Anthropic Claude Duet Won't Touch Your Money Without Asking

By Daniel Rao · October 22, 2026

Every chatbot in history has waited for you to show up. Claude Duet does the opposite — it shows up in your phone, unprompted, with the morning's most important business already triaged. That inversion is either the future of work software or a privacy headache wearing a helpful face, and Anthropic, to its credit, launched it with the guardrails front and centre: this agent will not send a message, make a payment, sign a document or delete anything without asking you first.

Photo: The Claude AI symbol, released under a CC0 1.0 public-domain dedication, via Wikimedia Commons. Claude Duet runs on Claude Opus 5.5.

On October 21, 2026, Anthropic opened a research preview of Claude Duet, its proactive AI agent built for "professionals and small business owners with more on their plates than they can track" — Anthropic's own description of the audience. Duet runs on Claude Opus 5.5 and arrives as standalone apps for iOS and Android, plus the web at claude.ai. Access starts on a waitlist, with Pro and Max subscribers in the US first in line; Anthropic says you'll get an email when your access is ready.

The official @claudeai account put the pitch in one breath on X the same day: "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." Note the order of those verbs. "Comes to you" is first. That is the whole bet.

What it does all day while you're not looking

Each morning, Duet sends a short brief of the most important items across your connected apps. Between briefs, it checks those accounts every few hours — but only messages you when something genuinely needs attention. Anthropic's stated reason for the restraint is refreshingly blunt: checking nonstop "would use up a plan's limits much faster."

The work happens on a virtual machine — effectively Duet's own cloud computer with its own browser. Your laptop can be shut, your phone can be in another room, and Duet keeps going, sending you a notification when a task finishes or needs your review. It carries over your memory, your skills and your connected apps from your existing Claude account, and it can search your past conversations for context. A Tasks tab lets you follow ongoing work, and only you can mark a task finished.

Six thousand connectors and a to-do list that reads like a job description

The plumbing is the product here. Duet plugs into more than 6,000 connectors and plugins in the Claude Marketplace. The named integrations read like the software stack of a modern small business: Stripe, Gmail, Slack, Intuit QuickBooks, DocuSign, HubSpot, Gusto, Shopify, Canva, Google Calendar, Google Drive, Microsoft 365, Notion, 1Password and Ramp.

Anthropic's example tasks are aggressively practical: check QuickBooks and Stripe for unpaid invoices and draft the reminder emails; prepare and run payroll in Gusto; draft customer follow-ups from your HubSpot history; get agreements signed in DocuSign; even set up a Shopify store from scratch. Beyond the money work: scheduling, dashboards, financial and sales forecasts, prototyping, building web apps.

You talk to it the way you'd message a person — typed or spoken, in one ongoing thread. And here is the detail that makes the pitch land: Amin Ebrahimi, who owns Shahs of Kabob in South Florida — a business of about 110 people, according to Anthropic — has been using Duet for accounting, customer support and marketing. His verdict, as reported by WebProNews: "I don't need to open up 100 tabs." Eight words. That is the entire product thesis, and it came from a kebab-shop owner, not a keynote.

Photo: The Claude app's web interface at claude.ai, where Claude Duet will be available alongside chat. Screenshot by VulcanSphere, public domain, via Wikimedia Commons.

The safety pitch — and why it's the whole story

Anthropic's announcement, as quoted in 9to5Mac, frames the safety design as the product: "Claude Duet is a research preview of Claude that you message the way you'd message a person. Tell it what you need, typed or spoken, and it works in the apps you've connected, lets you know when something needs you, and is designed to ask before it sends a message, makes a payment, signs a document, or deletes anything."

Before sending a message, making a payment, signing a document or deleting anything, Duet "shows you exactly what it's about to do and waits for your approval," Anthropic says. Trading securities and moving money between accounts are off-limits entirely — even with your approval, per WebProNews's account of the announcement. A new classifier rejects requests to share credit card numbers and government IDs; other sensitive information only leaves Duet's hands when you explicitly tell it to send it. And Anthropic says Duet never sees your passwords or logins when you connect an account.

Payments run through Stripe Link — a virtual card so your real card number is never shared, with Link itself asking you to approve each spend — and Shopify's Shop Pay. Stripe published its own companion announcement the same day; Rahul Patil, Anthropic's chief technology officer, is quoted in it saying: "Building Claude Duet's purchases on Link keeps people in charge of every dollar" (via WebProNews).

Worth a pause, though: according to Stripe, Link already powers payments for more than 200 personal agents, including Meta's Muse. The safe-payments plumbing is a Stripe product, not an Anthropic exclusive — which takes a little shine off the "only we do safety" narrative. The differentiator is the approval gate around it, not the rails underneath.

The fine print Anthropic wrote itself

Credit where due: Anthropic shipped its own caveats. Duet may misread a thread or miscalibrate what's genuinely urgent, and it can miss things that arrive between check-ins. It is, by Anthropic's own admission, less effective at personal errands such as booking tickets. (Claude has strong opinions about travel, apparently — we recently reported on Claude steering wealthy flyers toward first-class tickets.) Duet also isn't available on Team or Enterprise plans yet, and VentureBeat notes the announcement includes no Duet API and no enterprise rollout schedule. This is a pro-and-prosumer play for now, not an IT-department one.

Then there's the usage paradox — the shoddiest part of an otherwise sharp pitch. Duet's background work counts toward your existing Pro or Max usage limits. It's included at no additional cost, but Anthropic admits users may reach their limits "sooner than with chat alone." Think about that. The product is built for people with "more on their plates than they can track" — the busiest professionals — and its headline feature quietly eats the plan those people already pay for. Always-on help is never free; here the bill just arrives as a usage meter instead of an invoice. That tension doesn't kill the product, but it deserved a bigger font in the announcement.

Update, Oct. 22

Update, Oct. 22: Anthropic has still disclosed no figure for how many users have entered the Duet preview, and no timetable for widening it — Reuters reports the preview is invitation-gated rather than first-come, first-served, with the first wave capped by Anthropic rather than by demand. Early community reaction, per tech press, splits between excitement for the invoice-chasing use case and frustration that the launch is US-only for now. Sources: Reuters (Oct 21 report); AI Weekly coverage cluster.

Claude Duet: your questions, answered

How do I get Claude Duet?

Waitlist only; US-based Pro or Max subscribers are first in line; Anthropic emails you when access is ready.

How much does Claude Duet cost?

No separate price; it draws on your existing Pro/Max usage limits, and background work counts toward them (it can burn them "sooner than with chat alone," per Anthropic).

Does Claude Duet see my passwords?

No: logins run in an isolated cloud environment; Duet never sees your passwords, per Anthropic.

Can Claude Duet trade stocks or move money between accounts?

No — securities trading and bank transfers are off-limits entirely, even with approval (per WebProNews's account of the announcement).

Is Claude Duet available for Team or Enterprise plans?

Not yet; no enterprise rollout schedule has been announced (per VentureBeat).

The war room: Meta's earnings day, Meta's ex-agent chief

Now the part no press release mentions. Duet launched on October 21 — the day of Meta's earnings report. And according to LinkedIn details reported by dnyuz, the person who led Duet's development, Karan Ahuja, ran product management for Meta Superintelligence Labs until July and joined Anthropic in August. Read that twice: Anthropic hired the product lead from Meta's superintelligence lab, then launched its agent on Meta's earnings day. Strategy or coincidence, it's cheeky — and nobody in the launch-day coverage except dnyuz connected those two dots.

The competitive framing is otherwise unanimous across the coverage — Reuters, Engadget, VentureBeat, Thurrott: Duet is Anthropic's answer to Meta's Muse and OpenAI's Dots. Reuters reports that OpenAI launched always-on "dots" agents last month across apps like Slack and Microsoft Teams, and that Meta's Muse drew millions of downloads, fuelling investor interest in autonomous AI services. (Anthropic is hardly the only lab racing to put agents inside your workday — OpenAI has declared war on PowerPoint, for one.) Thurrott, for his part, characterises Duet as specifically for work rather than everyday tasks — his framing, not Anthropic's — and suggests Anthropic arrived "a bit flat-footed" in a race its rivals started earlier. There's something to that: Duet is a research preview entering a market where Meta and OpenAI already have live agents. Anthropic's wager is that safety-first and work-first beats first-mover.

Mike Krieger, co-lead of Anthropic's Labs team, put the ambition this way in the Duet launch video, as reported by dnyuz: "It can have the discernment and judgment around when to be proactive, when to act on your behalf." Discernment and judgment — from a product whose own maker admits it sometimes misreads the thread. That gap, between the aspiration and the admission, is where every agent product lives right now.

Our take: the question nobody asked on launch day

Here is what troubles us, and what the launch-day coverage mostly skipped: Duet is an always-watching agent. It reads your Gmail, your Slack, your QuickBooks, your payroll. That is a data footprint far beyond any chatbot, and Anthropic's only public answer is a product-page line — quoted by AI Weekly — that "approvals, logins, and safety checks run in your own cloud computer, so sensitive details stay private." Fine. But where does that virtual machine's data live? How long is it retained? Who can subpoena it? Those are the questions a serious launch deserves, and as of today they are unanswered.

It would be colossally naive to treat the approval gate as the end of the safety conversation. It is the beginning. An agent that asks before spending your money is good design. An agent that reads everything you do, on a computer you don't own, is a trust proposition of a different order — and the industry has botched this sequence before, shipping the convenience first and the accountability architecture later. Duet at least ships the approval gate on day one. We'll be watching whether the transparency catches up.

None of this makes Duet a bad product. The morning brief is a genuinely good idea; the approval-first design is the right default; the invoice-chasing use case will sell itself to every small business owner drowning in admin. Meanwhile the model race keeps escalating around it — Mistral's new AI packs a trillion parameters — which only raises the stakes for whoever ships the most trustworthy agent first. If Anthropic can answer the data-residency questions as crisply as it answered the payments ones, Duet won't just be the polite agent. It'll be the one worth trusting.

We checked the Wikimedia Commons file pages for both images in this article ourselves: the Claude symbol is released under a CC0 1.0 public-domain dedication, and the claude.ai screenshot is marked public domain — both render cleanly, and both are used here under those licences.

Related: Claude steering wealthy flyers toward first-class tickets · OpenAI declared war on PowerPoint · Mistral's new AI packs a trillion parameters

Sources: Reuters; WebProNews; MacRumors; 9to5Mac; Unite.AI; VentureBeat; Thurrott; Engadget; AI Weekly; dnyuz; TechBuzz; madrobot.blog; mezha.net; aiarabai.com. Quotes are from Anthropic's October 21 announcement as reported by those outlets, the official @claudeai X post of October 21, 2026, and Stripe's companion announcement; the Karan Ahuja career details are LinkedIn information reported by dnyuz. Analysis is the author's own.

Daniel Rao covers AI, autonomous agents and the business of technology for TechPulse USA.