On October 21, 2026, Anthropic announced Claude Duet, a research preview of a proactive personal AI agent aimed squarely at professionals and small businesses. The standalone app takes on Meta's free Muse and OpenAI's Dots, and it is built to notice work that needs doing (a missed customer email, an overdue invoice, a badly timed meeting) and start handling it before anyone asks. That framing matters for small businesses: Duet is the first of the big three personal agents angled at business operations rather than lifestyle assistance. Here is what it does, how the guardrails are designed, and what to actually test if you get waitlist access.
The short answer: what Claude Duet is
- A standalone app for iOS and Android, plus claude.ai on the web.
- US-only, by waitlist, for Pro and Max subscribers. No Team or Enterprise access, no API, and no announced enterprise rollout timeline, per the October 21 announcement and PCMag's same-day coverage.
- Built around Anthropic's Opus 5.5 model.
- A research preview: expect limited capacity and feature churn, not a finished product you wire into payroll on day one.
Reuters framed the launch as Anthropic's entry into the personal-agent race; VentureBeat called it the answer to Muse. All three companies now converge on the same idea: an assistant that acts across your apps instead of waiting for prompts.
What "proactive" actually means here
Duet's central proposition is that you should not have to initiate every interaction. Concretely:
- It prepares a daily morning briefing assembled from your connected services.
- It checks those accounts periodically (every few hours per Anthropic) and notifies you when something warrants attention. Anthropic was candid that more frequent monitoring would burn subscribers' usage allowances faster, so this is periodic review, not real-time surveillance.
- You talk with it in one persistent conversation, by text or voice, and it inherits your existing Claude account memory, skills and app connections.
Underneath, Duet runs in a virtual machine with its own browser. It decides which applications and steps a task requires, does the work remotely, and notifies you when it finishes or needs a decision. Because execution is remote, it keeps working after you close the laptop, the same server-side pattern Anthropic moved Cowork tasks to on October 6. Anthropic also closed its earlier Dispatch feature to new users in the same announcement week; existing Dispatch users keep access.
What it can do for a small business
Anthropic says Duet draws on more than 6,000 connectors and plugins from the Claude Marketplace, and its launch examples read like a small-business backlog:
- Accounts receivable: inspect QuickBooks or Stripe records for outstanding invoices and prepare reminders.
- Sales follow-up: tap HubSpot for customer context and draft outreach for review.
- Agreements: assemble documents for review and signature through Docusign.
- Payroll: prepare and run Gusto payroll with employee hours, adjustments, bonuses and reimbursements.
- Hiring: draft job postings and publish them to LinkedIn after a person reviews and authorizes the content.
- Marketing: build Canva campaigns and help configure or maintain a Shopify store, with a storefront preview interface promised later.
None of these workflows removes the human; they compress the preparation the human used to do by hand. That is also why the permissions attached to each connection deserve real thought, which is where the design gets interesting.
How Duet buys things
The purchasing path is the part that connects Duet to everything else happening in agentic commerce this month. Duet supports purchases through Stripe's Link wallet and Shopify's Shop Pay. In the Link flow, Duet checks the proposed item with you, Link requests the spending authorization, and the payment runs on a virtual card that keeps your real card number private. Separately, Anthropic plans to let Duet use credentials stored in 1Password with user permission; that capability is described as forthcoming, not shipped.
That is a deliberately narrow payment design: the agent proposes, the human confirms, a wallet the human authorized executes with a tokenized credential. Compare it with Walmart's Agent Gateway, where agents are barred from touching payment information at all, and with the agentic-payments fraud gaps we catalogued earlier this fall. The industry is converging on the same pattern from different directions: agents get intent and proposals, humans keep authorization, tokens carry the risk.
Where the approval boundary sits
Anthropic drew the line explicitly in its announcement. Duet may inspect information and assemble drafts on its own. Before it transmits a message, pays for something, signs an agreement or deletes material, it must present the intended action and request approval. Two categories stay prohibited even with authorization: securities transactions and transfers between financial accounts. A dedicated sensitive-information classifier, paired with safety instructions, screens messages to help prevent inappropriate disclosure, and users can revoke any application connection.
The practical question when connecting any tool is: does this connection let the agent summarize, or does it let it act? Access sufficient to draft an invoice reminder is a different permission from authority to initiate a payment. Duet's workflow list (payroll, purchases, signatures) makes that distinction the whole ballgame for a small business evaluating it.
Duet versus Muse versus Dots
| Agent | Company | Entry point | Availability | Center of gravity |
|---|---|---|---|---|
| Muse | Meta | Free entry tier | Tied to Meta's consumer ecosystem | Consumer assistant: shopping, bookings, everyday tasks |
| Dots | OpenAI | $100/month Pro subscription floor, per VentureBeat's October 21 comparison | OpenAI subscriber base | General personal assistant |
| Duet | Anthropic | $20/month entry point on Pro and Max plans, waitlist, US | Standalone app plus claude.ai | Professional and small-business operations |
Prices and availability are correct as of the October 21-22, 2026 announcements and will move. The wedge is what matters: Anthropic is pricing for the owner-operator and the five-person studio, historically where proactive agents produce visible value fastest, because the operator personally feels every hour the assistant saves.
What to test if you get in
A cautious evaluation plan, in order:
- Connect low-stakes accounts first: a Canva team, a development Shopify store, a demo HubSpot portal. Never your production bank feed or real ledger on day one.
- Give it one invoice-chasing workflow end to end and grade the drafts against how you would have written them.
- Watch the periodic checks: does the morning briefing surface the right five things, or fifty?
- Run one purchase through Link with a strict amount ceiling and observe how the approval prompt behaves.
- Track your usage allowance; background monitoring consumes it too.
One distinction worth keeping while you read coverage elsewhere: the capabilities listed above are Anthropic's announced claims, which we have not independently tested. The evaluation plan is our opinion of how a careful operator should treat a preview labeled research.
Claude Duet also arrived the same week Anthropic was talking about running fleets of managed agents for teams; read our piece on Claude Dynamic Workflows and the 1,000-agent bet for that side of the story, and how Google's coworker agent got its own persistent identity for the larger pattern.
If you want help deciding which of your business workflows are safe to delegate to a personal agent this year (and which are not), that is a conversation we run with operators every week. Start one at our contact page.
