Anthropic launches Claude Duet research preview for business automation

A proactive approach to work

Anthropic recently opened a research preview for Claude Duet, a new tool designed to function as a proactive agent for professionals and small business owners. Unlike traditional AI chatbots that wait for a user to type a prompt, this agent is designed to come to the user first. It determines what tasks require immediate attention and what can wait for later.

The system is built around the Claude Opus 5.5 model to help it decide on priorities across various connected applications. Every morning, the agent sends a short brief to the user regarding priority items. Throughout the day, it monitors those apps every few hours and only messages the user when something specific needs their attention.

Currently, the tool is available to US Pro, Max, and Ultra subscribers who join a waitlist. These users can access the standalone app on iOS and Android devices as well as on the claude.ai website. During this preview phase, the tool shares the existing usage limits of these specific subscription plans.

Background execution and tools

One of the core technical features of Claude Duet is its use of a virtual machine equipped with its own browser. This allows the agent to continue working on tasks even after a user has stepped away from their device. Because the work happens in a virtual environment, a user's laptop or phone does not need to stay powered on or connected to complete a task.

Users can describe their desired goals through voice or text in a single ongoing thread. The system then identifies which apps to use and maps out the necessary steps to complete the request. Once the agent finishes a task or hits a point where it needs a human to review the work, it sends a notification to the user.

The agent leverages over 6,000 connectors and plugins from the Claude Marketplace to interact with common business software. It can find missed messages in Gmail and Slack, check Intuit QuickBooks or Stripe for unpaid invoices, and draft reminders. It can also prepare Gusto payroll by filling in hours, bonuses, and reimbursements automatically.

Commerce and integrations

Beyond standard office communication, Claude Duet can manage complex commerce and marketing workflows. It can draft LinkedIn job postings, prepare Docusign agreements, and write follow-up emails in HubSpot. It also allows users to build social campaigns in Canva and manage various Shopify stores. For many small business owners, this centralizes several different platforms into one interface.

The agent can even perform purchases through Link, which is a digital wallet from Stripe. When a user confirms an item and approves the spend, the agent pays using a virtual card. This setup ensures that the user's actual credit card number remains unshared with the AI. It can also use Shop Pay to complete checkouts at various online stores.

Anthropic plans to expand these capabilities further in the near future. They have announced that a richer visual editor for Shopify storefronts is coming soon. Additionally, they will be adding permission-based sign-in support for 1Password to help manage secure access to various accounts.

Security and limitations

Despite its ability to work autonomously in the background, Anthropic has built in several gates to keep humans in control of major decisions. The agent can read, draft, and prepare work on its own, but it must wait for user approval before it can send a message, make a payment, sign a document, or delete any data. Certain high-risk actions, such as trading securities or moving money between accounts, are strictly off-limits.

The security architecture follows a design where the agent never directly sees passwords or sign-in tokens. An auto mode classifier approves actions, and the system uses browser-use research to defend against prompt injections. A dedicated safety system prompt also rejects requests to share sensitive information like credit card numbers or government identification.

While the tool is powerful for business, it has some current limitations. Anthropic notes that the agent may occasionally misread a thread or miscalibrate what is urgent. Because it is optimized for work, it is not as effective for personal errands like booking tickets. Additionally, the app cannot currently show its browser window and lives in a separate app from the main Claude interface.

Real-world testing and next steps

Early testers are already using the tool to streamline complex operations. One early tester, who owns a restaurant with 110 employees, used the preview for accounting, customer support, marketing, and store-by-store invoicing. He noted that the tool helps consolidate his workflow so he does not have to open hundreds of different tabs to manage his business.

Anthropic is continuing to refine the tool through red teaming, external penetration testing, and a bug bounty program. They are currently working on improving the system's ability to handle sensitive details like health or personal finances, which currently only pass through if a user explicitly instructs it. They are also working on better handling of the agent's autonomy.

The company has not yet specified standalone pricing for Claude Duet or an international release date. For now, the focus remains on the research preview for US subscribers. As the tool moves toward a wider release, users can expect more clarity on how the agent will handle complex, multi-step instructions across the growing library of marketplace connectors.

Try it

Check the Claude Marketplace to see which of the 6,000+ connectors are compatible with your current business tools.

Watch

The release of the visual Shopify storefront editor.

The addition of 1Password permission-based sign-in support.

Future pricing models for standalone Claude Duet access.