Claude Duet does not wait for a prompt. That changes the job of an AI assistant

What is Claude Duet?

Anthropic announced Claude Duet on October 21, 2026 as a proactive agent for professionals and small businesses. Instead of waiting inside a chat, it checks connected accounts at intervals, sends a morning brief, follows work through the day and starts tasks that you ask it to handle.

Duet keeps one main conversation for quick requests and longer projects. Separate channels can organize topics while sharing the same memory. A Tasks tab shows ongoing jobs and requests for input, while a Make tab holds larger outputs such as research briefs, financial models, presentations and dashboards.

The app runs on a cloud virtual machine with its own browser. It can continue after your phone or laptop turns off, use connected apps when a connector exists and operate a website through its browser when one does not.

Why this launch matters

Most assistant products compete on answer quality. Duet competes on initiative. It decides what deserves attention, when to notify you, which tools a task needs and when a decision must come back to you. A strong model is necessary, but interruption quality and restraint now matter just as much.

That is why Anthropic built Duet around Claude Opus 5.5. The model choice suggests Anthropic values judgment over the lowest possible cost. Users cannot select another model, so they cannot trade capability for cheaper background checks.

Duet also makes Claude's existing memory, skills and connector strategy more valuable. Those features stop being optional chat conveniences when an agent uses them to choose work before a person asks. The same context that improves a briefing also expands the consequences of a bad judgment or an overbroad permission.

How Claude Duet works during a normal day

• Morning brief: Duet reads connected email and calendar data, selects the items it considers most important and can prepare first steps such as draft replies.
• Periodic monitoring: it checks connected accounts every few hours rather than continuously, partly to control usage consumption.
• Background execution: it breaks a request into steps, chooses connectors or a cloud browser and continues while you are away.
• Proactive follow-up: it can message you when a task finishes, needs a decision or appears to have gone quiet.
• Human completion: Duet can suggest that a task is finished, but only the user can mark it complete in the Tasks tab.

Which apps can Claude Duet use?

Anthropic says Duet can use more than 6,000 connectors and plugins in the Claude Marketplace. Named examples include Gmail, Google Calendar, Google Drive, Slack, QuickBooks, Stripe, HubSpot, Docusign, Calendly, Gusto, LinkedIn, Canva and Shopify.

The practical range is broad. Duet can find missed messages, draft replies, identify unpaid invoices, prepare payroll, create a job post, prepare an agreement, build a campaign or work on a Shopify store. A connector still needs user authorization, and access can be disconnected later.

For sites without connectors, Duet can use its cloud browser. Users cannot watch that browser live or take it over during the research preview, which removes a useful debugging and trust signal available in some computer-use tools.

What requires approval, and what remains off limits?

Anthropic says Duet must show the proposed action and wait before it sends a message, makes a payment, signs a document or deletes anything. It can finish the rest of a task while an approval waits. Users can approve purchases through Link or Shop Pay, with a one-time card used for each Link purchase.

Some actions remain prohibited even with approval. Duet cannot trade securities or move money between accounts. It also does not currently make phone calls or send text messages to other people.

The Claude Duet help guide warns that the agent can misread a thread or misjudge urgency. An approval screen reduces the cost of a bad action, but it does not repair a bad summary, a missed deadline or an interruption that arrives too late.

Security is the hardest part of a proactive work agent

Duet reads emails, documents and web pages written by other people. Any of them can contain instructions intended to manipulate an agent. Anthropic says Duet treats retrieved content as information rather than authority, but its own help page states that no prompt-injection defense is perfect.

Anthropic added a dedicated classifier and safety prompt to inspect messages for sensitive-data leakage. Requests to share the strictest data, including credit-card numbers and government identification, are rejected by default. Health and personal-finance details require explicit user instruction before Duet sends them.

The approval boundary matters because Duet combines reading and acting. Teams should still begin with low-consequence access, separate workspaces and narrow connectors. Review source citations, drafts and approval payloads before expanding the agent's reach.

Claude Duet pricing, limits and availability

The research preview is available by waitlist to Pro and Max subscribers in the United States. Team and Enterprise plans are not supported yet. Access opens in waves, and Anthropic says features can change during the preview.

Duet is included in the supported paid plans rather than sold as a separate subscription, but its work is not free. Morning briefs, account checks and background jobs consume the same five-hour session and weekly limits used by ordinary Claude chats. Larger jobs require more model turns and can reach those limits sooner.

Users admitted to the preview receive one full five-hour and weekly limit reset, available until November 30, 2026. That reset is temporary launch support, not a permanent increase in plan capacity.

Claude Duet compared with a normal chatbot and Claude Cowork

A normal Claude conversation is reactive. The user opens a chat, provides a request and waits for the answer. Duet adds a persistent task list, scheduled checks, proactive notifications and a cloud computer that keeps working after the user leaves.

Claude Cowork focuses on collaborative work that a user starts inside Claude. Duet is positioned as the agent that also starts the conversation, watches connected systems and decides when work needs attention. The distinction may narrow as Anthropic learns from the preview, since the company says Duet currently lives in a separate app and that early use will shape how the products fit together.

The wider market includes proactive and personal agents from OpenAI, Meta and other vendors. TechCrunch's launch report notes that Anthropic is emphasizing professional and small-business work rather than personal errands. Anthropic also says ticket booking and similar personal tasks are not a current strength.

Should you join the Claude Duet waitlist?

Join if you already use Claude memory and connectors, have recurring work spread across email, calendar and business apps, and can start with a reversible task. The preview is most interesting for people whose main bottleneck is noticing and coordinating work rather than generating one answer.

Wait if you need Team or Enterprise administration, live browser visibility, model choice, phone or SMS actions, continuous real-time monitoring, or predictable usage consumption. Also wait if your first use case requires broad access to regulated or high-consequence systems.

A useful first test is a one-week briefing experiment. Connect only a calendar and a low-risk inbox, define what deserves interruption, compare Duet's brief with your own priorities and record false positives, omissions, corrections and usage consumed. Expand access only when the evidence supports it.

• Choose one bounded workflow with a clear success condition.
• Connect the minimum set of apps and review their existing permissions.
• State what Duet should monitor, how often it should notify you and which events it should ignore.
• Inspect the original email, file or spreadsheet behind each important claim.
• Review every consequential approval and measure the plan usage spent on the workflow.
• Disconnect access and stop the trial if the agent repeatedly misreads urgency or scope.

Frequently asked questions

• Is Claude Duet a new model? No. It is a proactive-agent app built around Claude Opus 5.5.
• Is Claude Duet free? No. The preview is included for eligible US Pro and Max subscribers and shares their existing usage limits.
• Can Claude Duet keep working when my computer is off? Yes. It runs on its own cloud virtual machine and browser.
• Can Claude Duet send emails or make purchases by itself? It can prepare those actions, but Anthropic says it asks for approval before sending or paying.
• Can I watch Claude Duet use its browser? Not during the current research preview.
• Does Claude Duet work for Team or Enterprise accounts? Not yet. Anthropic says support is planned later.
• Can Claude Duet trade stocks or transfer money? No. Those actions remain off limits even with user approval.