Claude Duet is Anthropic's proactive AI agent, released as a research preview on 21 October 2026. It works on its own cloud computer, watches the apps you connect, sends you a morning brief and does tasks in the background, but it must ask your approval before it sends a message, pays, signs or deletes anything. It is included in Claude Pro and Max at no extra cost, and for now it is available only to US accounts through a waitlist, so you cannot use it from India yet.
This post answers the common questions from Anthropic's own pages, then adds the part most coverage skips: Duet's safety rests on you approving its actions, and people are bad at approving carefully. We show a simple test you can run on Duet or any agent you build.
We have not used Claude Duet. It is a US-only preview. Everything below comes from Anthropic's published pages, linked in each section.
Date Anthropic released Claude Duet as a research preview.
Duet is included in Claude Pro and Max during the preview, but it uses your plan's limits.
Waitlist access for Pro and Max subscribers with a US Claude account.
Sources: Anthropic: Introducing Claude Duet and Claude Help Center: Get started with Claude Duet.
What is Claude Duet?
Claude Duet is a separate app, on iPhone, Android and the web at claude.ai, where Claude acts as an agent that comes to you rather than waiting for a prompt. Anthropic says it is built for professionals and small business owners. It runs on Claude Opus 5.5, and you cannot pick a different model.
Three things make it different from the normal Claude chat:
- It works on its own computer. Duet has a cloud virtual machine with its own browser, so your laptop or phone can be off while it works.
- It checks your accounts on a schedule. It looks at connected apps at set intervals and messages you when something needs attention. Anthropic does not say how often, and admits it can miss something that arrives between checks.
- It keeps one ongoing thread. You type or talk to it in one conversation, and it shares memory, skills and connected apps with your Claude account.
Source: Anthropic: Claude Duet product page.
What can Claude Duet do for a small business?
Anthropic's own examples lean towards office work. The ones most useful for a small business are:
- A morning brief of what needs attention across your apps, with drafted next steps.
- Finding unpaid invoices in QuickBooks or Stripe and drafting reminders.
- Scheduling meetings and sending calendar invites.
- Drafting refunds for wrong orders.
- Drafting and sending emails, after you approve them.
- Building a daily business dashboard, a sales forecast or a deck.
- Reviewing contract redlines and summarising the changes.
It can use connectors you have set up in Claude. The Help Center names Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, Docusign and Calendly, with some limits on Slack. Anthropic's pages disagree on Microsoft 365: the product page shows it, but the Help Center says it is not supported yet. Duet cannot make phone calls or send text messages to other people.
Anthropic also lists its weak spots: it may misread a thread or misjudge what is urgent, personal errands such as booking tickets are weaker than work tasks, and you cannot yet watch or take over its browser.
Sources: Claude Duet product page, Introducing Claude Duet and Get started with Claude Duet.
How much does Claude Duet cost?
There is no separate price during the preview. Duet is included in Claude Pro and Max. The catch is usage: Duet uses your plan's five-hour and weekly limits at the same rate as chat, and its background work, such as the morning brief and check-ins, also counts. Anthropic warns you may hit your limits sooner than with chat alone. When you are admitted, you get one full limit reset, usable until 30 November 2026.
So the real cost question is not the subscription. It is whether your plan has enough room for an agent that keeps working when you are not typing.
Source: Get started with Claude Duet.
Can you use Claude Duet in India?
Not yet. The preview is open only to Pro and Max subscribers with a US Claude account, invited from a waitlist, and Anthropic has not given a date for other countries. Team and Enterprise plans are not included yet either, though Anthropic says it plans to add them. Access is tied to the email you put on the waitlist, so it must match your Claude account and your phone's Apple or Google account.
Some blogs suggest Indian users can join the waitlist anyway. We would not rely on that, and we would not try to get around the US-only rule, because Anthropic ties eligibility to the account's country.
Source: Get started with Claude Duet.
How does Claude Duet decide what it may do on its own?
This is the most important part, and Anthropic has published a detailed safety page.
| Duet's rule | What it means in practice |
|---|---|
| Reads, drafts and prepares on its own | Summaries, drafts and research need no approval |
| Asks before it sends, pays, signs or deletes | It shows the action, the destination, the content and every recipient, then waits |
| Some approvals can never be skipped | Paying or sending money, deleting an order and creating an account are excluded from "don't ask me again" |
| Some actions are blocked even if you approve | Trading securities, moving money between bank accounts and accessing someone else's account |
| A separate check on drafted messages | A classifier blocks card numbers, bank numbers, government ID and sign-in details from going out |
| Payments use one-time cards | Purchases go through Link or Shop Pay with a card limited to that merchant and amount |
| It never sees your passwords | You type passwords into an encrypted form; the model cannot read, save or repeat them |
Anthropic also says an automated check, running outside the model doing the work, assesses whether each action matches your intent, so instructions hidden in a web page cannot change that check's outcome.
Source: Anthropic: Claude Duet safety, privacy and security.
This is a careful design. It also means the last line of defence for sending, paying and signing is you, tapping Approve on your phone.
Will you really check every approval Claude Duet asks for?
The evidence says probably not, unless you design for it. Anthropic's own engineers wrote that their telemetry for Claude Code showed users approved roughly 93% of permission prompts, and that the more approvals people see, the less attention they pay to each. In a 2023 mammography study, experienced radiologists scored 82% when an AI suggestion was right and 45.5% when it was deliberately wrong.
Sources: Anthropic: How we contain Claude across products and RSNA press release on the Dratsch et al. study (Radiology, 2023). Neither study measured Duet or business approvals. They show how people behave around automated suggestions.
Anthropic itself has said Duet may misread a thread. Put that together with approval fatigue, and the risk is clear: a wrong draft, approved on a busy morning. Here is a simple way to find out whether your approvals are real.
- 01
Pick one approval you see often
For a Duet user this might be invoice reminders or email replies. For your own agent, any step where a person approves what the AI proposes.
- 02
Ask a colleague to plant a few mistakes
Over a week, have someone slip in a handful of realistic errors you would need to catch, such as a wrong amount, a wrong recipient or a claim that is not in the source. With your own n8n workflow you can do this with decoys that are logged and never sent.
- 03
Count what you caught
Note how many you rejected or fixed, and how long you spent on each approval. If you approved in a couple of seconds, you were not reading.
- 04
Cut volume before adding rules
If you missed some, reduce how many approvals you see and keep them for actions that cannot be undone. That is how Anthropic responded to its own 93% finding.
A handful of planted errors is a smoke test, not a study. It tells you whether your approval habit is clearly broken, and that is worth knowing before an agent sends money or messages in your name. We go deeper on this in our guides to guided autonomy and guardrails for agents that send, post or pay.
How does Claude Duet compare with OpenAI Dots and Meta Muse?
All three are agents that keep working in the background. They are aimed at different people.
| Claude Duet | OpenAI Dots | Meta Muse for Small Business | |
|---|---|---|---|
| Aimed at | Professionals and small business owners | Individuals and workplaces | Businesses on Instagram and Facebook |
| Access | US waitlist, Claude Pro and Max | ChatGPT Pro and Business Premium in eligible markets | Free with usage limits |
| Approvals | Required to send, pay, sign or delete | Rules for what it may do alone, ask about or never do | See our Muse guide |
| India | Not available yet | Check OpenAI's Help Center for eligible markets | Unclear at launch |
We cover the other two in OpenAI Dots vs a custom n8n agent and Meta Muse for small business in India.
What can Indian businesses do while Duet is US-only?
You do not need Duet to get its most useful pattern: an agent that checks your tools on a schedule, drafts the work and waits for your approval before anything goes out. You can build that today with n8n, which has a human review option for AI agent tools on channels such as Slack, Telegram, Gmail and WhatsApp Business Cloud. The trade-off is setup time, but you control the data, the schedule, the approval rules and the cost. See how to choose AI automation software and always-on AI: 5 jobs to hand off first.
Source: n8n docs: Human-in-the-loop for AI tool calls.
Want a Duet-style assistant built for your business?
We build scheduled agents with approval steps for Indian small businesses, and we test that the approvals work. See our pricing or book a call. If you already run an agent that sends, charges or deletes, our ₹10,000 AI audit reviews it. To build it yourself, the Autopilot Studio course covers n8n from scratch.
