Anthropic 向美国 Pro 和 Max 用户开放 Claude Duet 预览

Anthropic has opened a research preview of Claude Duet, a proactive agent that surfaces work needing attention, carries out assigned tasks in the background, and asks before acting on a user’s behalf. Access is available by waitlist to US Claude Pro and Max subscribers. Duet runs as a standalone app on iOS and Android and on the Claude website, while remaining separate from the main Claude app during the preview. It is included in those plans and shares their existing usage limits.

Built around Claude Opus 5.5, Duet is aimed at professionals and small business owners managing crowded workloads. It sends a concise morning brief based on connected apps, checks those accounts every few hours, and messages the user when it judges that something needs attention. Tasks can range from meeting scheduling to dashboards, financial or sales forecasts, prototypes, and web apps. One ongoing text or voice thread handles both quick requests and weeklong projects, while a virtual machine with its own browser keeps working when the user’s devices are off.

The service carries over memory, skills, and connected apps from existing Claude accounts and can use more than 6,000 connectors and plugins in the Claude Marketplace. Announced use cases span Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva, and Shopify. Duet can also make business purchases through Stripe’s Link or Shopify’s Shop Pay, with approval steps and virtual-card handling for Link. Support for 1Password logins and richer Shopify visual editing is planned.

Anthropic is positioning human review as a core boundary. Duet can read, draft, and prepare work independently, but must show the proposed action and wait before sending a message, making a payment, signing a document, or deleting anything. Securities trading and transfers between accounts remain prohibited. Its credential design follows Claude Managed Agents, while action controls draw on the classifier used for Claude and Claude Code. A separate classifier and safety prompt blocks sharing credit card numbers and government identification, while other sensitive material such as health or personal finance information requires explicit instructions. Anthropic says red teams and an external security firm tested the system, which is also covered by its bug bounty program.

The preview remains limited. Duet may misread threads or rank urgency poorly, personal errands are less developed, and users cannot yet view its browser. Background checks consume plan limits, while larger projects require more model turns. Anthropic says early feedback will guide how Duet and the main Claude app work together.

Earlier, we also reported that Anthropic is working on its own proactive agent, and here it comes!

Sources and related context

• Claude Duet product page and waitlist: Supports: The product page confirms the US Pro and Max preview, Opus 5.5, morning briefs, and carryover of memory, skills, and connected apps.
• Scaling Managed Agents: Decoupling the brain from the hands: Related context: Anthropic explains the credential-isolation architecture behind Managed Agents, the design cited for Duet’s handling of sign-in credentials.