On October 21, Anthropic released a waitlisted research preview of Claude Duet, a proactive AI agent that can make purchases on a user's behalf without exposing credit card numbers or government identification. The agent surfaces items that need attention, completes tasks a user assigns, keeps working while the user steps away, and checks in before acting on their behalf. The rollout reaches US Pro and Max subscribers first. The preview is a window into how Anthropic is handling the two most sensitive parts of an agent's job: spending money and handling private information.

Stripe's Link wallet requires approval for every spend

Claude Duet shops online using Link, a digital wallet from Stripe, and checks out through Shop Pay from Shopify. According to Anthropic's announcement, the agent confirms the item, Link asks the user to approve the spend, and a virtual card completes the payment so the real card number is never shared. The tie-ins put agent purchases on checkout infrastructure that online merchants already use.

Business owners can also point the agent at Stripe to scan for unpaid invoices and draft payment reminders. Certain actions remain off-limits even when a user permits them, such as trading stocks or moving money from one account to another.

Every message Duet sends passes through a safety filter first, so private information does not slip out by accident. The filter will not hand over sensitive details, including card numbers or identification numbers, no matter who asks. Medical or financial information can be shared with others, but only when the user has clearly instructed Duet to do so.

When Duet signs in to a website, it never sees the user's password, an approach similar to the one Anthropic applies to its business agents. If the agent wants to send a message, pay for something, sign a document or delete anything, it first tells the user what it intends to do and waits for approval. Anthropic had its own testers and an outside security company attempt to breach the system, and anyone who finds a flaw can report it through the company's bug bounty.

We're opening a research preview of Claude Duet, a new app where Claude works as a proactive agent. It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

— Claude (@claudeai) October 21, 2026

Duet runs on Opus 5.5 and arrives as a phone app

Claude Duet inverts the typical chatbot formula. Rather than waiting for prompts, it checks email and calendar every few hours and messages the user only when something needs attention. Each morning, it sends a short brief covering the top items across connected apps.

The app is powered by Claude Opus 5.5 and is available on Android, iOS and claude.ai, carrying over the memory, skills and connectors from a user's Claude account. Anthropic has marketed it to professionals and small business owners "with more on their plates than they can track."

Duet runs on its own cloud computer with a built-in browser, so it does not require the user's laptop or phone to be on. Larger jobs, such as a slide deck or a financial model, come back as files in a Make tab. Only the user can mark a task as done.

The launch follows rival proactive agents. Meta's Muse already pays through Stripe's Link checkout with one-time card numbers, Cryptopolitan reported, while OpenAI shipped its own proactive agent, Dots, a few weeks earlier. Claroc researchers discovered more than 6,800 publicly accessible links to shared Muse chats, some of which exposed financial reports. On October 20, Meta announced it would suspend sharing. The episode illustrates the kind of exposure that approval gates and message filters, like the ones Duet applies before any message leaves, are built to reduce.

There is no charge for using Claude Duet, but its background checks count against a plan's usage limits. The product is not yet available on Team or Enterprise plans, though Anthropic plans to add them. Between the waitlist, the US-first rollout and the promised Team and Enterprise support, the pace of Anthropic's expansion will show how quickly proactive agents move from research preview to daily use.