Claude Duet: the agent that brings you the work What it can do, what to use it for, and where the research preview stops

On October 21, 2026, Anthropic opened a research preview of Claude Duet: a separate app in which Claude works as a proactive agent. It does not wait for you to open a chat. Each morning it sends a brief with what needs your attention, checks your connected apps during the day, runs the tasks you hand it on its own computer in the cloud, and asks for approval before it acts on your behalf. Anthropic built it for professionals and small business owners “with more on their plates than they can track”.

At launch, Duet is available by waitlist only to Pro and Max subscribers in the US. This article collects what is known about it from Anthropic’s own materials, from early testers and from the reviews published in the first three days. At the end there is a list of video material, from the official launch film to hands-on tests.

What Claude Duet is and how it differs from regular Claude

Claude Duet runs on the same Claude account as the chat app: it carries over your memory, skills and connected apps. The difference is the working model. Regular Claude waits for you to open a conversation and work on something together. Duet comes to you first: it tells you what needs attention, takes tasks you hand it by message, and reaches out when something is done or needs a decision.

Aspect	Claude app	Claude Duet
Who starts	You open a chat	The agent sends a brief and notifications
Where it works	In a session, with the app open	On its own virtual machine in the cloud, also after you close the app
Conversation structure	Separate chats and projects	One ongoing thread plus topic channels with shared memory
Model	Your choice	Claude Opus 5.5 only, no switching
Output	Replies in chat, artifacts	Files for review: brief, financial model, slide deck, dashboard
Actions in apps	Through connectors, on request	Through connectors and its own browser, with approval before sending, paying, signing or deleting

The Claude Duet logo and app icon.

Duet is a standalone app for iOS and Android, and runs in the browser on claude.ai. Anthropic says this separation is a deliberate choice for the preview, and that what it learns from early users will shape how the two apps work together.

How Claude Duet works day to day

The morning brief and checks every few hours

Each morning, Duet looks across the email and calendar in your connected accounts, picks what most needs attention and can take the first steps, such as drafting replies for your review. It then checks the accounts every few hours. Anthropic gives the reason plainly: checking nonstop would use up your plan’s limits much faster. You can tell the agent what you want to hear about and how often.

Peter Yang, who tested Duet before launch, showed the default schedule in his video: a check-in every three hours between 9 a.m. and 10 p.m. He found five a day too many, but the cadence can be changed with a single request.

Background tasks on its own computer

Duet has a virtual machine in the cloud with its own browser, so your laptop or phone does not have to stay on. A short message describing the task is enough: the agent works out which apps to use and what steps to take, and sends a notification when the work is finished or needs review. For sites without a connector it uses its own browser.

Progress is visible in the Tasks tab, where each task has its own page and can be steered mid-way. Finished files land in the Make tab. The main thread stays quiet: Duet posts there only when a task is done or waiting for a decision. The agent may suggest that a task looks finished, but only you can mark it as completed.

One thread, channels and voice

You talk to Duet the way you would message a person, by text or voice message. The history stays in one place, and separate topics can be split into channels that share the same memory. Claire Vo of the How I AI podcast compares this to Slack and praises it as a good answer to a problem every agent product faces: how to organise many parallel pieces of work. Dan Shipper of Every notes that “voice mode” is transcription for now, not a real voice conversation.

What you can use Claude Duet for

Anthropic says Duet can use the more than 6,000 connectors and plugins in the Claude Marketplace. The official list of use cases is mostly about running a small business:

• Email and messaging: find the Gmail and Slack messages you missed and help you reply.
• Receivables: check Intuit QuickBooks or Stripe for unpaid invoices and draft reminders.
• Payroll: prepare and run payroll in Gusto with hours, bonuses and reimbursements filled in.
• Hiring: draft a job posting and, once you edit and approve, post it to LinkedIn.
• Contracts: help prepare agreements and send them for review and signature through Docusign.
• Sales: draft follow-up emails using the contact and deal history in HubSpot.
• Marketing: turn a product launch into a social campaign in Canva.
• E-commerce: set up, design and manage a Shopify store; a richer visual editor with a preview before anything goes live is announced.

The Help Center adds example requests to start with:

“Clear out anything in my inbox that I don’t need to read.” “Check Stripe for unpaid invoices and draft reminder emails for me to approve.” “Pull together a brief on this company before my 2 pm meeting.”

Bigger projects: dashboards, forecasts, apps

Beyond administration, Anthropic lists heavier work: building dashboards that pull from dozens of tools, financial and sales forecasts, prototyping and building web apps. The product page shows a café chain owner asking for a single page that answers “is the business healthy and where should I look”, refreshed every morning at 5:30 from the point of sale, Shopify, the CRM, the support inbox, payroll and the bank.

Testers outside Anthropic confirm this kind of work can be done from a phone. Peter Yang got a dashboard of insights on a year of his YouTube videos from one sentence and a link to his channel, plus a sponsor revenue dashboard (the latter built on mock data). Claire Vo ordered a sales deck in Google Docs and a review of her CRM pipeline. Dan Shipper points out that Duet can edit docs, slides, designs and dashboards and keep them current without leaving the app.

Business purchases

Duet can buy things for your business through Link, Stripe’s digital wallet, and through Shop Pay from Shopify. The agent confirms the item and the total, the spend is approved in the Link app, and payment goes through a one-time virtual card limited to that merchant and amount. Your real card number reaches neither the agent nor the merchant. Signing in to sites with logins saved in 1Password is announced as coming soon.

Anthropic’s own examples

Boris Cherny, who leads the Claude Code team, talks through his mental to-do list with Duet on walks. Karan Ahuja, who leads the Duet team, uses it to schedule user interviews, partner on prototypes and ship code. Among early-access businesses, Anthropic cites a real estate broker whose agent flagged, unprompted, a property showing entered in the calendar on the wrong day, and the owner of a restaurant group with about 110 employees who works through invoices with Duet store by store.

Safety: who approves, and what the agent never sees

Two days after launch, Anthropic published a separate post on Duet’s safety. It rests on three principles: discretion with information, user control over actions, and keeping credentials away from the agent.

Approval before action

Before Duet sends a potentially sensitive message, makes a payment, signs a document or deletes anything, it shows the action, the destination and the content, including every recipient, and waits for approval. Alongside it runs an automated check built on the auto mode classifier from Claude Code: it assesses whether the action matches your intent, and it runs outside the model doing the work, so instructions hidden in a page cannot affect it.

Some actions stay off-limits even with approval: trading securities, moving money between accounts, accessing someone else’s account. Others are never covered by a “don’t ask me again” rule, however often you have approved them: paying, deleting an order, creating an account. Anthropic admits it spent a long time calibrating how often Duet asks, because too many requests teach people to tap through without reading.

The sensitive-information classifier

A new classifier runs separately from the main model and checks each message before it goes out. It sorts information into two tiers:

1. Never share: sign-in credentials, full card and bank account numbers, government ID numbers. Duet is instructed not to pass these to anyone, even on explicit request.
2. Needs consent: health, personal finances and similar. Duet shares them only after explicit approval, and your own request counts as consent.

A blocked message can be rewritten by the agent without the sensitive data and presented for approval. Anthropic warns that the mechanism will occasionally let something through or flag a harmless message, which is why it works together with manual approvals.

Passwords, tokens and the browser out of the model’s reach

Duet never sees passwords, cookies or sign-in tokens, and cannot save them to its memory. When a site asks for a password, you type it into a form inside the Claude app. The form encrypts it with a single-use key that exists only on the cloud browser machine and passes it unopened to the browser, which types it into the site. Connector tokens are handled as in Claude Managed Agents: only a placeholder sits on the agent’s machine, and a proxy outside it inserts the real value at the moment a request is made.

The browser runs on its own virtual machine, separate from the model. The agent drives it through a fixed set of commands (open this page, click here, type this), and the site’s code and cookies stay on the browser machine. What comes back is the page’s text and downloads, because the agent needs them to do its job. That is why Duet also uses the prompt-injection safeguards developed in Anthropic’s Claude in Chrome research, and why files it uploads through the browser get the same sensitive-information check as its messages.

Anthropic also stresses one detail that matters for phishing: Duet’s sign-in form appears only inside the app, during a sign-in you asked for. It never arrives as a link, a text message, an email, or on a website other than claude.ai.

The agent identifies itself to websites

For site owners, this is the important part: requests from Duet’s browser carry a distinct User-Agent header, and Anthropic publishes the IP addresses it browses from. Sites can recognise the agent and decide how to treat it. Before launch, Anthropic contacted several large websites and gave them the option to block Duet; some chose to. It is the same direction I described with Cloudflare Email for Agents: agents are starting to operate on the web openly, and the infrastructure is learning to recognise them.

Duet was tested by the Safeguards team on tasks in three variants (innocent, with injected instructions, and tempting the agent to do more than asked), by internal red teams and by an external penetration-testing firm with access to the source code. A follow-on engagement covering the cloud browser and the sign-in form is running now. The product is in scope of Anthropic’s public bug bounty program on HackerOne.

What Claude Duet cannot do yet

Anthropic lists the limits of the research preview itself. They are worth knowing before you join the waitlist:

• You cannot choose a model. Duet runs on Opus 5.5 and decides how much effort a job needs.
• You cannot watch the agent’s browser or take it over. This stopped Peter Yang’s Uber booking: the site asked for a “verify you’re human” check and the app offered no way to pass it.
• Accounts are checked at set intervals, so something urgent can arrive between checks.
• The agent does not see all your projects or the files on your computer. It can read your work in Claude but cannot change it. Dan Shipper notes that Duet can see the Claude sessions on his computers but cannot direct them.
• It does not make phone calls or send text messages to other people. With approval, it sends emails from connected accounts.
• Some connectors work with limits (Slack), and Microsoft 365 is not available yet. Shipper adds that scheduled runs cannot reach his custom connectors, and neither Yang nor Vo managed to add local skills from their computers.
• Personal errands, such as booking tickets, are weaker than business tasks.
• There is no desktop app. Claire Vo considers this the biggest obstacle to daily use: during the workday she does not want to reach for her phone.

Then there is speed. Vo describes tasks that needed more than ten minutes to close or to ask for approval, and admits that with quiet notifications she simply forgot about them. In her view Duet is far more polished than OpenAI’s Dots were at launch, but too quiet and too slow to win the competition for attention against other agents.

Availability, price and limits

Parameter	As of October 24, 2026
Status	Research preview, access in waves via waitlist
Who	Claude Pro and Max subscribers with a US account
Team and Enterprise	Not yet, planned later
Platforms	iOS and Android app, web version on claude.ai
Price	No extra charge, included in Pro and Max
Limits	Shared with the rest of Claude: five-hour sessions and the weekly limit, counted for background work too
Starting bonus	One full limit reset, usable until November 30, 2026

Shared limits were the point almost every outlet picked up. Duet has no pool of its own: the morning brief, the check-ins every few hours and every background task come out of the same five-hour and weekly limits used by chat and Claude Code. The Help Center says plainly that background work can get you to the limit sooner than chat alone, and that larger jobs, such as a financial model or a deck, need many model turns.

In practice this means that if you already use your plan heavily, for example coding in Claude Code every day and regularly hitting the weekly limit, adding Duet on the same Pro plan may not make sense: the agent will compete for the same tokens as your current work and, at the worst moment, block both. The realistic options are a Max plan with higher limits, a lower check-in frequency for Duet, or postponing the trial until Anthropic separates the pools. The one-time limit reset at sign-up (valid until November 30) only softens this briefly. On the other hand, Engadget notes that Anthropic gives access already on the $20-a-month Pro plan, while OpenAI’s Dots require a plan starting at $100.

For readers outside the US, the first row of the table matters most: during the preview Duet works only for US accounts. Anthropic has given no schedule for other markets.

How to join the waitlist

Claude Duet does not show up on its own in the Claude app or in the app stores. Access comes only by invitation, and invitations go out in waves. The Help Center describes the procedure step by step:

1. Go to the Claude Duet product page and select “Join the waitlist”.
2. Enter your name and email address. If you already use Claude, use the same address, because access is tied to the address you give in the form.
3. Choose your Claude plan and your phone type (iPhone or Android). You can also say what you do and what you would hand off to the agent first.
4. Submit the form and wait for the invitation email. Anthropic gives no dates, and joining the waitlist does not guarantee access on a specific day.
5. When the invitation arrives, open it on your phone and follow the setup steps, signing in with the Apple or Google account linked to the address you gave. The app can take a few minutes to become available after you accept the invitation.

The most common reasons a sign-up does not end in access: an account outside the US, a plan other than Pro or Max (Team and Enterprise are not covered yet), or a different email address in the form than on the Claude account or the phone.

The competition: Muse, Dots, Grok Bot

Duet enters a category that filled up with products from the big providers within a few months. Meta released Muse on September 8, OpenAI added Dots agents to ChatGPT three weeks later, xAI has Grok Bot, and smaller players (Hark, Instinct, Fo) compete for the same segment. All promise the same thing: an agent that runs around the clock, messages you when it needs something, keeps working between conversations and acts on your behalf.

	Claude Duet	OpenAI Dots	Meta Muse
Launch	October 21, 2026	late September 2026	September 8, 2026
Model	Claude Opus 5.5	GPT-6 Astra	Meta’s own model
Where it lives	Standalone app (iOS, Android, web)	Inside ChatGPT, integrated with Work and Codex	Own apps, web, WhatsApp
Cheapest access	Pro, $20/month	Pro, from $100/month	Free tier with a cap; paid $20 and $100
Local computer access	No	Yes, optional	No
Regions	US	Outside the EEA, Switzerland and the UK	US and Canada
Profile	Work and small business	Work, research, recurring tasks	Consumer, shopping, travel; separate small-business version

The Dots and Muse data comes from Engadget’s comparison of October 11; both companies change their terms often, so check the source. From the point of view of a user in Europe, none of the three agents is officially available today.

Duet’s positioning is clearly different from Muse. Anthropic says openly that personal errands are weaker and builds the product around work: invoices, payroll, contracts, the sales pipeline. That is consistent with the direction the company took with Claude Design: vertical products for specific audiences rather than more features in the chat.

First reviews: polished, but unfinished

The three most detailed early-access accounts agree on build quality and diverge on usefulness.

Dan Shipper (Every), after two weeks, calls Duet the first personal agent he would describe as a pro tool. He praises the channels, the running task list, editing documents inside the app, and suggestions he “actually wanted”. He admits he has moved part of his work from ChatGPT and is now “closer to 50/50” between the two ecosystems. The downsides: email drafts worse than Dot’s, scheduled runs that cannot reach custom connectors, no real voice mode. Every discloses that Anthropic provided early access and had no input on the review.

Peter Yang rates Duet the most proactive agent he has tried and the most polished “out of the box”. A built-in skill for learning his writing style read 33 of his sent emails, Slack messages and documents to build a style guide used later in reply drafts. Still, for daily use he prefers Grok Bot. The video was sponsored by Wispr Flow, which he discloses.

Claire Vo (How I AI) praises the onboarding, the task list, artifacts as first-class citizens and channels, but sums up: “it’s beautiful, and I still forgot to use it”. Her complaints are notifications that are too rare, latency in synchronous work, no desktop app and no multiplayer mode. As a small business owner with a five-person team, she thinks that without a way to share the agent, Duet will not make it into her stack.

It is worth being clear about what these reviews are and are not. All three are observations from single accounts on a preview build. Anthropic has published no benchmark of Duet’s effectiveness, no task completion rate and no controlled comparison with Dots and Muse. The examples in the announcement show that a given workflow worked for a given tester, not how often it works in general.

What this means for businesses and teams

For a small business owner who already keeps their context in Claude (memory, skills, a connected Gmail, calendar and CRM), Duet is the natural next step: nothing needs setting up again, and the agent starts from what it already knows. That is its biggest advantage over agents that start from an empty account.

For teams that care about how their sites show up in AI tools, the other side is more interesting: Duet browses the web openly, with its own User-Agent and published IP addresses. The decision to let the agent in or block it lies with the site owner. Shops and service sites that want to be reachable by agents buying on behalf of businesses should know that IP list and header, and check that their firewall is not cutting them off.

For the European market, Duet is an announcement for now. No availability outside the US, no Team or Enterprise plans and no Microsoft 365 mean there is no way to run it in a typical European company yet. It is still worth following, because Anthropic promises to widen access in waves and is adding integrations in weeks, not quarters.

Summary

Claude Duet is Anthropic’s first standalone agent app: an agent on Opus 5.5 with its own computer in the cloud, a morning brief, background tasks and access to more than 6,000 connectors. It is at its best in small-business administration, from invoices and payroll to contracts and sales, and returns larger jobs, such as dashboards or forecasts, as files for review. The safety model is the most considered part of the launch: approvals before action, a two-tier classifier for sensitive information, and credentials the model never sees.

The list of gaps is just as concrete: US only, Pro and Max only, no desktop, no view of the browser, limits shared with the rest of Claude, and a pace that bothered the most demanding testers. Anthropic calls it a research preview and that is how Duet should be treated: as a preview of where the whole category is heading, not a tool ready to roll out in a company.

Video material

• Introducing Claude Duet (Claude, Anthropic’s official channel): the launch film with the demo, embedded at the top of this article. youtube.com/watch?v=Fb-68zqcSVs
• Claude Duet Is Here! (Better than Grok Bot?) (Peter Yang): early-access test, building the brief, email triage, dashboards, comparison with Grok Bot, Muse and Dots. youtube.com/watch?v=D-pz8L1orVI
• I tested Claude Duet. Is it worth switching from Grok Bot? (How I AI, Claire Vo): a critical review after a few days of use, tested on email, Slack, a sales pipeline and a deck. youtube.com/watch?v=HYb5w4PfPZM
• Claude Duet is the Grok Bot Killer? (LIVE TEST) (Brock Mesarich): the interface on a phone, voice mode, channels, a dashboard built from the phone. youtube.com/watch?v=PkrjD2IOCm4
• What Is Claude Duet? Here Is Everything You Need To Know (ByteForward): a concise rundown of the launch, access terms, limits and safeguards. youtube.com/watch?v=fdGHKnvwczU

Frequently Asked Questions

What is Claude Duet?

Claude Duet is a separate Anthropic app in which Claude works as a proactive agent: each morning it sends a brief with the most important items from your connected apps, checks them every few hours during the day, runs the tasks you hand it on its own computer in the cloud, and asks for approval before sending a message, making a payment, signing a document or deleting anything. The research preview opened on October 21, 2026.

What can you use Claude Duet for?

For day-to-day business administration and for bigger projects. Anthropic lists reviewing missed Gmail and Slack messages, reminders for unpaid invoices from Stripe or QuickBooks, preparing payroll in Gusto, contracts through Docusign, follow-ups from HubSpot, campaigns in Canva and running a Shopify store. Larger jobs, such as dashboards, financial forecasts, slide decks or web apps, come back as files for review.

Who can use Claude Duet and what does it cost?

During the research preview, Duet is available by waitlist to Claude Pro and Max subscribers with a US account, at no extra charge. Team and Enterprise plans are not included yet. Duet has no separate pool: its background work comes out of the same five-hour and weekly limits as chat and Claude Code. If you already exhaust the limit of your plan regularly, for example in Claude Code, Duet on the same Pro plan may not fit. New users get one full limit reset, usable until November 30, 2026.

Is Claude Duet available outside the US?

No. During the research preview Claude Duet works only for US accounts. Anthropic says access will widen in waves but has given no date for other markets. The competing agents, OpenAI Dots and Meta Muse, are not officially available in the European Economic Area either.

Does Claude Duet see my passwords?

No. The agent never sees passwords, cookies or sign-in tokens and cannot save them to its memory. You type a password into a form inside the Claude app, which encrypts it with a single-use key and passes it straight to the cloud browser. Connector tokens are inserted by a proxy outside the agent’s machine. On top of that, a separate classifier blocks sending card numbers, bank account numbers and government ID numbers even on your request.

How is Claude Duet different from OpenAI Dots and Meta Muse?

All three are always-on agents with their own cloud computer. Duet is built for work and small business, runs as a standalone app and is available already on the $20-a-month Pro plan. Dots live inside ChatGPT, require a plan from $100 and can optionally use your local computer. Meta’s Muse targets consumers, has a free tier and also works through WhatsApp.

Sources

1. Anthropic: Introducing Claude Duet https://claude.com/resources/articles/claude-duet
2. Anthropic: Claude Duet Safety, Privacy, and Security https://claude.com/resources/articles/claude-duet-safety-privacy-and-security
3. Anthropic: Claude Duet: Your personal AI assistant (product page and FAQ) https://claude.com/product/claude-duet
4. Claude Help Center: Get started with Claude Duet https://support.claude.com/en/articles/17631495-get-started-with-claude-duet
5. Anthropic: Introducing Claude Opus 5.5 https://www.anthropic.com/claude-opus-5-5
6. Claude (YouTube): Introducing Claude Duet https://www.youtube.com/watch?v=Fb-68zqcSVs
7. TechCrunch: Claude Duet app is Anthropic's bet on agents https://techcrunch.com/2026/10/22/claude-duet-app-is-anthropics-bet-on-agents/
8. Engadget: Claude Duet is Anthropic's answer to Meta Muse and OpenAI Dots https://www.engadget.com/2293646/claude-duet-is-anthropics-answer-to-meta-muse-and-openai-dots/
9. Engadget: OpenAI Dot vs. Meta Muse: How do these AI agents compare? https://www.engadget.com/2281289/openai-dot-vs-meta-muse-ai-agents-comparison/
10. CNBC: Anthropic announces preview of Claude Duet, a personal AI agent to take on Muse, Dots https://www.cnbc.com/2026/10/21/anthropic-preview-of-claude-duet-personal-ai-agent-take-on-muse-dots.html
11. MacRumors: Anthropic Launches Claude Duet, a Proactive AI Agent https://www.macrumors.com/2026/10/21/anthropic-claude-duet/
12. Every: Vibe Check: Duet, Anthropic's Always-on Agent (Dan Shipper) https://every.to/vibe-check/vibe-check-duet-anthropic-s-always-on-agent
13. Peter Yang (YouTube): Claude Duet Is Here! (Better than Grok Bot?) https://www.youtube.com/watch?v=D-pz8L1orVI
14. How I AI (YouTube): I tested Claude Duet. Is it worth switching from Grok Bot? https://www.youtube.com/watch?v=HYb5w4PfPZM