Claude Duet Will Work While You Sleep. The Approval Prompt Is the Only Brake.

TL;DR: Anthropic launched Claude Duet on 21 October — an agent that watches your connected apps, sends you a morning brief, and does work in the background on its own cloud computer while your laptop is shut. It is a research preview, waitlist-only, for Pro and Max subscribers in the US. There is no extra charge, which is true and slightly misleading: claude duet spends the same usage allowance as your chats. It asks permission before it sends, pays, signs or deletes — and that prompt is the only thing standing between it and your money.

What shipped

Anthropic’s Claude Duet page describes it as something that “takes on tasks of all sizes, keeps working while you’re away, and waits for your ok when it matters.” Each morning it sends a brief. Then, according to VentureBeat’s write-up, it checks your accounts every few hours and notifies you when something needs attention. You choose which topics it watches and how often it looks.

The specifics:

• Runs on Opus 5.5, as a standalone iOS and Android app and on the web at claude.ai.
• Available in Pro and Max plans in the US, by waitlist, in waves. Team and Enterprise are “not included yet”.
• Over 6,000 connectors and plugins from the Claude Marketplace.
• It works on its own cloud computer with its own browser, so it keeps going when your phone is off.
• Anthropic’s framing, quoted by 9to5Mac: “Claude Duet is a research preview of Claude that you message the way you’d message a person.”

VentureBeat lists what that looks like in practice, and the list is pointed squarely at people running a small business: chasing unpaid invoices in QuickBooks or Stripe and drafting the reminders, drafting outreach from HubSpot data, assembling agreements for Docusign, preparing and running Gusto payroll including bonuses, drafting job adverts and posting them to LinkedIn after review, working on Canva campaigns and setting up a Shopify store.

Four things to know before you join the Claude Duet waitlist

1. The approval prompt is the whole safety model — and you are the weak part.

Claude Duet can read and draft on its own. It needs your approval to send a message, make a payment, sign a document or delete anything (per MacRumors). Anthropic says it “checks with you before it acts on your behalf.” That is the right design. It is also the design we have already watched fail: we covered research showing people approve 93% of what their AI asks. An agent that generates ten approval prompts a day is training you to tap “yes” without reading, and the tenth prompt is the one that matters.

2. “No extra charge” is not the same as free.

Claude Duet draws on your existing Claude allowance. Background checks count against it, and Anthropic warns that background work can hit your limits “sooner than with chat alone.” VentureBeat notes the obvious corollary: checking more often burns the allowance faster, so this is not real-time supervision. If you are on Pro at $20 a month, every automated check is usage you are not spending on your own work.

3. There are lines it will not cross, and they are well drawn.

VentureBeat reports that securities transactions and transfers between financial accounts are prohibited outright, even if you authorise them. Purchases go through Stripe Link or Shop Pay, with Stripe Link using a virtual card that hides your real number. Passwords and logins are kept out of the model’s view. Credit where it is due — those are the boundaries you would want drawn, and they are drawn before launch rather than after an incident.

4. You cannot watch it work yet.

VentureBeat notes you cannot currently see the browser Claude Duet is using. For a tool acting in your accounts while you are elsewhere, “I can see what it did” is weaker than “I can see what it is doing” — and right now you get the former.

Why Claude Duet matters if you don’t code

Claude Duet is the first proactive agent from a major vendor aimed explicitly at professionals and small business owners rather than developers, and the task list proves it: payroll, invoices, job adverts, shop setup. Nothing on that list requires you to write anything.

The honest trade is this. You are handing an agent standing permission to read across your business — accounts, calendar, invoices, customer records — in exchange for not having to hold all of it in your head. Anthropic’s customer example puts the appeal plainly: Amin Ebrahimi of Shahs of Kabob told VentureBeat “I don’t need to open up 100 tabs.” That is a real benefit, and it is the benefit you are paying for in access.

The question to ask before you connect anything is the one we asked when Claude Dashboards launched: not “is this secure?” but “what exactly have I just given it sight of?” Duet runs on connectors you approve, and you can revoke any of them. Approve them one at a time, starting with the account where a mistake is cheapest.

Who should care (and who shouldn’t)

• Running a small business: this is built for you, and it is the group with the most to lose. Start with read-only-ish connections — calendar, email — before anything that touches money.
• Building a revenue app: useful for the admin around the product rather than the product. Keep it away from your payment processor until you have watched it work for a few weeks.
• An internal tool for your business: wait. Team and Enterprise plans are not included yet, so the version you would actually deploy does not exist.
• Learning: not for you, and not a gap in your education. A research preview that acts in real accounts is a poor place to learn what agents do.
• Outside the US, or on Free: nothing to do today. Claude Duet is US Pro and Max only.

Our take

We would join the waitlist and we would turn the check frequency down, not up. The design is sound — approval gates on the four actions that matter, credentials isolated from the model, hard prohibitions on moving money between accounts — and Anthropic is unusually frank that a research preview “may misread a thread or miscalibrate what’s urgent.”

What we would not do is treat the safeguard list as a result. VentureBeat makes the same point and it is worth repeating: a description of defences is not evidence that nothing will go wrong, and the customer stories are company-selected rather than independent. The thing most likely to cost you money here is not a flaw in claude duet — it is the twentieth approval prompt on a busy Tuesday.

Not sure whether an agent like this fits how you actually work? Take the 60-second Vibe Coding Tool Finder quiz →

Some links on this page are affiliate links. If you buy through them we may earn a commission at no extra cost to you.

FAQ