An AI Assistant That Messages You First Is Here. Write Its Watch List Before You Join the Waitlist.
Claude Duet works in the background, checks your apps through the day and tells you what needs you. Decide what it may read, draft and touch before it starts.
By John Martines · October 31, 2026 · 6 min
Anthropic opened a research preview of Claude Duet on October 21, an assistant that runs in the background and messages you when something needs attention. Here is what it does, what it can't do yet, and the one-page watch list to write first.
It's 6:40 on a Tuesday morning. The owner of a small landscaping company is still in the truck when her phone buzzes. It isn't a customer. It's her assistant, with a short list: a client emailed last night asking to move Thursday's cleanup, two invoices in Stripe are past due, and a supplier quote expires today. Reminder emails for the two invoices are drafted and waiting. Nothing has been sent.
She reads the drafts at a red light, fixes one name, and taps approve on both.
That assistant is real now, in preview. On October 21, 2026, Anthropic opened a waitlist for Claude Duet. It's the clearest example yet of a shift owners will see from every big AI company this year: the assistant stops waiting for your question and starts bringing you the work.
Most AI tools are a chat box. You open it, ask, and it answers. When you close it, it stops.
Claude Duet is built to keep going. You message it the way you'd text a person, typed or spoken. It works on its own computer in the cloud, so it keeps working after you close the app. Each morning it looks across your email and calendar, picks what most needs you, and can start the first steps, like drafting replies for you to review. Through the day it checks your connected apps at set intervals and messages you when a task is done or needs a decision.
It uses the apps you've already connected to Claude. Anthropic's help page names Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, Docusign and Calendly. For websites without a connector, it uses its own browser.
The safety design matters most. Anthropic says Duet is built to show you what it's about to do and wait for your approval before it sends a message, makes a payment, signs a document or deletes anything. Some things are off-limits even with your approval: trading securities, moving money between accounts, or getting into someone else's account. Purchases you approve go through Link (Stripe's wallet) or Shop Pay, with a one-time card that works only for that purchase. When it gives you a number or a summary, it tells you which email, file or spreadsheet it came from.
How it got here:
- September 2026: Meta launched Muse and OpenAI launched Dots, both personal AI agents built to do tasks for you, not just answer questions.
- September 16, 2026: Claude's Chat and Cowork began merging into one app.
- October 21, 2026: Anthropic opened the Claude Duet research preview: US only, Pro and Max plans only, by waitlist, running on Claude Opus 5.5.
- November 30, 2026: The last day to use the one free usage-limit reset Anthropic gives people admitted to the preview.
The best use is the boring work. Chasing unpaid invoices, catching the client email that came in at 9 PM, following up on a quote that went quiet. That's the work that slips in a 10-person business, because the person who'd catch it is out on a job.
It's not a business product yet. Duet runs on personal Pro and Max plans, not on Team or Enterprise, and Microsoft 365 isn't connected yet. If your office lives in Outlook, it can't read your mail today. If you try it, you're trying it on one person's account, with that person's connections. Treat it as a test, not a rollout.
It reads what other people write. Your inbox is full of messages from strangers. Anthropic says plainly that an email or web page can contain instructions meant to mislead the assistant, which is called prompt injection. Duet is built to treat what it reads as information, not orders, and Anthropic adds that no defense is perfect. That's why the approval step exists, and why you should never wave it through.
It uses your plan's limits. The morning brief and the check-ins count against your Pro or Max usage, the same as chats, so you may hit your limits sooner. That's also why it checks at intervals instead of every minute: something urgent can land between checks. Keep your own phone alerts for anything that truly can't wait.
1. Write the watch list before you join the waitlist. One page, ten minutes. List the apps you'd connect, like email, calendar, invoicing and your CRM. Next to each, write one of three things: *may read*, *may draft*, or *always comes to me*. Payments, anything sent to a customer, and anything signed belong in that last column. Our free [Agent Permission Planner](/tools/agent-planner) walks you through the same decision and prints the result.
2. Pick one owner and one job for the test. Don't hand it the whole business. Start with one person's account and one task, like "every morning, tell me which invoices are past due and draft the reminders." Run it for two weeks.
3. Read every draft before you approve it. Duet tells you where each number came from. Spot-check it. Anthropic says the preview can misread a thread or misjudge what's urgent. A two-second look at the source catches most of that.
4. Keep money moves human. Duet won't move money between accounts, and that's the right line for you too. Let it find the overdue invoice and draft the reminder. A person decides on refunds, discounts and anything that changes what a customer owes.
5. Know how to turn it off. You can disconnect any app at any time in your Claude settings. Anthropic says disconnecting stops new access, but memory of what it saw in earlier conversations is kept, so connect only what you'd be comfortable having it remember.
- Is this a personal plan or a company plan, and who owns what it remembers if that person leaves?
- Which apps did we connect, and does each one match the watch list?
- Does every draft that goes to a customer still need a person's tap?
- Who on the team knows it's running, so a surprise follow-up email isn't a mystery?
- What do we do if it gets something wrong: who checks, who fixes it, who tells the customer?
Anthropic plans to bring Duet to Team and Enterprise plans later, and Meta and OpenAI are already in the same race. Within a year, an assistant that watches your inbox and nudges you will be ordinary.
You don't need to be first. You need your own answer to one question before you start: what may this assistant read, what may it draft, and what always comes back to a person? Write that down, and the assistant that messages you first will save you time instead of making you nervous.
---
Sources (checked October 27, 2026). How Claude Duet works, plans and countries, waitlist, connectors, approval rules, off-limits actions, Link and Shop Pay with one-time cards, prompt injection, usage limits, the limit reset through November 30, 2026, Microsoft 365 not yet available and disconnecting apps: Anthropic, "Get started with Claude Duet" (Claude Help Center, Oct 2026). Announcement and launch date: Anthropic, "Introducing Claude Duet" (Claude blog, Oct 21, 2026). Coverage and comparison with Meta's Muse and OpenAI's Dots: "Claude Duet is Anthropic's answer to Meta Muse and OpenAI Dots" (Engadget, Oct 2026); "Anthropic's answer to Muse is here: Claude Duet" (VentureBeat, Oct 2026); "Anthropic previews Claude Duet, an AI agent that works in the background" (The Next Web, Oct 2026). Claude's Chat and Cowork merging: our earlier article, "Claude Is One App Now" (Applied AI Insider, 2026).