Claude Duet works while you’re away and asks before acting

Anthropic on October 21 opened a waitlist for Claude Duet, a standalone app where Claude watches your connected accounts, brings you what needs attention and keeps working on a virtual machine after your devices are off. Before it sends a message, pays, signs or deletes anything, it has to show you the action and wait.

Anthropic’s own post on Threads keeps it short: a research preview for Pro and Max plans in the US, built around Claude Opus 5.5, which “gives it the judgment to decide what needs you, what can wait, and when to check with you first.” Memory, skills and connected apps carry over from your Claude account, “so there’s little to set up.” Most of the operating detail comes from Testing Catalog’s write-up of the product page and launch materials.

What it does between your check-ins

Testing Catalog says Duet is aimed at professionals and small business owners with crowded workloads. Each morning it sends a short brief built from your connected apps. After that it checks those accounts every few hours and messages you when it decides something needs you. The tasks it describes run from scheduling meetings to dashboards, financial or sales forecasts, prototypes and web apps. Quick requests and weeklong projects both happen in one ongoing thread, by text or voice. The work itself runs on a virtual machine with its own browser, so nothing stops when your laptop or phone goes offline.

Some of this has been seen before. A Morning Brief showed up in August as a scheduled task inside Cowork, and on September 16 Anthropic began rolling Cowork into every Claude conversation, carrying over chats, Cowork tasks, projects, connectors and skills, according to its release notes. What’s new in Duet is the direction. You don’t open a task and wait for the result. The agent decides when to message you.

Mike Krieger said on X that he had used it “for a few weeks, at work and at home,” and that “it drafted questions for a fireside chat, watched Slack over the weekend.” His framing of the launch: “It’s our take on giving everyone a trusted partner with your work.”

Where it has to stop and ask

Anthropic’s post says Duet “never directly sees your passwords,” that a new classifier guards against sharing sensitive information, and that “the important decisions stay with you.” It promises more on its safety approach soon.

Testing Catalog fills in the rules. Duet can read, draft and prepare work on its own, but it must show the proposed action and wait for you before it sends a message, makes a payment, signs a document or deletes anything. Securities trading and transfers between accounts are off the table entirely. The credential handling follows the design of Claude Managed Agents, where the model doing the reasoning stays separated from the sign-in secrets the tools use. Action controls draw on the classifier already used for Claude and Claude Code.

Sensitive data sits under a separate layer. A dedicated classifier plus a safety prompt blocks Duet from sharing credit card numbers and government identification. Other sensitive material, such as health or personal finance information, needs your explicit instruction before Duet will pass it on. Anthropic says red teams and an outside security firm tested the system, and Duet falls under its bug bounty program, per Testing Catalog.

Paying with Link and Shop Pay

Purchases are where the approval rule gets its first real test. Testing Catalog reports that Duet can make business purchases through Stripe’s Link or Shopify’s Shop Pay, with approval steps and virtual-card handling on the Link side. Shopify said Shop Pay works for every purchase in Duet, adding: “It’s the first time we’re opening Shop Pay in agents to stores not on Shopify.”

Duet can also reach more than 6,000 connectors and plugins in the Claude Marketplace. The launch use cases name Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva and Shopify. Support for 1Password logins and richer visual editing for Shopify is planned but not shipped.

Before launch, Testing Catalog posted an early look at the Stripe Link shopping flow, noted hashtags under the agent that it read as a possible hint of Slack-style threaded conversations, and reported Duet’s internal name: “aibo”.

Who gets in, and what it costs

The preview runs as a standalone app on iOS and Android and on the Claude website, separate from the main Claude app for now. It’s included in Pro and Max with no extra fee, and it shares those plans’ existing usage limits, per Testing Catalog. The background checks every few hours count against those limits too, and bigger projects take more model turns, so a busy Duet eats into the same budget as your chats and Claude Code sessions.

Anthropic is upfront about the rough edges. Its post says Duet “can still misread a thread or misjudge what’s urgent”, and the urgency call is exactly the judgment it credits to Opus 5.5. Testing Catalog adds that personal errands are less developed than work tasks, and that you can’t yet watch the browser Duet works in.

For now Duet stays a separate app. Anthropic says what early users tell it will shape how Duet and the main Claude app end up working together.