What is Claude Duet? Anthropic's proactive AI agent, explained

What it does in a day

Anthropic’s announcement aims Duet at professionals and small business owners with more to track than time to track it. Each morning it sends a short brief on what matters across your email, calendar and other connected apps, with drafted next steps. Through the day it checks those accounts every few hours, and you set what you want to hear about and how often. You give it work in a short message, typed or spoken, and it decides which apps to use and what steps to take, then tells you when the job is done or needs your review. Anthropic’s examples run from scheduling meetings and chasing unpaid invoices to preparing payroll for review, getting contracts ready to send, and building dashboards, forecasts, decks, prototypes and small web apps.

You talk to it in one ongoing thread, or in channels for separate topics that share one memory. Larger jobs come back as files in a Make tab, and a Tasks tab lists what it is working on and what needs your input; only you can mark a task complete. It brings your Claude account’s memory, skills and connected apps with it, and can search your past Claude chats unless you turn that setting off.

The computer behind it

Duet works on a cloud computer of its own, with a browser of its own, which is why it can keep going after you close the app. It uses the browser for sites that have no connector. Anthropic’s safety page adds one detail about the setup: the browser runs on its own dedicated machine, kept apart from the model and from the computers where the agent does its work. It says nothing else about them. You can’t watch its browser or take it over, and it can’t see your local files. There is no model picker: it runs on Claude Opus 5.5 and decides how much effort a job needs.

What it can reach, and when it asks first

Duet uses the connectors and plugins in the Claude Marketplace, more than 6,000 by Anthropic’s count. The ones it names include Gmail, Google Calendar, Google Drive, Slack, HubSpot, QuickBooks, Stripe, Gusto, Docusign, Calendly, LinkedIn, Canva and Shopify. Microsoft 365 isn’t available yet. You start each connection and can disconnect any app at any time, though disconnecting stops future access without erasing what Duet already remembers. It doesn’t make calls or send texts; with your approval it can send email from a connected account.

It reads, drafts and prepares on its own. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it will do, recipients included, and waits for your approval, and the same rule holds for routines that run while you’re away. Some actions are off-limits even if you approve them: trading securities, moving money between bank accounts, and getting into someone else’s account. Purchases go through Link, Stripe’s wallet, with a one-time card limited to that merchant and amount, or through Shopify’s Shop Pay, so your real card number isn’t shared. Signing in to sites with logins saved in 1Password is listed as coming soon.

Safety, and what Anthropic keeps

Anthropic’s safety page describes three layers beyond the model’s own training. Passwords are typed into a form in the app and never reach the model, and Duet never stores passwords, cookies or sign-in tokens. A separate classifier checks every outgoing message: card and bank numbers, government ID numbers and sign-in details are never shared, and health or financial details are shared only when you ask. A check that runs outside the working model judges whether each action matches what you asked for, so a prompt injection hidden in a web page or an email can’t approve an action on its own. Anthropic is plain that this isn’t complete: the safeguards “will occasionally miss something or flag a harmless message,” and Duet “may misread a thread or miscalibrate what’s urgent,” or miss something that arrives between its checks.

Duet keeps your conversation history, its memory and its task list until you delete the history or reset the memory, and deleting your Claude account removes everything it stored. Anthropic’s Duet pages don’t say how long that data is kept otherwise, or whether Duet’s conversations are used to train models.

Who can use it, and what it costs

As of October 26, 2026, Duet is a research preview by invitation. You join the waitlist on its product page, invitations go out in waves, and access is tied to the email you gave, so it should match your Claude account and the Apple or Google account on your phone. It is open to US accounts on:

• Claude Pro: $20 a month, or $17 a month billed yearly.
• Claude Max: from $100 a month, billed monthly.

Free accounts don’t get it, and Team and Enterprise plans are promised later. Duet costs nothing extra during the preview, but it draws on the same five-hour and weekly usage limits as the rest of Claude, at the same rate as chat, and its background work counts too: the morning brief, the checks every few hours, and the extra turns a big job takes. Anthropic gives each person it admits one full reset of those limits, usable until November 30, 2026.