Claude Duet: Anthropic’s always-on agent for your apps

A research preview that keeps working after you close the app, and asks before sending, paying, signing or deleting.

On 21 October 2026 Anthropic introduced Claude Duet, an agent that works across your connected apps and keeps going in the background after you close it. It is a research preview for Pro and Max subscribers in the US, and it is built to ask before it sends messages, makes payments, signs documents or deletes anything.

Key facts

Announced by Anthropic on 21 October 2026 as a research preview.
Rolling out in waves to Pro and Max subscribers in the US, through a waitlist.
Available as iOS and Android apps and on claude.ai for invited users.
Included in the plan, but background work counts toward usage limits.
Designed to ask before sending a message, making a payment, signing a document or deleting anything.

What happened

Claude Duet is a new way to use Claude: instead of answering one message at a time, it works as an agent that stays on. It runs on its own computer in the cloud, so it can keep working on a task after you close the app. It can follow your email and calendar, act inside the apps you have connected, and use its own browser for websites that do not have a connector.

Duet shares memories, skills and connectors with the regular Claude app and can search your past conversations, so it starts with the context you have already built. A Tasks tab shows what it is working on, and only you can mark a task as finished. Integrations mentioned at launch include Stripe, for checking unpaid invoices and drafting payment reminders, and Link and Shop Pay, for purchases that you approve.

Access is gradual. Anthropic is sending invitations in waves to Pro and Max subscribers in the US. Team and Enterprise plans are not included yet, though Anthropic plans to add them later.

Why it matters

Most people still use AI like a very smart search box: ask, read, copy. An always-on agent changes the rhythm. You describe a goal once and the agent keeps moving it forward, checking in when it needs a decision.

That makes the quality of your instructions more important, not less. A vague request to a chat costs you one bad answer. A vague request to an agent that works for hours can cost you a messy inbox or a wrong reminder drafted for a client. The safeguards help, since Duet asks before messages, payments, signatures and deletions, but the clearer the brief, the fewer interruptions and corrections you get.

There is also a practical cost: background work uses your plan’s limits, so it pays to give the agent focused tasks rather than open-ended ones.

How to try it

If you do not have access yet, you can still practise the habit that makes agents work well: writing a clear weekly brief. Open our free prompt Plan my week and fill the brackets with three things: your fixed commitments (meetings, school runs, deadlines), your goals for the week, and your limits (for example “no work after 7 pm”).

A filled request could look like this: “Fixed: team call Monday 10:00, dentist Wednesday 15:00. Goals: finish the client proposal, two gym sessions, reply to all supplier emails. Limits: no meetings on Friday afternoon.”

Paste it into Claude, ChatGPT or Gemini and check three things in the answer: that every fixed commitment is in the right slot, that there is free buffer time each day, and that each day has one clear top priority. The same brief, with your apps connected, is exactly the kind of instruction an agent like Duet needs.

Try it

Plan my week

Frequently asked questions

Who can use Claude Duet?

At launch it is rolling out to Pro and Max subscribers in the US, in waves and through a waitlist. Team and Enterprise plans are not included yet, according to the launch coverage.

Does Claude Duet cost extra?

It is included in the Pro and Max plans at no extra cost, but the work it does in the background counts toward your plan's usage limits, so you may reach them sooner than with chat alone.

Can Claude Duet send emails or pay on its own?

Anthropic says Duet is designed to ask before it sends a message, makes a payment, signs a document or deletes anything, so those actions wait for your approval.

Sources