Anthropic launches Claude Duet, an always-on agent that comes to you first

Anthropic has opened a research preview of Claude Duet, a standalone app that reframes the assistant as a proactive agent rather than a chat window. It ships on iOS and Android and on claude.ai, with invitations going out in waves from a waitlist.

Duet is included in Pro and Max plans at no extra cost, but it draws on the same usage limits as the rest of Claude. Anthropic warns that background work counts against those limits, so heavy users may hit them sooner than they would with chat alone. Team and Enterprise plans are excluded for now.

The product is organized around attention. Each morning Duet scans connected email and calendar accounts and sends a short brief on what most needs the user, and it can take first steps such as drafting replies. Through the day it follows those threads and messages only when something needs a decision, while larger jobs come back as research briefs, financial models, slide decks or dashboards.

Underneath sits a virtual machine with its own browser in the cloud, so work continues after the app is closed. Anthropic says Duet checks connected accounts every few hours rather than continuously, because constant polling would burn through plan limits much faster. Notifications land in the main conversation when a task finishes or needs review, and only the user can mark a task complete.

Duet is built around Claude Opus 5.5 and carries over memory, skills, connectors and past conversations from the user's Claude account. Anthropic frames its judgment — deciding what needs the user, what can wait and when to ask first — as the core of the product rather than any single capability.

Safety gets unusual space in the announcement. Credential handling follows the Claude Managed Agents design so Duet never sees passwords or sign-in tokens; actions are gated by the same auto mode classifier used for claude.ai and Claude Code; prompt-injection defenses come from Anthropic's browser-use research. The company says red teams and an external firm penetration tested the product, which is covered by its bug bounty.

A new classifier and safety system prompt check every message and block disclosure of the strictest tier of information, including credit card numbers and government IDs, while health or financial details pass only on explicit instruction. Duet also connects to Stripe to flag unpaid invoices and to Link and Shop Pay for purchases made with approval.

The competitive framing is explicit: Duet lands as Anthropic's answer to always-on agents such as Meta's Muse. The test will not be the demo but two harder numbers — how often Duet is right about when to interrupt, and whether subscribers keep paying once background work eats into their limits.