares the 5-hour and weekly usage limits with regular Claude, and work it does in the background also uses those limits. New users get one reset of their usage limits that can be used through Nov. 30. It checks connected services periodically rather than constantly because doing so uses these limits. Anthropic says it may misread exchanges or misjudge whether something is urgent.
OpenAI's "dots" and SpaceXAI's "Grok Bot" are also always-on agents that act ahead of users, and competition is intensifying.
