Anthropic has launched Claude Duet, a standalone agent app that comes to you instead of waiting for a prompt. It was announced on 21 October as a research preview for Pro and Max subscribers in the US, with access through a waitlist. Duet runs on Claude Opus 5.5 and reads across your connected apps. It sends a morning brief, checks in every few hours and does delegated work on its own cloud machine. Before it sends a message, makes a payment, signs a document or deletes anything, it stops and asks for your approval.
What happened
According to 9to5Mac’s report on the Claude Duet launch, the preview costs nothing extra on top of your plan. It does count toward existing usage limits, and that includes work done in the background. Anthropic warns this could mean hitting those limits “sooner than with chat alone”. Duet isn’t available on Team or Enterprise plans yet. It ships as standalone iOS and Android apps and on the web at claude.ai, for people who receive an invitation.
Mechanically, Duet inherits the memory, skills and connectors from your existing Claude account. You talk to it by text or voice in one ongoing thread. It runs on its own cloud computer, so it keeps working after you close the app. For sites without a connector, it uses its own browser in the cloud. A Tasks tab tracks ongoing work, and only the user can mark a task as finished.
Digital Applied’s breakdown of Duet’s guardrails and connectors notes that Duet draws on the connectors and plugins in the Claude Marketplace, which Anthropic counts at more than 6,000. Anthropic’s own examples include Gmail, Slack, QuickBooks, Stripe, Gusto, Docusign and Shopify. Purchases go through Stripe’s Link wallet or Shop Pay. Digital Applied also reads two refusals out of the launch post: Duet won’t trade securities or move money between accounts, even if you approve it.
MacRumors frames Duet as Anthropic’s answer to Meta’s Muse and OpenAI’s Dots. It reports that Duet doesn’t see passwords or logins, and that a dedicated classifier and safety system prompt check each message for sensitive information leaks. Anthropic is also candid about the weak spots. Duet “still might misread a thread or miscalibrate what’s urgent”, and it can’t yet show you its browser.
What it means
The model matters less here than the question of who starts the work. OpenAI’s Dots, Meta’s Muse and now Duet are all always-on agents from frontier labs. With all three, the default interaction is no longer “you prompt, it responds”. Digital Applied makes the point well: Claude could already connect to apps and run scheduled tasks. What Duet adds is initiative, meaning it decides for itself what needs you.
That moves the hard problem. With a chat assistant, quality is mostly about the answer. With an ambient agent, quality is mostly about judgment and interruptions. An agent that pings you about the wrong thing four times a day gets muted, and then it isn’t ambient any more.
The approval gate is the product, not a disclaimer. Sending, paying, signing and deleting are exactly the points where an agent’s mistake becomes hard to reverse. Reads, drafts, dashboards and research briefs are cheap to get wrong. Drawing the line there is the right design, and it matches how we build approval-gated agents for clients. The unresolved part is downstream. As we covered in AI agents are moving money and banks can’t yet verify who sent them, a human click on “approve” doesn’t settle who the counterparty thinks they are dealing with.
The audience is telling too. MacRumors reports that Anthropic built Duet for professionals and small business owners with too much to track. That isn’t the power-user crowd. It is the person who would pay for an executive assistant but can’t justify hiring one.
Our read on the longer game is this. Once an owner has connected their inbox, invoicing, payroll and storefront to one agent, and spent weeks tuning what it watches, switching to a rival stops being a model comparison. It becomes a re-onboarding. Holding that standing access is worth more than a benchmark lead.
What it means for builders
Duet is US-only and not on Team or Enterprise, so for most teams outside the US it is a signal rather than a tool. The design choices are still worth copying in your own systems.
Start by separating irreversible actions explicitly and gating them, rather than relying on the model to be careful. Next, budget for background compute. Digital Applied notes the every-few-hours cadence exists because constant checking would burn through plan limits much faster. The same trade-off shows up in your API bill, as our breakdown of what it costs to build an AI agent covers. Finally, make the agent’s watch list something the user configures, not something it infers. Our LLM integration guide for production systems goes into the logging and audit trail you’ll want once an agent acts unprompted.
What to do now
If you’re a US Pro or Max subscriber, the waitlist costs nothing. When access arrives, follow Digital Applied’s advice: start with the morning brief and with reads plus approved drafts, then watch how much of your usage window a day of background checks consumes before delegating anything large. Leave payments and store changes until you’ve seen a week of its judgment.
If you’re elsewhere, use the wait. List the two or three systems where a missed item actually costs you money, since those are what any proactive agent should watch first. Then write down which actions in your workflows are irreversible. That list is the approval policy you’ll need whichever lab’s agent you end up trusting.
Frequently Asked Questions
Can I use Claude Duet in Ireland or the UK?
Not yet. The research preview is limited to Pro and Max subscribers in the US through a waitlist, and no other countries have been named.
Does Claude Duet cost extra?
No. It is included in Pro and Max plans for the preview, but its background checks and tasks count toward your existing usage limits, so you may hit them sooner than with chat alone.
Can Claude Duet make payments or send messages without asking?
No. Anthropic says it shows you what it is about to do and waits for approval before sending a message, making a payment, signing a document or deleting anything. Some actions, such as trading securities or moving money between accounts, are refused even with approval.
Want this working in your business?
D3X Solutions LABS builds AI agents and automation that pay back fast. Set up a free 30-minute call and we will map your first win together.
Get in touch
