---
title: "What can you hand off to Claude Duet? Compared with dots"
description: "Claude Duet, a Claude agent in research preview since October 21, 2026, reaches out first and keeps working while you're away. How it works, approvals, usage limits, gaps and dots."
url: https://udon-dev.com/en/notes/claude-duet/
lang: en
published: 2026-10-25
topic: [ai]
tags: ["claude", "anthropic", "ai-agent", "claude-duet", "dots"]
---

# What can you hand off to Claude Duet? Compared with dots

- Claude Duet is an always-on agent that Anthropic opened as a research preview on October 21, 2026. It runs on Claude Opus 5.5 and, through a morning summary and check-ins every few hours, contacts you only when it needs you
- It carries over your Claude memory, skills and connectors, and works in its own virtual machine and browser in the cloud. You use it through standalone iOS and Android apps and claude.ai
- It always asks for approval before sending a message, making a payment, signing a document or deleting something. It won't trade securities or transfer money between accounts, even with approval
- It is for Pro and Max users in the US, by invitation from a waitlist. Team and Enterprise aren't covered yet. There is no extra charge, but everything it does, including background check-ins, counts toward your plan's usage limits
- The biggest differences from dots: you can't take over its browser, you can't connect your own computer, you can't write per-action permission rules, and it counts toward usage limits

On October 21, 2026, Anthropic [opened a research preview](https://claude.com/resources/articles/claude-duet) of Claude Duet. Instead of waiting for you to start a conversation, Claude reaches out with "take a look at this," and it keeps working while you're away. Anthropic says it built Duet for professionals and small business owners who can't keep track of everything on their plate.

（図: Image from the Claude Duet product page）
Source: [Claude Duet (Claude official site)](https://claude.com/product/claude-duet)

[MacRumors](https://www.macrumors.com/2026/10/21/anthropic-claude-duet/) framed Duet as a response to Meta's Muse and OpenAI's [dots](/en/notes/openai-dots/). OpenAI announced dots on September 29, and I covered how they work on this blog.

## How it differs from a Claude chat

The [FAQ on the product page](https://claude.com/product/claude-duet) explains the difference in terms of who speaks first. The Claude app waits for you to open a chat, and you work alongside Claude. Duet doesn't wait; it sends you messages on its own. According to [the announcement](https://claude.com/resources/articles/claude-duet), every morning it sends a short summary of what matters in your connected apps. After that, it checks those apps every few hours and messages you only when something needs your decision.

You send requests by text or voice in a single ongoing thread. The examples range from admin work such as scheduling meetings to building a dashboard that uses dozens of tools, forecasting finances and sales, and prototyping a web app. According to the [Help Center](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet), you can also create separate channels by topic, and memory is shared across all channels. Work in progress is listed in the "Tasks" tab, and finished documents, spreadsheets and slides appear in the "Make" tab. Duet may tell you a task looks done, but only you can mark it complete.

## A cloud virtual machine, and the settings it carries over

Duet works in a virtual machine in the cloud that has its own browser. Work continues even when your computer and phone are turned off. You describe a task in a short message, Duet decides which apps to use and in what order, and you get a notification when it's done or when it needs you to confirm something.

Because Duet uses your Claude account as is, the memory, skills and connectors you set up in Claude are available from the first message. It also searches past chats to pick up context from earlier work. The search covers Duet chats and Claude chats that aren't in a project. Turning off "Search and reference chats" in Claude's settings stops this search.

It can connect to more than 6,000 connectors and plugins in Claude's marketplace. Anthropic's examples include finding missed messages in Gmail and Slack and helping you reply, and finding unpaid invoices in QuickBooks or Stripe and drafting reminders. Preparing payroll in Gusto and sending contracts out for signature in Docusign are also among the examples. For sites without a connector, it uses the cloud browser.

For purchases, it can pay with Link, Stripe's wallet, or Shopify's Shop Pay. Both ask for approval on every purchase and pay with a single-use card number valid only for that purchase, so your real card number never reaches the merchant.

## What needs approval, and what it won't do

Duet handles reading, drafting and preparing on its own. For four actions (sending a message, making a payment, signing a document and deleting something) it shows you exactly what it will do and waits for your approval. If you can't approve right away, it finishes the rest of the work first and brings the approval request back later. It won't trade securities, transfer money between accounts or access other people's accounts, even with your approval.

[The announcement](https://claude.com/resources/articles/claude-duet) says the safety mechanisms come from existing products. Whether an action is allowed is decided by the "auto mode" classifier also used in Claude Code. Credentials are handled with the same design as Claude Managed Agents, and Duet never sees passwords or login tokens directly. When a site asks for a password, you type it into a dedicated field on your own device, and the value goes straight to the browser. The Help Center says this is designed so the password doesn't remain in your chat with Duet.

What's new for Duet is a classifier that prevents sensitive information from leaking, plus a safety system prompt. It checks the content of each message and refuses requests to share information in the strictest category, such as credit card numbers and government IDs. It passes health and personal financial information only when you explicitly tell it to send it.

On prompt injection (attacks that hide instructions in emails or web pages to make the agent follow them), the Help Center says Duet is built to treat what it reads as information. It then notes that no defense is perfect, and gives this as one reason Duet asks for approval before high-impact actions.

## Pricing and usage limits

During the research preview, Duet is included in Pro and Max plans at no extra charge. However, according to the [Help Center](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet), its usage limits are shared with the rest of your Claude usage. It uses both the 5-hour limit and the weekly limit at the same rate as Claude chats.

The morning summary and the check-ins every few hours also count toward the limits, so unlike chat, your limits go down even at times when you haven't asked for anything. Larger jobs such as financial models and slide decks involve more model calls and background work. Anthropic writes that it set the check-in interval to every few hours so you don't use up your limits too quickly. When you're invited to Duet, you can fully reset your 5-hour and weekly limits once. This reset is available until November 30, 2026.

## What it can't do yet

The [Help Center](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet) lists these limitations for the research preview:

- You can't choose the model. It runs on Claude Opus 5.5, and Duet decides how much effort to put into each task
- You can't watch its browser work or take over control
- Because it checks at set intervals, something urgent can arrive between check-ins
- It can't see all of your projects or files on your computer. It can read your work in Claude but can't change it
- It doesn't make phone calls or send text messages to other people. With your approval, it can send email from a connected account
- Some connectors, such as Slack, have limited functionality. Microsoft 365 isn't available yet

The announcement also says Duet may misread a thread or misjudge whether something is urgent. Because it was built mainly for work, it isn't good yet at personal errands such as booking tickets. For now, Duet is a separate app from the Claude app, and Anthropic says it will decide how to bring the two together based on feedback from early users.

## How it differs from dots

dots are also agents that keep working between conversations on a dedicated cloud computer and reach out only when they need a decision. The goals are similar; the differences show up in how finely you can delegate and where you can use them. The dots specifications are in [Meet dots](https://learn.chatgpt.com/docs/dots) and [Control your dot](https://learn.chatgpt.com/docs/dots/controls).

| Point of comparison | Claude Duet | dots |
|---|---|---|
| Model | Claude Opus 5.5 (not selectable) | GPT-6 Astra |
| Where you talk to it | Dedicated iOS and Android apps, claude.ai | ChatGPT desktop and mobile apps, desktop browser, Slack, Microsoft Teams. Voice calls are also available inside ChatGPT |
| Browser during work | Can't watch it or take over | You can open the screen and take over with "Take over" |
| Your own computer | Can't connect (can't see your files either) | Connect one computer and let it use local files and apps |
| Actions that need confirmation | Sending, paying, signing and deleting always need approval. Won't trade securities or transfer money | An automated review decides. You can set custom rules at four levels for each action |
| Usage limits | Counts toward plan limits, including background check-ins | Conversations with dots don't count (work started in Work or Codex does) |
| Plans | Pro and Max in the US (by invitation). Team and Enterprise not covered | Pro 100, Pro 200, Pro 500, Business Premium, Enterprise. Individual Pro excludes the EEA, the UK and Switzerland |

What stands out on Duet's side of the table is how little you get to decide. With dots, you can set how each action is handled, for example confirming before sending a message to a customer and leaving deletion of shared files to you. With Duet, Anthropic has decided in advance which actions need approval and which it won't do at all. You can tell Duet how often to contact you and what to tell you about, but I couldn't find any setting in the announcement or the Help Center that widens what it can do without approval.

The other difference is how it connects to work tools. dots can be called from Teams and can use your own computer. Duet, as the Help Center's list of limitations shows, doesn't connect to Microsoft 365 yet. At companies that work in Microsoft 365, the range of work you can hand off to Duet today is likely to be narrow.

## Can you use it from Japan?

Not right now. The Help Center lists three requirements: a Pro or Max plan, a US Claude account, and an invitation from the waitlist. Invitations go out in stages, and joining the waitlist doesn't guarantee a date. Anthropic says it plans to offer Duet to Team and Enterprise later.

About all users in Japan can do for now is get the memory, skills and connectors that Duet carries over in order. Connectors you've set up and skills you've made in [Claude Cowork](/en/notes/claude-cowork-feature-map/) and elsewhere will be used as they are once Duet becomes available.