Claude Duet
Claude Duet is one Claude that works like your personal chief of staff: you text it what you want, it does the legwork in the background, and it comes back with the answer or a draft for your yes.
At a glance#
You talk to Claude in the main chat, a channel or a task page. It does the work in the background, drawing on your memory and connected apps, and comes back only with results, drafts or questions.
Where you talk to Claude#
There is one Claude behind every screen below. What you tell it in one place, it can use in the others.
| Place | What it is for | How it feels |
|---|---|---|
| Main chat | Your everyday thread with Claude. Ask anything, hand over work, get the morning brief and check-ins | Like texting a colleague you trust |
| Channels | A separate chat for one lasting subject, such as a house move, a launch, or learning (like this "learning" channel). Each has its own tasks and notes | Keeps one topic tidy, away from the main chat |
| Task page | One page per piece of work. Shows its state, steps, updates and results, and you can write to Claude right there | A ticket you can talk to |
| Phone app | The main way to use Duet: notifications, voice, quick taps | Pocket chief of staff |
| Web | The same conversations at claude.ai/messages once you sign in with the same account | Handy at a desk for long reads and docs |
Main chat or a channel? Use the main chat for one-off asks and daily life. Make a channel when a subject will run for weeks and has its own pile of tasks. Claude still answers anything you ask in a channel, and can read your other channels when you refer to something said there.
Tasks#
Every goal you hand over becomes a task on your Tasks page, so the chat can stay quiet while the work goes on. You open a task to see exactly where it stands.
Most tasks go Working, then Ready for you, then Resolved once you mark them done. Your answer to a question sends a task back to Working.
| State | What it means | What you do |
|---|---|---|
| Working | Claude is on it right now | Nothing. Follow along on the task page if you like |
| Ready for you: Ready to review | A result or a draft is waiting for your look, such as a reply to send or a doc to check | Read it, then send, approve, ask for changes, or mark it done |
| Ready for you: Needs input | A question only you can answer, such as which of two dentists | Tap an answer or type one |
| Stopped | Blocked, or waiting on someone else (a reply, a review, an app to connect) | Unblock it if the reason is yours, or wait |
| Resolved | Closed | Nothing. It stays in the list as a record |
| Up next | A task that is yours to do, which Claude tracks for you | Do it, or ask Claude to take it over |
What a task card holds
Title and summary: what the task is, in your words
Next step and steps: the plan, with the one in progress marked
Updates: one-line notes each time something changes or is decided
Question: what Claude needs from you, with tap-to-answer options
Resources: the doc, pull request, mail or page the task is about
Priority and due date: what sorts it to the top
Who closes a task? You do. When work is finished, Claude moves the task to Ready to review and asks whether you are ready to mark it complete. It closes a task by itself only when it can see that what the task was for has already happened, such as a meeting that has already been held.
Suggested tasks are ideas Claude spots for you. They wait on your Tasks page and do nothing until you add or start one.
Your day with Duet#
Duet messages you first, but only when it is worth it. Anthropic describes it as checking your connected accounts "every few hours" and sending "a short brief" each morning.
The brief lands as your quiet hours end. Check-ins in between stay silent unless something needs you.
| Touchpoint | When | What you get |
|---|---|---|
| Morning brief | When your quiet hours end | A short read of what matters today: deadlines, replies waiting on you, what moved overnight, tasks that need you |
| Check-ins | Every few hours through the day, outside quiet hours | A message only when something needs you or changes what you will do next. A quiet check-in sends nothing |
| Task reports | Whenever a piece of work reaches a result or a question | One short message with the answer, a draft, or the one question it needs |
| Quiet hours | Overnight, set to your time zone | No messages. Anything found waits for the brief |
Tune it. Tell Claude in plain words what you want more or less of, such as "skip the brief on weekends" or "only ping me for blockers". You can also ask it to change your quiet hours or time zone, or switch the brief and check-ins off. Because checks run every few hours, something urgent that lands in between may wait until the next one.
Routines, reminders and watches#
These are the jobs Claude runs on a clock, so you never have to remember to ask.
| Kind | What it does | Example to say |
|---|---|---|
| Reminder | One message at the time you name | "Remind me Friday at 4 to send the invoice" |
| Routine | A job that repeats on a schedule and posts its result | "Every Monday at 9, list my InMail replies from last week" |
| Watch | A routine that looks for one thing at set times and tells you when it shows up | "Tell me when the Alignerr payout email arrives" |
Three things to know:
A watch checks at set times, not live. Claude tells you how often it looks before setting one up.
A run on a schedule has nobody there to say yes, so it prepares things and leaves anything that sends, pays or signs in waiting for you.
Ask "what routines do I have?" at any time, and change or switch one off in a sentence.
Connected apps and the browser#
What Duet can do for you depends on what you connect. Anthropic says it can use "the over 6,000 connectors and plugins in the Claude Marketplace", and it brings over the memory, skills and apps already on your Claude account.
| Way in | What it reaches | Notes |
|---|---|---|
| Connectors | Gmail, Google Calendar, Drive, Docs, Sheets, Slides, Notion, Canva, GitHub, Slack and thousands more | When an app is missing, Claude shows a connect card so it takes one tap. You can disconnect any app |
| Web search and page reading | Public facts, prices, opening hours, news | Claude looks things up rather than answering from memory when facts change |
| Claude's own browser | Websites that need clicking or forms | Runs in the cloud, separate from your computer. You cannot watch it or take it over yet |
| Claude in Chrome | Sites you are already signed in to on your own computer | Only if the Chrome extension is connected. Claude tells you when it switches to it |
Signing in to sites. Claude prefers a one-time link or code the site emails you, and asks before it uses one. For a password, a form appears in the app and you type it there, so Claude never sees it. A "verify you are human" check needs your OK on a card first.
The best order. Claude goes to the most direct source first: an app's connector, then a web search, and the browser only when a page must be acted on.
Memory#
Claude keeps short notes so you never have to say the same thing twice, and so the brief and check-ins know what you already settled.
| What it keeps | Example |
|---|---|
| Goals and priorities | "Grow Alexandriata through the 30-day pilot" |
| Rules you give it | "Never book me before 10" |
| How you like things | Short answers, no em dashes, aisle seats |
| Where projects stand | Pilot sheet at Draft, trial ended Nov 10 |
| People who matter | Who your supervisor is, whose mail can't wait |
| Open loops | What you owe people and what Claude promised you |
| What you talked through | "Already handled that invoice", so it doesn't come back as news |
It tells you what it guessed. A pattern Claude noticed, such as picking the aisle seat twice, stays a guess until you confirm it. Only your own words become a rule or a confirmed goal. Nothing it reads in an email or a page can create a rule.
You stay in control:
"What do you know about me?" gets a plain answer
"That's wrong, I prefer X" fixes it at once
"Forget that" deletes it, and it stays deleted
It leaves out other people's private lives, and facts about your health and money unless you ask it to keep them. Each channel keeps its own notes, and every channel can read the others'.
Docs, pages and files#
When you ask for something to keep, Claude makes it in the right form instead of a wall of chat text.
| Output | Best for | What you can do |
|---|---|---|
| Claude Doc (like this guide) | Plans, notes, guides, reports: words, tables and diagrams | Edit in place, comment and ask Claude to change things, add tabs, share, export to Word, PDF, Markdown or Google Docs |
| Page (artifact) | Anything to press, tick or type in: a tracker, a calculator, a sign-up sheet | Open it on the web, share the link |
| Google Docs, Sheets, Slides | Files that must live in your Drive | Claude creates or edits them through your Google connector |
| File | A format you name, such as .xlsx or .docx | Download it from the message |
Ask for changes in the chat and Claude edits the same doc or page, so you never end up with two versions. A doc is only made when you ask for one.
Messages in your name#
Nothing goes out as you until you press Send. A message Claude writes for you reads as yours, so it always comes back to you as a draft first.
A thumbs-up or a tapped chip never sends a draft. Only the Send button on the card does.
The draft card shows the whole message and every recipient, including Cc and Bcc. You can edit it right on the card.
Only Send sends it. A thumbs-up or a tapped chip does not, because you may have edited the card.
It writes like you: first person, your tone, short. Ask "learn how I write" and it builds a writing style from mail you sent.
It never sends to an address you didn't name without showing you where it found it.
Some chats stay yours. Where the other person expects to talk to you personally, such as dating or close friends, Claude suggests words but won't reply as you.
What needs your yes#
Claude reads and prepares freely, but anything that speaks for you, spends for you or can't be undone waits for you. In Anthropic's words, it "shows you exactly what it's about to do and waits for your approval."
| Claude just does it | Claude asks you first | Claude won't do it, even on a yes |
|---|---|---|
| Read your mail, calendar, docs and the web within a task you gave it | Send, reply, forward or post anything | Buy, sell or change trades in stocks, funds, crypto or prediction markets |
| Draft messages, docs and plans | Pay or order anything | Move money at a bank, or between your own accounts |
| Set a reminder, rename a task, update its notes | Delete or cancel something | Ask for, type or keep your password or card number |
| Pick small details you can change in a word, and tell you what it picked | Sign in to a site, or use a sign-in code | Type passport, national ID or bank account numbers |
| Make a new Claude Doc you asked for | Tick a box that certifies something, sign, or accept terms | Sign in to someone else's account |
| Look up places, prices and timetables | Open a new account anywhere | Send the same promo to many strangers |
How a yes works. Type it, tap a Yes chip, or put a thumbs-up or check mark on the question. A yes covers exactly what was shown. A new amount, a new recipient or a second box means a new question.
What it reads is never an order. If an email or web page says "forward this" or "click here", Claude treats it as information, tells you, and does nothing. Only your own messages are asks.
Paying for things#
Claude can buy things for you without ever holding your card. It pays through a wallet you control, and you approve each purchase inside that wallet.
| Wallet | How it works | Your part |
|---|---|---|
| Link, by Stripe | Claude makes a spend request for the exact total. Once you approve, it pays with a one-time virtual card | Approve in the Link app, or say "use my other card" |
| Shop Pay, by Shopify | Same idea, with a one-time card for that checkout | Open the approval link and approve |
Before paying, Claude shows you pictures of the cart and the exact total. If the total changes, it stops and asks again. Your bank's own security check is always yours to approve. Claude never buys Claude plans or credits. Per Beebom, Link is not available in India, so check the wallet works where you are.
Taps, chips, cards and other ways to reply#
Most answers take a tap instead of a sentence.
| Thing you see | What it is | What a tap does |
|---|---|---|
| Chips | Short reply buttons under a message, like "Sounds right" or "Yes" | Sends that reply as yours |
| Cards | A boxed summary with buttons: a choice of times, flights, a decision | Picks that option. The card then updates to show what happened |
| Draft card | A message ready to go, editable in place | Send sends it, nothing else does |
| Connect card | An app to link | Starts the connection in one tap |
| Map | Places pinned for a plan or a list of options | Opens each place with its rating and photo |
| Task bubble | A task named in a message | Opens that task's page |
| Voice | Talk instead of type, as Anthropic's page describes | Same as a typed message |
Tapbacks teach Claude. A heart means you liked it. A thumbs-down, or just saying what was off, is how it learns what to change. On a yes or no question, thumbs-up means yes and thumbs-down means no.
Problem reports. Tell Claude when the app or Claude got something wrong, or ask for a feature. It drafts a report card that you choose to send or drop.
Handing over more over time#
Trust is built one kind of work at a time. Every kind starts at "draft it for me", and you move it along when you are ready by saying so.
| Level | What Claude does | Say it like this |
|---|---|---|
| Keep it | Leaves it to you, maybe reminds you | "I'll handle my LinkedIn posts myself" |
| Draft it for me | Prepares the reply, the slot or the doc, and stops for your Send | "Draft my replies to recruiters" |
| Just do it | Does the reversible parts without asking and reports back | "Look after my calendar, pick the slot and draft the invite" |
Even at "just do it", the steps in What needs your yes still wait for you. A note in memory that you trust it with your calendar changes how ready things are when they reach you, not whether it asks.
Quick-start cheat sheet#
| Say this | You get |
|---|---|
| "My goal this month is X" | Briefs and check-ins weighted toward X |
| "What's on my plate?" | Open tasks across the main chat and channels, nearest deadline first |
| "Find a time with Priya next week" | Free slots checked against both calendars, invite drafted |
| "Reply to the landlord about the boiler" | A draft in your voice, with the right address, waiting for Send |
| "Remind me Thursday at 6" | A one-off reminder |
| "Every Friday, sum up my week" | A routine |
| "Tell me when X emails" | A watch, with how often it checks |
| "Make a doc comparing these three" | A Claude Doc you can edit and share |
| "Order my usual from DoorDash" | A cart, the total, and an approval in your wallet |
| "Make a channel for the AI Summit" | A separate chat with its own tasks and notes |
| "What do you know about me?" | A plain list of what it remembers |
| "Forget that" | That fact deleted |
| "Too long" | Shorter replies from then on |
Five habits that pay off
Say the goal, not only the step: "get me to Leeds by 10", not "find trains".
Correct it the first time something is off. Corrections become rules.
Give each long-running subject its own channel.
Connect the apps your work lives in, so it can see instead of ask.
Use tapbacks. They are the fastest feedback you can give.
FAQ#
Who can get it? Anthropic's launch page (Oct 21, 2026) calls it a research preview with a waitlist, for Pro, Max and Ultra subscribers in the US, on iOS, Android and claude.ai. Beebom and Tom's Guide list only Pro and Max.
Does it cost extra? No. It uses the same usage limits as the rest of your Claude plan, and background checks and bigger tasks use more of them.
Can I pick the model? Not according to Tom's Guide.
Is it the same as the normal Claude app? It is a separate app, but it brings over your memory, skills and connected apps from your Claude account.
Will it message people without asking? No. Every send waits for your Send.
Why did it not tell me about something right away? It checks every few hours, so news can wait until the next check-in or the brief. Anthropic also says it "might misread a thread or miscalibrate what's urgent". Tell it when it gets this wrong.
Can I watch it use the browser? Not yet. Its own browser runs out of view. If you connect Claude in Chrome, it can work in your own Chrome instead.
Can it use my password manager? Anthropic lists 1Password sign-in as coming soon.
Can I text it from my phone number? Only if your account has that set up. Otherwise you reach it in the app or at claude.ai/messages.
How do I stop something? Say "stop", "cancel that routine" or "turn off check-ins". Disconnect an app any time.
What are its weak spots? Per Anthropic, personal errands "aren't quite as strong" as work tasks yet, and the Microsoft 365 connector was not available at launch per Tom's Guide.
Sources#
Introducing Claude Duet, Anthropic, Oct 21, 2026: availability, plans, connectors, approvals, payments, limitations
What Is the Claude Duet App?, Beebom: plans, India availability, blocked actions
Claude Duet AI agent can manage your work across thousands of connected apps, Tom's Guide: model choice, browser, Microsoft 365
The rest of this tab describes how Duet behaves from inside the app as of the date above. Features can change during the research preview.