Anthropic wants its AI to stop waiting for prompts. Claude Duet, now in a waitlisted research preview, runs tasks while you are away and pings you when something needs a decision.
Access starts with paying Claude subscribers in the United States (Pro, Max and Ultra), with the wider rollout left unstated. It is bundled into those plans at no extra cost, but its work counts against your normal Claude usage allowance.
A helper with its own computer
According to Anthropic’s announcement, Duet runs on Claude Opus 5.5 inside a cloud machine with a browser of its own, so your devices need not be switched on. You message it by text or voice in a single ongoing thread. Each morning it compiles a brief from the apps you have linked, and it checks those apps every few hours, not continuously, to save your allowance.
The company lists everyday jobs: booking meetings, reminding customers about unpaid invoices in QuickBooks or Stripe, preparing payroll in Gusto, writing HubSpot follow-ups, or assembling a sales forecast. Your Claude memory, skills and connected apps carry over, and it can use any of the 6,000-plus connectors in the Claude Marketplace.
Approval before anything irreversible
Duet shows you what it intends to do and waits before it sends a message, pays, signs or deletes. Some actions are blocked outright, including buying or selling securities and shifting money between your accounts. Anthropic also says the agent never directly sees passwords or sign-in tokens.
To limit leaks, a new classifier refuses to pass on the most sensitive data such as card numbers and government ID. Health and personal finance details are shared only when you tell it to.
Rough edges
Anthropic concedes the preview is tuned for work, so errands such as buying tickets are weaker, and you cannot watch its browser yet. Big jobs like financial models or decks eat more of your allowance. Duet is a separate app on iOS and Android, and also on claude.ai, rather than a mode inside the main Claude app.
Waitlisted users will get an email when their turn comes. Anthropic says it is still adjusting how often the agent should interrupt.
