Claude Duet keeps working. Your usage limits still apply.

Anthropic’s Claude Duet can keep working while your devices are off, but its background checks consume the same allowance as the rest of Claude. The research preview starts with a US Pro and Max waitlist. My read: its value depends on whether the attention it saves justifies the allowance it consumes. [1] [2]

Who can use Claude Duet at launch?

Claude Duet is a research preview for Pro and Max subscribers in the US, through a waitlist. Anthropic’s product FAQ explicitly excludes Team and Enterprise for now. Those are the access terms I verified on 23 October 2026, rather than a promise of access to every paying Claude user. [2]

The launch happened on 21 October 2026. Both Anthropic’s announcement and the Claude Help Center release notes use that date; 23 October is this article’s publication date. Duet has separate iOS and Android apps and a web interface at claude.ai. Anthropic says it is powered by Opus 5.5. [1] [2] [3]

The model matters, but it is not the main change here. Opus 5.5’s capability and cost concern the work a model can complete. Duet adds a product decision on top: Claude can choose to start checking for useful work before you send another message. That changes who decides when part of your allowance gets used.

Does Claude Duet work with your laptop closed?

Yes. Anthropic says Duet has its own virtual machine and browser, so your laptop and phone can both be off while a task continues. It brings across your Claude memory, skills and connected apps, with text and voice messages in one continuing thread. [1]

That arrangement removes a practical dependency: you can leave a job running without maintaining a local session. But task execution and noticing new events are different functions. Anthropic says Duet checks connected accounts every few hours because constant checking would exhaust plan limits faster. A morning brief and an occasional useful interruption are the intended experience. [1]

I would treat it as periodic assistance when deciding which work to delegate. A meeting brief prepared before I return is useful even if it takes a while. An alert that must arrive immediately needs a different expectation. Anthropic’s FAQ acknowledges that Duet can miss something between checks and misjudge urgency. [2]

This is also where persistent assistants such as ChatGPT dots become interesting to compare. For me, continuity matters when the saved context makes the next useful task easier to finish, rather than simply keeping a conversation open.

Do Claude Duet’s background checks use your plan limits?

Yes. Anthropic says background checks count toward the same usage limits as the rest of Claude during the research preview. Larger jobs, including financial models and slide decks, require more model turns and background work. Inclusion in Pro or Max therefore gives you another way to spend the existing allowance. [2]

That makes a quiet agent different from an idle chat. You can stop typing while Duet keeps consuming usage to look for work. The product pages do not publish a fixed allowance cost per check, so they cannot tell us how much chat or coding capacity a day of monitoring will leave. The documented interval alone is not enough to calculate it.

I would judge that trade-off by the interruptions and decisions it removes. If Duet finds a missed deadline early enough to help, some background usage can be a good exchange. If it repeatedly summarizes low-priority messages, the same spending is harder to justify. That is my evaluation criterion, not a measured result from the preview.

The distinction between access and capacity also matters when reading Claude’s other usage-limit changes. A feature being included answers whether you can use it; the shared allowance determines how much room it has alongside your existing work.

What can Duet do without your approval?

Duet can read, draft and prepare work independently, but Anthropic says it waits for approval before sending a message, making a payment, signing a document or deleting anything. Trading securities and moving money between accounts remain off-limits even with approval. Preparation can continue while you are away; those decisions still require you. [1]

The approval boundary makes the proposed action the useful thing to inspect. I would check the recipient and final wording of an email, or the amount and purpose of a payment, before approving it. A well-written summary of the plan is less useful than seeing exactly what will happen.

Anthropic’s announcement says Duet presents the action before asking. The same announcement also says the preview cannot yet show you its browser. Those are different levels of visibility: reviewing a proposed action is available, while watching its browser session is not. [1]

I would start with preparation I can verify

My first Duet task would be a narrow brief from a small set of connected sources, with a clear deadline and an output I can check. That would let me judge whether background work saves time before making it responsible for more of the day.

For example, I would ask for a client-meeting brief with links to the relevant messages, open questions and a draft follow-up. I would compare the result with the source material and note whether the brief arrived when useful. I would also watch how much plan capacity remained for my normal Claude work. These are proposed checks; I have not tested Duet hands-on.

I would keep context management deliberate even when the assistant brings its own memory. Knowing which sources a task should use makes both the result and the review easier to understand.

The work I would give Duet next depends on that first result. If I can trust a brief after a quick source check and still have enough allowance for the work I do myself, I have a reason to delegate more. If reviewing its choices becomes another morning task, I would narrow what it watches.