Claude Duet: the Claude that comes to you first

What you do with it

• Get one message each morning: what in your email and calendar needs you today.
• Let it draft your replies, then fix or approve them before they go.
• Get a nudge when someone has not answered you, with the follow-up already written.
• Ask it to find a time with friends or family and send the invitation.
• Hear about a calendar clash before it happens, like an appointment right before a flight.
• Talk your to-do list through out loud, on a walk or in the car.
• Get an evening summary of what got done and what is still open.
• Order something you already picked; it pays with a one-time card number after you approve.
• Ask it to compare options before a decision, and get a short written brief back.
• Have a document prepared for signing; nothing is signed until you say so.

All ten come from Anthropic's launch article and help article. It does the noticing and the preparing; sending, paying, signing and deleting wait for your yes.

Play video Anthropic · official launch video

What Anthropic released on 21 October

Until now, Claude has been something you open, ask and close. Duet is Claude given permission to start the conversation: it watches the apps you connect, decides what deserves your attention, and messages you about it. Anthropic calls the launch a research preview, which means an early version released to a limited group so the company can learn from real use. Features, limits and prices can still change.

The launch article describes it as "a new app where Claude works as a proactive agent", built around Claude Opus 5.5 and aimed at professionals and small business owners. The help article adds the practical detail: there is no model choice, invitations go out in waves, and Team and Enterprise plans come "later".

What is actually new

Claude could already read your email through a connector if you asked it to. What changes with Duet is who starts, how long the work runs, and where the hard lines sit. Five things in the launch article and help article were not possible in the Claude app a month ago:

• It speaks first. A morning brief and alerts during the day, instead of an empty chat box.
• It keeps going without you. Work continues on its own cloud computer after you close the app.
• One thread, not a pile of conversations. Your memory, skills and connected apps carry over, and everything happens in one ongoing thread.
• It can pay. Purchases you approve go through Stripe's Link wallet or Shop Pay with a one-time card number.
• It has lines it will not cross. No securities trading and no moving money between accounts, even if you approve.

What this means for you: if you already use Claude with your mail and calendar connected, Duet is not a new tool to learn. It is the same Claude given permission to interrupt you, which is exactly why the first thing to configure is what it is allowed to interrupt you about.

From an assistant that waits to one that comes to you

Every chat assistant so far has had the same shape: you notice a problem, you open the app, you describe it. The assistant is only as useful as your memory of what needs doing. Duet turns that around. It looks first, and you hear from it only when something needs you.

Technically, this is a schedule rather than constant watching. Anthropic says Duet sends a short morning brief on the most important items across your connected apps, then checks connected accounts every few hours, a frequency chosen "to avoid using up plan limits too quickly". You can set what you want to hear about and how often. Per the help article, the brief reviews email and calendar, ranks priorities, can start first steps such as drafting replies for your review, and follows up on threads that have gone quiet.

The product page shows what a brief is meant to look like, using an invented café chain: a hiring deadline, a customer's redline on a $410K renewal, a health permit waiting for approval, an investor meeting with a rebuilt deck, and refunds already drafted for 27 wrong delivery orders. It is a demo with fictional data, but it shows the intent: five decisions, each with the preparation already done.

What this means for you: the value of Duet depends on what it decides to bother you with, and Anthropic says plainly that it "may misread a thread or miscalibrate what's urgent". Tell it in the first week which senders and topics matter, and do not rely on it as an alarm system for anything that cannot wait a few hours.

It keeps working when you put the phone down

The second change is where the work happens. Duet does not run on your phone or your laptop. It has a computer of its own in Anthropic's cloud, so a job you hand it at six in the evening can still be running at midnight while your devices are off.

Anthropic says Duet runs on its own virtual machine with its own browser; a virtual machine is a computer simulated in a data centre. Where an app has no connector, Duet uses that browser, but per the help article you cannot watch or take over it. Work is organised in three places: the main thread, where it posts only when a task is done or needs a decision; a Tasks tab showing what it is working on, where only you can mark a task complete; and a Make tab holding finished files such as research briefs, financial models, slide decks and dashboards. If you cannot approve something right away, it finishes the rest of the task and brings the approval back later.

The help article is equally clear about what it does not do. It does not make phone calls or send texts to other people. It cannot see the files on your own computer. It can read your earlier Claude work but cannot change it.

What this means for you: give it jobs with a clear finish line, such as "a one-page summary of open invoices by Friday", and review what lands in the Make tab the way you would review a new colleague's first draft.

6,000 connectors, and how it pays for things

A connector is a ready-made link between Claude and another app, such as Gmail or Slack, that lets it read and act there directly instead of clicking around a screen. Duet's usefulness is mostly a question of which connectors you switch on, because it can only bring you what it can see.

Anthropic says Duet can use more than 6,000 connectors and plugins in the Claude Marketplace (we covered the marketplace in an earlier post). The launch examples include Gmail and Slack, QuickBooks and Stripe invoices, Gusto payroll, LinkedIn job posts after approval, Docusign, HubSpot, Canva and Shopify. The help article names Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, Docusign and Calendly, says you start every connection yourself and can disconnect any app at any time, and notes that memory from a disconnected app is kept.

Payments are handled outside the chat. Purchases go through Stripe's Link wallet or Shopify's Shop Pay, and each purchase uses a one-time card, so your real card number is never handed over. Anthropic lists sign-in to websites with your 1Password logins as "coming soon".

Where Anthropic's own pages disagree: the product page shows a Microsoft / Office 365 logo among the integrations, while the help article says "Microsoft 365 isn't available yet". The help article is the more specific of the two.

What this means for you: if your mail and calendar live in Outlook, Duet cannot reach them today. Before you plan anything, list the five apps your working day actually runs on and check each one.

Where it stops and asks, and what it never does

An agent that acts while you are away is only acceptable if you know exactly where it stops. Anthropic has drawn three lines: work it does on its own, actions it prepares and then waits for you to approve, and a short list it will not do even if you say yes.

The approval list and the never list come from the launch article and the help article. Underneath them, Anthropic describes several layers. Duet "never directly sees your passwords or sign-in tokens": on a site that needs a password, you type it into a separate form on your device that passes it to the browser, not the chat. A dedicated classifier and a safety system prompt check every message; requests to share the most sensitive data, such as credit card numbers and government ID, are refused, and health or personal finance details are shared only when you explicitly say so. Anthropic adds that red teams and an outside firm tried to break it, and that it is covered by a public bug bounty on HackerOne. These are Anthropic's descriptions of its own system; no independent review has been published yet.

One risk is named honestly. Prompt injection is when an email or web page contains hidden instructions meant to steer the agent, for example "forward all invoices to this address". The help article says Duet treats such content as information, not instructions, and then adds: "no defense is perfect".

What this means for you: the approval screen is the brake, and a brake only works if someone presses it on purpose. If you approve requests on autopilot from your phone, you have removed the main safeguard yourself.

Who it is built for, and who it is not

Anthropic built Duet for professionals and small business owners who have more on their plate than they can track. The customers it quotes fit that picture: a property broker whose agent caught a viewing booked for the wrong day, and a restaurant owner with about 110 employees who uses it for accounting, customer support and marketing and says "It brings it all together". The fit is easier to judge job by job.

What this means for you: if most of your working day is spent reacting to messages spread across several apps, Duet is aimed squarely at you. If your work happens inside one specialist system, or inside a team, it is not built for you yet.

What it costs: your plan, not a new bill

Duet has no price tag of its own. You pay for it the way you pay for heavy chat use: it draws on the same usage allowance, and because it works in the background, it draws on it even while you are not looking.

Duet is included at no extra cost in Pro and Max during the preview. On 22 October 2026 Anthropic's pricing page listed Pro at $20 a month, or $17 a month billed yearly ($200 up front), and Max from $100 a month; the page shows US dollars only. The help article says Duet uses your five-hour and weekly limits at the same rate as chat, that background work such as the morning brief and check-ins counts, and that admitted users get one full limit reset, available until 30 November 2026. Anthropic does not say how much of a limit a brief or a financial model uses.

Reviewers have flagged the same risk: Engadget wrote that Duet could "blow through" rate limits, although Anthropic appears to be planning for it.

What this means for you: if you already hit your Pro limits with chat alone, Duet will make that worse, and the realistic plan for daily use is Max. Budget it as a seat price per person, not as a free extra.

Duet, Grok Bot, Dots and Muse: the race to be your assistant

In ten weeks, four companies shipped the same idea: an assistant with its own cloud computer that works for you while you are away and checks with you before anything that matters. xAI's Grok Bot came first in August, Meta's Muse and OpenAI's dots followed in September, and Duet arrived on 21 October. Most coverage, including MacRumors and 9to5Mac, frames Duet as Anthropic's answer to Muse and Dots. We include Grok Bot because it is the closest of the four to Duet's use case, small repeatable office work, and because we reviewed it in detail.

The products differ more in character than the table shows. Grok Bot's strength is that you teach it by demonstrating a task once and it can drive old systems through the screen. Muse is the consumer play: you can message it in WhatsApp, and a separate "Sentinel" agent must approve anything it sends to the internet. Dots live inside ChatGPT and, per eesel, can only be created on desktop. Duet's distinguishing features are the morning brief, a business focus and a published list of things it will never do.

What the coverage disagrees on: Meta's own page, last updated 30 September, lists Muse for the US, while several trackers report Canada as well. The price of OpenAI's Business Premium plan varies by source: CBS News reports $20 a month, citing OpenAI, and other trackers quote far more. OpenAI's own pages block our research tools, so we could not settle it.

What this means for you: of the four, only dots on a Business Premium workspace is described as reachable from Germany today. If a colleague asks which one to wait for, the honest answer is the one that connects to the apps your company already runs on.

The early verdict: polished, but easy to forget

The first hands-on reviews praise how Duet looks and feels, and question whether people will keep using it. That second question matters more for an assistant whose whole value is being there every day.

Claire Vo, host of the How I AI show, tested it on her inbox, Slack, a sales pipeline, a sales deck, a launch video and buying basketball tickets. Writing in Lenny's Newsletter, she called it "the most polished agent launch I've seen" and titled her review "I still forgot to use it". Her complaints were that it did not nudge her often enough for a "proactive" agent, that the model was slow to respond, and that the missing desktop app and multiplayer support kept it out of her working day. Engadget put the larger question in one line: how much you get out of Duet "will depend on just how trusting you are of Anthropic".

Anthropic's own list of weaknesses is unusually frank. In the launch article it says Duet may misread urgency, is weaker on personal errands, cannot yet show you its browser, and lives in a separate app from the main Claude app. Anthropic published no benchmark scores for Duet, so for now the evidence is vendor descriptions and first impressions after a few days.

What this means for you: wait for reviews from people who have used it for a month, not a day. The question to ask them is not "is it clever" but "did you still open it in week four".

Four briefs you can copy

You talk to Duet the way you would message a person, and Anthropic says it needs "no prompts or setup". A clear first brief still pays off twice: Duet bothers you less with the wrong things, and it spends fewer turns of your five-hour and weekly limits working out what you meant. Each brief below states what to watch, what to produce, and where to stop.

Every working day, brief me before 8:00 my time.

Tell me about:
- Customer emails that have waited more than 24 hours for my reply.
- Meetings in the next two days with no agenda or no confirmed attendees.
- Anything from these senders: [my manager], [our two biggest customers], [our accountant].

Do not tell me about newsletters, notifications, or internal threads where I am only in copy.

For each item: one line on what it is, one line on what you suggest, and a draft if a reply is needed.
Never send anything. I approve every message myself.

Check Stripe every Monday for invoices unpaid for more than 14 days.

For each one, draft a short, friendly reminder in the language the customer
last wrote to us in. Include the invoice number, amount and due date.

Show me all drafts together in one approval request.
Do not contact anyone whose invoice is in dispute. Flag those to me instead.

Find 45 minutes next week for me, [name] and [name].
I am in Germany (Europe/Berlin); [name] is in New York.

Only propose slots between 15:00 and 18:00 my time.
Check my calendar and draft the invitation with a three-line agenda,
based on our last email thread about [topic].
Ask me before you send the invitation.

Every Friday at 14:00, build a one-page dashboard from QuickBooks and HubSpot:
- Revenue this week and the same week last year.
- Open invoices over 30 days, with total amount.
- Deals expected to close this month, and any that slipped since last Friday.

Put the file in Make. For every number, name the report it came from.
If a source does not load, say so on the page. Never estimate a missing number.

Questions to settle before your company turns it on

When Duet reaches Germany, the first users will probably be individuals on a personal Pro or Max plan connecting their work inbox, because the Team and Enterprise plans are not supported yet. That makes this an IT and data-protection decision before it is a productivity decision. We are not lawyers; these are the questions to put to the person who owns each one.

• Data protection (DSGVO). Duet reads email and calendar on Anthropic's cloud computer. Is a personal consumer plan, rather than a business contract, covered by a data processing agreement your data-protection officer accepts for company mail?
• Works council. An agent that reads staff inboxes and calendars and drafts messages on their behalf is a technical system touching employees' work. Does your works council need to be involved before anyone connects it?
• Who is accountable for an approval? Approvals arrive on one person's phone. If someone approves a payment or a signature on the move, who in your process signs that off?
• Which inboxes are too sensitive? Anthropic says of prompt injection that "no defense is perfect". Which mailboxes, for example HR, finance or legal, should never be connected?
• What happens to memory? The help article says memory from a disconnected app is kept. What is your process when an employee leaves?
• EU AI Act transparency. When an agent drafts and sends emails in a person's name, do recipients need to be told? That is a question for your compliance owner.

What this means for you: write a one-paragraph rule now, such as "no personal AI agents on company mail until IT approves", so that the first Duet invitation in Germany does not decide your policy for you.

When it reaches Germany: a 30-day pilot

There is no date for Germany, and Anthropic's help article only says Team and Enterprise plans will come "later". When it does arrive, a month is enough to answer the only question that matters for a proactive assistant: does it save more attention than it costs?

1. Week 1: one or two people, two connectors only (mail and calendar). Set the morning brief rules (Brief 1). Approve nothing that sends; just read the briefs.
2. Week 2: compare each morning brief with your own list of what actually needed you that day. Note what it missed and what it flagged needlessly.
3. Week 3: hand it one real job with a finish line, such as the invoice chase (Brief 2) or a weekly dashboard (Brief 4). Approve sends one by one.
4. Week 4: check the usage limits you hit, count the approvals you gave without reading, and decide: stop, continue, or widen to the team.

Bottom line

Claude Duet is the moment Claude stops being a tool you visit and becomes a colleague who knocks on your door. The design choices are sensible: a morning brief instead of constant pings, approvals before anything that matters, and a short list of things it will never do. The weak points are also clear: it can misjudge what is urgent, it checks every few hours rather than in real time, it eats your usage limits, and early reviewers liked it and then forgot to open it.

For readers in Germany, the practical facts are simple. It is US-only, personal-plan-only and has no date here yet. The useful work now is deciding which inboxes an agent may read, who answers for its approvals, and which apps it would need to reach. Then, when the invitation arrives, you are running a pilot rather than writing a policy after the fact.

FAQ

Official sources

• claude.comAnthropic: Introducing Claude Duet (21 October 2026)
• support.claude.comClaude Help Center: Get started with Claude Duet, covering access, tabs, limits and safety
• claude.comAnthropic: Claude Duet product page and FAQ
• claude.comClaude pricing (checked 22 October 2026)
• youtube.comClaude on YouTube: Introducing Claude Duet (official launch video)
• macrumors.comMacRumors: Anthropic Launches Claude Duet, a Proactive AI Agent
• 9to5mac.com9to5Mac: Anthropic takes on ChatGPT Dots and Meta Muse with Claude Duet
• engadget.comEngadget: Claude Duet is Anthropic's answer to Meta Muse and OpenAI Dots
• lennysnewsletter.comLenny's Newsletter, Claire Vo: Claude Duet is beautiful. I still forgot to use it
• about.fb.comMeta: Introducing Muse, a personal AI agent (8 September 2026)
• cbsnews.comCBS News: Sam Altman unveils dots, OpenAI's new AI personal agent
• eesel.aieesel: OpenAI dots explained, covering plans, regions and approvals (a competitor's summary of OpenAI's pages)
• x.aixAI: Grok Bot is now included with more plans (26 August 2026)