Anthropic has introduced Claude Duet, an application in which Claude operates as a proactive agent. Available as a research preview on iOS, Android, and claude.ai, it is currently accessible only via a waitlist, to Pro and Max subscribers in the United States. Duet is included in these subscriptions and shares their usage limits.
An agent that flags what needs attention
Claude Duet is an artificial intelligence agent that works within the user’s connected applications, such as their email, calendar, or management tools. Anthropic intends it for professionals and small business leaders. It fulfills two functions: independently spotting topics that require action, and carrying out tasks assigned to it, even when the user is not in front of their screen.
Concretely, Duet sends a short summary each morning of the important items spotted in the connected applications. It then follows these topics throughout the day and only contacts the user when a topic warrants it. This ability to take initiative distinguishes it from Claude Cowork, which also works in the background, but only on tasks the user assigns to it.
Among the tasks that can be delegated to it, Anthropic mentions scheduling meetings, creating dashboards, developing financial or sales forecasts, as well as prototyping and creating web applications.
An application linked to the Claude account
Duet is used like a messaging app. The user writes or speaks to it in a single conversation thread, and it is through this same thread that the agent contacts them when a topic requires it. Getting started requires little setup, as Duet retrieves the memory, skills, and connected applications from the Claude account.
Once a task is assigned, the agent executes it on its own virtual machine, equipped with a browser. The user’s device therefore does not need to stay on, and a notification alerts them when the task is complete. To identify topics to flag, Duet checks the connected accounts every few hours, not continuously, to respect usage limits.
The agent has access to over 6,000 connectors and plugins from the Claude Marketplace, including Gmail, Slack, HubSpot, and Shopify.
Mandatory approvals and a filter for sensitive data
Duet acts on the user’s behalf in their tools: it can read their emails, send messages, or make purchases. Anthropic has therefore implemented several safeguards. Before sending a message, making a payment, signing a document, or deleting an item, the agent presents the planned action and waits for the user’s approval. Some operations remain prohibited even with this approval, such as stock trading or transfers between accounts.
Purchases go through Link, Stripe’s digital wallet. The expense must be approved, and the payment is made with a virtual card, without transmitting the actual card number. Credentials are managed according to the Claude Managed Agents model, so Duet never directly sees passwords.
Actions go through the classifier of Claude Code’s automatic mode, and the application has undergone penetration testing, conducted internally as well as by an external firm. A new classifier also checks each message to prevent leaks of sensitive information. It blocks the sharing of the most critical data, such as bank card numbers and identity documents. Other sensitive information, related to health or personal finances, is only transmitted upon explicit instruction from the user.
