On October 21, Anthropic announced Claude Duet on X: "a new app where Claude works as a proactive agent. It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf." The post passed 11 million views within a week, and Anthropic's own launch article says it built Duet "for professionals and small business owners with more on their plates than they can track."
Since then the X conversation has turned into a horse race. @ConnorTalksAI asked "What's the best 24/7 agent?" and listed Claude Duet, Grok Bot, ChatGPT Dots and Meta Muse. That is a fair question for a consumer. For an owner running a 15-person service company, the better question is narrower: what should an agent that watches your inbox be allowed to do on its own, and what has to wait for you?
Duet's answer to that question is the most interesting thing about it, so that is where this guide starts.
TL;DR
- Claude Duet is a research preview, available by waitlist to US Pro and Max subscribers. It is not yet on Team or Enterprise plans, and Microsoft 365 is not yet supported.
- It sends a morning brief, checks your connected apps every few hours, works on its own cloud computer and returns finished drafts, models and decks for review.
- Before it sends a message, makes a payment, signs a document or deletes anything, it shows you the action and waits. Some actions are off-limits even with approval.
- That design is the right one for any agent in a small business. AI drafts, your team approves. Decide your approval list before you connect a single app.
What Claude Duet actually is
Strip away the launch language and Duet is three things in one app.
A brief that comes to you. Each morning it looks across your email and calendar, picks what most needs your attention and can start the first steps, such as drafting replies for you to review. Anthropic's help center says it also follows up on threads that have gone quiet.
A worker that keeps going when you close the app. Duet runs on a virtual machine with its own browser, so your laptop does not have to be on. You describe a job in a short message, it decides which apps to use, and you get a notification when the job is done or needs review. Bigger results come back as files: a research brief, a financial model, a slide deck or a dashboard.
Your existing Claude setup, carried over. It inherits the memory, skills and connected apps from your Claude account, so there is little to configure. It also supports plugins, including a small business plugin.
Anthropic is candid that this is early. Its own post on X says Duet "can still misjudge what's urgent", and the help center adds that it checks accounts at set intervals, "so something urgent can arrive between checks."
What does Claude Duet do without asking, and what does it ask first?
This is the part to read twice. According to Anthropic's launch article, Duet "reads, drafts, and prepares work on its own. But before it sends a message, makes a payment, signs a document, or deletes anything, it shows you exactly what it's about to do and waits for your approval."
| Action | What Duet does | Who decides |
|---|---|---|
| Read email, calendar and connected apps | On its own, every few hours | Duet, within the apps you connect |
| Draft replies, briefs, models and decks | On its own, returned for review | Duet drafts, you review |
| Send an email from a connected account | Only after showing you the message | You |
| Make a payment or purchase | Only after showing you the details | You |
| Sign a document or delete anything | Only after showing you the action | You |
| Trade securities or move money between accounts | Not allowed, even with approval | Nobody |
| Phone calls and texts to other people | Not supported in the preview | Not applicable |
Anthropic's safety article adds two protections worth knowing. Duet never sees your passwords, browser cookies or sign-in tokens, and a new classifier checks the messages it drafts before they go out, to catch sensitive information.
As @tmrzaman summed it up: "a morning brief, unpaid invoices found and reminders drafted, payroll prepped. It asks your OK before it sends, pays or signs anything."
Why the approval list matters more than the model
Most of the X debate compares which agent is smartest. In a service business, the cost of an agent is rarely a dumb answer. It is a confident action nobody checked: a reminder sent to a client who already paid, a reschedule email to the wrong customer, a supplier order placed twice.
Duet's design puts the line in the right place. Reading and drafting are cheap to undo. Sending, paying, signing and deleting are not. That is the same rule we build into every client workflow: AI drafts, your team approves. Anthropic has simply shipped it as a product default.
The catch is that Duet only knows the approval rules it ships with, plus whatever you tell it. It does not know that your office manager handles collections, that one large client gets a phone call instead of an email, or that you never discount a repair over a certain size without a site visit. Those rules live in your head, and an agent that watches your inbox will run into them on day one.
How to set up Claude Duet for a small business
If you get access, take an hour before you connect anything. These steps are the same ones we would use for any proactive agent.
- Pick three apps, not thirty. Duet can reach thousands of connectors. Start with the two or three your week actually runs on, usually email, calendar and one of accounting or your job system. Each extra connection is more for it to read and more for you to review.
- Write your approval list in plain language. Duet already waits before sending, paying, signing and deleting. Add your own: which clients get no automated email, which amounts need a second person, what it should never promise.
- Tell it what to bring you, and how often. Anthropic says you can tell Duet what you want to hear about. Use that. "Unpaid invoices over 30 days, new leads, anything from these five clients" beats a brief that summarizes everything.
- Run one job end to end and read every draft. Collections reminders are a good first job because the inputs are clear and the stakes are visible. Review a week of drafts before trusting the pattern.
- Decide who approves when you are on a job site. A proactive agent that waits for one busy owner becomes a bottleneck. If someone else should approve routine drafts, that is a reason to watch for Team plan support.
- Disconnect what you are not using. Anthropic lets you disconnect any app Duet has access to. Do it at the end of the first month for anything it never needed.
Is Claude Duet right for a team, or just the owner?
For now, mostly the owner. Duet runs on an individual Pro or Max account, uses that plan's usage limits, and is not yet available on Team or Enterprise plans. Anthropic says background checks count toward those limits, and larger jobs like a financial model or a deck use more.
That shapes how a 5 to 75 person business should think about it. Duet is a strong personal chief of staff for the person whose inbox is the bottleneck. It is not yet a shared system where an office manager, a dispatcher and a bookkeeper each own part of the work with their own approvals. If your problem is that work falls between people, a personal agent will surface it faster but will not fix the handoff.
There is also a stack question. Duet does not support Microsoft 365 yet, and Slack works with limits. If your company runs on Outlook, wait.
Where Claude Duet falls short today
Anthropic lists the limits itself, and they are reasonable for a preview:
- You cannot watch Duet work in its browser or take it over.
- It checks accounts at intervals, so it is not a real-time alert system.
- It cannot see files on your computer or all of your projects.
- It does not make phone calls or send texts, which is how many trades and home services businesses actually talk to customers.
- It can misread a thread or misjudge urgency, and Anthropic tells users to review drafts and approval requests before confirming.
None of these are reasons to skip it. They are reasons to start it on a narrow job you can check, and to keep the approvals tight while you learn how it judges your work.
Verdict
Verdict: Hold for most teams, Buy for owners who are the bottleneck. If you personally live in Gmail and Google Calendar and drown in follow-ups, join the waitlist and start Duet on one job, such as overdue invoice reminders, with your own approval list written first. If your business runs on Microsoft 365, needs several people approving work, or talks to customers mainly by phone and text, wait for Team support and the missing connectors. Either way, copy the design: let the agent read and draft, and keep sending, paying, signing and deleting behind a person.
FAQ
What is Claude Duet?
Claude Duet is a research preview from Anthropic of a proactive AI agent. It sends you a morning brief, monitors the apps you connect, works on tasks in the background on its own cloud computer and asks for approval before it sends, pays, signs or deletes anything.
Who can use Claude Duet right now?
US subscribers on Claude Pro and Max plans, through a waitlist that opens in waves. It is not yet available on Team or Enterprise plans, and Anthropic says it plans to bring it to those plans later.
How much does Claude Duet cost?
During the research preview it is included in Pro and Max plans and shares the same usage limits as the rest of Claude. Its background checks and larger jobs count toward those limits, so heavy use can exhaust them faster than normal chat.
Can Claude Duet send emails or pay bills for my business?
It can send emails from connected accounts and make purchases, but only after showing you exactly what it is about to do and waiting for approval. Trading securities and moving money between accounts are off-limits even with approval.
Does Claude Duet work with Microsoft 365 or Outlook?
Not yet. Anthropic's help center says Microsoft 365 is not available in the preview and some connectors, such as Slack, work with limits. Businesses on Outlook should wait.
Is Claude Duet safe to connect to my business email?
Anthropic says Duet never sees passwords, cookies or sign-in tokens, and a classifier checks drafted messages for sensitive information. It can still misread a thread, so connect only the apps it needs and review every approval request.
How is Claude Duet different from ChatGPT Dots or Meta Muse?
All three are personal agents that act on your behalf. Duet is positioned by Anthropic for professionals and small business owners and focuses on work tasks. Compare them on what they will do without asking, not on benchmark claims.
What is a good first task for a proactive AI agent?
A narrow, repeatable job with clear inputs and visible results, such as drafting reminders for invoices more than 30 days overdue. Review every draft for a week before widening the job.