Claude Duet is a proactive agent that messages you first: what Anthropic says it does

Claude Duet is a new Anthropic app in which Claude acts as a proactive agent: it sends you a morning brief, checks your connected apps every few hours and messages you only when something needs attention. Anthropic opened it as a research preview on October 21, 2026, by waitlist, for US Pro and Max subscribers.

That is the short version of what Claude Duet does. The longer version matters more, because an agent that contacts you first and acts on your accounts raises different questions from a chatbot you open on purpose.

What Anthropic says Claude Duet does

Anthropic's announcement describes the app as aimed at "professionals and small business owners with more on their plates than they can track," according to 9to5Google's report. Three behaviors define it.

• A daily brief. Each morning it summarizes the most important items across connected apps such as email and calendar.
• Background monitoring. During the day it checks those accounts every few hours and sends a message only when something needs attention.
• Work that continues without you. It runs on its own virtual machine with a browser, so your devices do not need to stay on, and you get a notification when a task finishes or needs review.

MacRumors reports the app is built on Claude Opus 5.5 and that everything happens in one ongoing thread. 9to5Mac adds that larger jobs come back as editable files in a Make tab, that a Tasks tab tracks ongoing work and that only the user can mark a task finished.

The approval rules

The design choice that matters most is where the agent must stop. Anthropic's page says Duet drafts and prepares work on its own but asks before sending messages, making payments, signing documents or deleting anything. Two actions are off-limits even with approval: trading securities and moving money between accounts.

Purchases are possible. Anthropic says Duet can buy through Stripe's Link wallet using a virtual card, so your real card number is not shared. 9to5Mac also lists Shop Pay support, with the user approving each purchase.

How Anthropic says it handles security

Anthropic says Duet uses its trustworthy-agent framework, an auto mode classifier that approves actions, and a dedicated classifier that blocks sharing the most sensitive data, such as credit card and government ID numbers. It also says Duet never directly sees passwords or sign-in tokens.

Those are design claims from the vendor. Nobody outside Anthropic has published a test of how well the classifiers hold up against a hostile email or web page that tries to steer the agent, which is the standard attack on any tool that reads your inbox and can act on it.

Where it fits in the agent race

Coverage frames Duet as Anthropic's answer to Meta's Muse and OpenAI's Dots, which also reach out before you ask. 9to5Google says OpenAI launched Dots a few weeks earlier. The pattern is clear: the labs are moving from assistants you consult to agents that watch on your behalf.

The practical gap between those products and a chatbot is the loop. A chatbot waits. A proactive agent decides when to speak, which means its judgment about urgency becomes the product. Anthropic concedes the weak spot: it says Duet may misjudge urgency, is weaker at personal errands such as booking tickets, and cannot yet show you its browser.

For background on how these systems are assembled, see our explainer on the agent harness, which covers the tools, memory and controls around a model, and what counts as an AI agent.

What it costs and who can use it

There is no separate price. Access is tied to Pro and Max subscriptions, and usage counts against existing plan limits. MacRumors and 9to5Mac both note that background checks count toward those limits, so Duet can reach them sooner than chat alone. Team and Enterprise plans are not included yet, and Anthropic says invitations go out in waves.

Sources differ on platform details. Anthropic's page lists a standalone iOS and Android app plus claude.ai, while MacRumors mentions a standalone iOS app. Anthropic's own page is the one to trust, and details may change during the preview.

What we would watch

Three things will tell us whether Duet is useful or merely novel.

• Alert quality. How often does it interrupt you for something that did not matter, and how often does it miss something that did? Anthropic has promised to keep tuning based on early users.
• The approval prompt. An approval step protects you only if you read it. Agents that ask too often train people to click through.
• Independent security testing. The first red-team write-ups on prompt injection against an agent with inbox access will say more than the launch page.

The tension is plain. The more an agent does without being asked, the more it is worth, and the more a single mistake costs. Anthropic has drawn its lines at money movement and trading, which is sensible. What it has not shown yet is evidence that the lines hold when someone is trying to push the agent across them.

How it compares with a normal chat

In ordinary Claude chat you start the conversation. With Duet, per Anthropic's page, the agent can start it, using memory, skills and connectors shared with your regular Claude account. MacRumors adds that it can search past conversations. The page lists more than 6,000 connectors and plugins, including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva and Shopify. That reach is the appeal and the exposure in one list.

Our position

A research preview behind a waitlist is the right way to ship this. Our advice for anyone who gets an invitation: connect the minimum, keep sensitive accounts out until others have tested it, and read every approval. We will update this piece when independent evaluations appear.