FR Newsletter

Automate

Claude Duet

Anthropic's proactive agent: it starts the conversation and asks before acting.

Official siteVendor information

What the publisher documents

A proactive agent from Anthropic, in research preview since 21 October 2026. It lives in its own app, separate from the main Claude app, on iOS, Android and claude.ai. What separates it from a chatbot is the direction of the conversation: it starts it. It checks the accounts you connect at set intervals, sends a morning brief across email and calendar, and follows up on threads that have gone quiet.

It runs on Claude Opus 5.5 and you cannot choose another model. It works from its own computer in the cloud, with its own browser, so a task does not stop when your phone is off. It arrives with the memory, skills and connected apps of your Claude account from the first message, can use the connectors and plugins of the Claude marketplace, from Gmail and Slack to QuickBooks and Shopify, and can search its own chats as well as your Claude chats outside projects. It can read your work in Claude but cannot change it.

Your data

Approval is the design, not a setting. Before it sends a message, makes a payment, signs a document or deletes anything, Anthropic documents that it shows you the action, the destination and the content, every recipient included, then waits. A second check runs outside the model and tests whether the action matches what you asked for, which is also what makes it resistant to an instruction hidden in a web page: text on that page cannot change the outcome of a check the model does not run.

Some actions stay closed even with your approval: trading securities, moving money between accounts, and reaching into someone else's account. A classifier is instructed never to share sign-in credentials, full card and bank account numbers or government ID numbers, in Anthropic's words even if you ask, and to share health or financial details only with your explicit approval; when a message is blocked, it can rewrite it without the sensitive part and show you the new version.

Passwords never reach the model. You type them into a form inside the Claude app, encrypted with a single-use key that exists only on the cloud browser's machine; the model sees neither passwords, nor browser cookies, nor sign-in tokens, and cannot save them to memory. One sentence from that page is worth keeping as a defence against phishing: the sign-in form never arrives as a link, a text message, an email, or a website other than claude.ai.

On retention, the documentation is clear about what disconnecting does not do. Claude Duet keeps its conversation history and its memory across tasks until you delete that history or reset the memory; deleting your account removes everything it stored; and disconnecting an app stops further access but leaves what it already learned in memory. To stop it reading your other Claude chats, the switch is the one Claude already had, Search and reference chats. Whether Duet conversations follow the same training rules as the rest of Claude is not stated on the pages we read.

Plans and what you own

Research preview, by waitlist, in the United States only, on Pro, Max and Ultra. Anthropic states it is not available on Team or Enterprise plans yet and costs nothing extra during the preview. It draws on the same usage limits as the rest of Claude, the five-hour sessions and the weekly cap, and the work it does in the background counts against them.

Documented limits

  • United States only, by waitlist, during the research preview; not available on Team or Enterprise plans.
  • It checks your accounts at intervals, and Anthropic warns that something urgent can arrive between two checks. The interval is not published.
  • You cannot watch its browser or take it over.
  • It does not make phone calls and does not send text messages to other people. It can send emails from a connected account, with your approval.
  • Its background work eats into your plan's usage limits, so you can reach them sooner than with chat alone.
  • Anthropic writes that the safeguards will occasionally miss something or flag a harmless message, and that approval frequency, the sign-in method and the accuracy of those checks are all still changing.
  • It can misread a thread or misjudge what is urgent, which is why drafts and approval requests are worth reading before confirming.
  • Microsoft 365 is not available yet, and some connectors, Slack among them, have documented limits.

What we still have to verify

We have not tested Claude Duet: the research preview is open in the United States only, and we are not there, so nothing here speaks to quality. What a test would have to settle: the real interval between two checks, which is not published; whether Duet conversations fall under the same training rules as the rest of a Claude account, which the pages we read do not say; and how often it asks for approval, since Anthropic states it is still calibrating that.

Sources

  1. Anthropic, Introducing Claude Duet
  2. Anthropic, Claude Duet Safety, Privacy, and Security
  3. Claude Help Center, Get started with Claude Duet

Publisher pages read on November 1, 2026. We do not publish prices: they change often and the official page is the only reliable one.

Our test

We have not tested Claude Duet yet. This page only reports what the vendor publishes. Our screenshots and verdict will appear here after the test.

Screenshot coming after our hands-on test
Claude Duet : 1/3
Screenshot coming after our hands-on test
Claude Duet : 2/3
Screenshot coming after our hands-on test
Claude Duet : 3/3

Alternatives

Guides that use it