What is Claude Duet, and how do you control a proactive AI agent that messages you first?

Anthropic’s Claude Duet, launched October 21, 2026 as a research preview for U.S. Pro and Max subscribers, runs on its own cloud machine, checks your connected apps every few hours and sends a morning brief — and it must ask before it sends, pays, signs or deletes anything. Here is how to control it.

You are used to AI that waits politely for a prompt. Claude Duet is the opposite — it starts the conversation. Short answer: Claude Duet is a proactive AI agent from Anthropic, launched October 21, 2026 as a research preview with a waitlist for U.S.

Pro and Max subscribers on iOS, Android and claude.ai. It runs on Claude Opus 5.5 on its own cloud virtual machine with a browser, so it keeps working while your phone and computer are off. Each morning it sends a brief of what needs attention across your connected apps, and it re-checks those apps every few hours — not in real time, because constant checking would drain your plan’s usage limits, which Duet counts against like regular Claude use.

Why it matters: Duet plugs into more than 6,000 app connectors and carries over your existing Claude memory and skills, so it starts out knowing your projects. The safeguard that makes it usable: it must ask your approval before it sends, pays, signs or deletes anything.1. Join the waitlist.

Open Claude on your phone or claude.ai and sign up for the Duet preview with a Pro or Max plan.2. Choose which apps it may watch. Connect the few apps your morning brief truly needs — email, calendar, documents — and leave the rest disconnected.3.

Set what it messages you about. Pick the topics worth an interruption, and let everything else batch into the morning brief.4. Review its activity log.

Skim what it read, drafted and proposed each day — especially in week one, while you are calibrating trust.5. Revoke a connector any time. If an app stops earning its place, disconnect it in settings and Duet’s access ends.Pro tip: Start with read-only connectors, and enable anything that can pay or send only after a week of watching its briefs — trust the log, not the launch video.Ad’s take: An assistant that speaks first is only helpful if it also knows when to stay quiet.

The approval step before anything sends, pays, signs or deletes is what makes me willing to try it.