Anthropic opened a research preview of Claude Duet, a new application that turns its Claude AI into a persistent, proactive agent capable of monitoring business tools, triaging priorities, and executing tasks in the background without requiring a device to stay on.
The app is available as a standalone download on iOS and Android and through claude.ai. During the research preview, it is included in Pro and Max subscriptions and draws from the same usage limits. U.S.-based Pro and Max subscribers can join a waitlist for access.
An agent that comes to you first
Claude Duet is built around Claude Opus 5.5, Anthropic’s latest flagship model, and is designed to flip the typical chatbot interaction. Rather than waiting for a prompt, it checks connected apps every few hours and sends a morning brief highlighting the most important items across a user’s accounts. Throughout the day, it messages the user only when something genuinely requires attention.
Users can also hand it tasks ranging from scheduling meetings to building financial models, sales forecasts, or prototype web apps. Because Claude Duet runs on a dedicated virtual machine with its own browser, work continues even when a phone or laptop is off. The user receives a notification when the task is finished or needs review.
Over 6,000 integrations through the Claude Marketplace
The app taps into Anthropic’s Claude Marketplace, which offers more than 6,000 connectors and plugins. The announced integrations span a wide range of business software suc has Gmail and Slack for communications, Intuit’s QuickBooks and Stripe for invoicing, Gusto for payroll, LinkedIn for job postings, DocuSign for agreements;, HubSpot for CRM follow-ups, Canva for marketing assets, and Shopify for storefront management.
Claude Duet can also make purchases on a user’s behalf through Stripe’s Link digital wallet and Shopify’s Shop Pay, using a virtual card so the real card number is never shared. Anthropic says 1Password integration for saved website logins is coming soon.
Guardrails on consequential actions
Before Claude Duet sends a message, makes a payment, signs a document, or deletes anything, it shows the user exactly what it plans to do and waits for explicit approval. Certain actions like trading securities and moving money between accounts are blocked entirely, even with approval. Users can also disconnect any app from Claude Duet at any time.
Anthropic says the agent’s safety architecture draws on several of its existing frameworks. Credential handling follows the design used for Claude Managed Agents, so the system never directly sees passwords or sign-in tokens. An auto-mode classifier, the same safeguard used in Claude Code, prevents the agent from taking steps the user didn’t request. Prompt-injection defenses are derived from Anthropic’s browser-use research.
A new dedicated classifier and safety system prompt guard against leaks of sensitive information. The strictest tier — credit card numbers and government identification — is always blocked from being shared. Other sensitive categories, such as health or personal financial data, are transmitted only when the user has explicitly instructed it.
Red teams and an external firm have penetration-tested Claude Duet’s security infrastructure, and it is covered by Anthropic’s public bug bounty program on HackerOne.
Early usage and known limitations
Anthropic acknowledges the product has limitations. It may misread a conversation thread or misjudge what’s urgent. Its core functionality is oriented around work, so personal errands like booking tickets aren’t as capable yet. Users also can’t currently view the agent’s browser session. Background monitoring and larger tasks consume plan usage at a higher rate than a typical Claude conversation.
For now, Claude Duet lives in its own app, separate from the main Claude experience. Anthropic says what it learns from early users will shape how the two eventually work together.
Leave a Reply