Anthropic opens the waiting list for Claude Duet, an agent who works alone and notifies when something requires your attention

Anthropic opens the waiting list for Claude Duet, an agent who works alone and notifies when something requires your attention

Do you remember the cute OpenAI Dots? Well, now Anthropic has just presented its proposal in this regard. Is called Claude Duet and it is an application in which Claude acts as a proactive agent, that is, without waiting for you to ask him for anything.

For now, it starts as a research preview with a waiting list, limited for now to plan subscribers Pro and Max in the United Statesand comes as an independent application for iOS and Android, in addition to being available on the web.

It works from the context you already have in your Claude account, because it inherits memory, skills and connected applications, so the initial configuration is minimal. From there, it sends a summary every morning with the most relevant features of those applications and, for the rest of the day, it checks the connected accounts every few hours to write to you only when it detects something that requires your intervention.

This interval is not coincidental, since a continuous check would exhaust the limits of the plan much sooner. Everything happens in a single thread of conversation, in writing or by voice, which can accommodate both a quick response and a multi-week project.

Keeps working even if you close your laptop and asks for permission before clicking send

Claude Duet works in the same way as the AI ​​agents we have been getting to know in recent months. It is even similar to OpenClaw, although in this case it is proactive. Plus, it works in the cloud, not on our PC. That’s why, It has a virtual machine with its own browserso tasks continue in the background without the need for your computer or mobile to be on.

You only receive a notification when it ends or when it requires your review. The distribution of responsibilities is marked, because he reads, writes and prepares on his own, although before sending a message, making a payment, signing a document or deleting something, he shows exactly what he is going to do and waits for your approval. Just in case, there are actions that are prohibited even with your approval, such as trading securities or moving money between accounts.

In terms of integrations, the application can draw on the more than 6,000 connectors and plugins on the Claude Marketplace, with examples ranging from locating pending messages in Gmail and Slack to reviewing unpaid invoices in QuickBooks or Stripe, preparing payrolls in Gusto or managing a Shopify store.

Purchases are made through Link, Stripe’s digital wallet, which requests approval of the expense and pays with a virtual card so as not to expose the real number. Anthropic also anticipates that later you will be able to log in to websites with the credentials saved in 1Password.

The company recognizes that it is an early version that can misinterpret a conversation or make a mistake when gauging what is urgent, that it is worse at personal errands than at work, and that it still can’t show you its browser. It also warns that its background reviews consume plan limits and that large tasks consume significantly more than a normal conversation.