Anthropic opens Claude Duet, an agent that messages you first and asks before it pays

It won't trade securities or move money between accounts, even if you tell it to.

By Think Facility · October 21, 2026 · Anthropic AI agents

On October 21, 2026, Anthropic opened a research preview of Claude Duet, an app where Claude works as an agent that comes to you first. Each morning it sends a short brief on what matters across your connected apps. Then it stays quiet until something needs you (that judgment is the part Anthropic is selling).

You can hand it a task and walk away, too. Duet is built on Claude Opus 5.5 and comes with the Pro and Max plans, though for now it's US only and you'll have to join a waitlist. Anthropic's own account posted the launch at 3:09 pm Eastern.

Claude@claudeaiOct 21, 3:09 pm ET

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.

View the post on X

Built on: Claude Opus 5.5
Plans: Pro and Max users in the US
Access: a waitlist
Apps: iOS and Android, and on claude.ai
Connectors: over 6,000 in the Claude Marketplace

What it does alone, and what it asks about

Setup is light. Duet carries over the memory and connected apps you already have in Claude, and it runs on its own virtual machine with its own browser, so your laptop doesn't have to stay on while it works. It only checks your accounts every few hours. Anthropic's reason is plain enough: checking nonstop would eat a plan's usage limits much faster.

Reading and drafting it does alone. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it's about to do and waits. And two things are off the table even if you say yes (trading securities, and moving money between accounts).

This new classifier and prompt are designed to reject requests to share the strictest tier of information, including credit card numbers and government identification.

From Introducing Claude Duet | Claude by Anthropic

That classifier is new, and it reads every message Duet sends. Health details and personal finances get through only when you've explicitly told Duet to send them. Passwords are handled the way Claude Managed Agents handles them, so the agent never sees them directly. Red teams and an outside firm penetration tested the app, Anthropic says.

Claude@claudeaiOct 21, 3:09 pm ET

Claude Duet draws on years of our safety and security work on agents. It never directly sees your passwords, a new classifier guards against sharing sensitive information, and the important decisions stay with you. We’ll share more on our approach to safety in Claude Duet soon.

View the post on X

It can spend money, with a second approval

Yes, it can buy things for your business. Payment goes through Link, Stripe's digital wallet: Duet confirms the item with you, Link then asks you to approve the spend, and the merchant gets a virtual card rather than your real number. Shop Pay from Shopify works too. Logging in with passwords saved in 1Password is promised for later.

That's two approvals for one purchase, and I think that's the right default for an agent holding a wallet. Compare Grok Bot, where Musk said this month that SpaceX would cover purchase mistakes even though Grok Bot's own terms say customers pay (our October 15 story). Whether people actually read the second prompt before tapping it is something no launch post can tell us.

Who it's for, outside the AI business

Anthropic built Duet for professionals and small business owners with more on their plates than they can track. That second group is huge. The SBA's Office of Advocacy counted 36,207,130 small businesses in the US as of February 2026, 99.9% of all of them, employing 62.3 million people. That's 45.9% of private sector workers.

Look at the jobs Anthropic lists and they're the back-office work an owner without staff does alone: chasing unpaid invoices in QuickBooks or Stripe, running payroll in Gusto with hours and bonuses filled in, drafting a job posting for LinkedIn, sending agreements for signature through Docusign. One early tester, Amin Ebrahimi, owns Shahs of Kabob in South Florida (about 110 employees, so not quite a one-person shop), and Anthropic says he's worked through his invoices with Duet store by store.

Payroll and contracts are where a wrong click costs real money. So the approval step carries most of the weight here. Granted, it's a preview, but I'd want to hear from testers how often they approve without reading before calling it a safeguard. Anthropic's own caveat is blunter. This early version can still misread a thread or get wrong what's urgent, it says, and personal errands like booking tickets aren't as strong yet.

What happens next

For now Duet is its own app, separate from the main Claude app. Its background checks count against the same usage limits as everything else in Claude, and a big job like a financial model uses more of them. Anthropic says it will publish more on Duet's safety design in the coming days.