What Is Claude Duet? Access, Background Tasks and Key Limits

What is Claude Duet, and does it add something you need beyond Claude? Anthropic's proactive agent can follow work across connected apps and return while you're away. The decision depends on access, the work you want it to notice, and which parts of your day you are ready to delegate. Those details matter more than the label “always on.”

What Is Claude Duet Designed to Do?

A continuing agent for work that needs attention

Anthropic's Claude Duet announcement describes an agent that can notice work across connected services, prepare results and return when it needs you. It has its own cloud computer and browser, so a delegated task does not depend on your phone or laptop staying on.

The appeal is continuity. A question today may connect to an email tomorrow or an upcoming meeting. Duet is intended to carry that context forward. That is the product's stated design, not a finding that it consistently recognizes the right priority.

It also changes the question you need to answer before trying it. With a writing tool, you can judge a draft when you request one. With a proactive agent, you also need to judge what it decides to surface, what it leaves alone and whether its interruptions arrive at a useful time.

Official Claude Duet announcement, checked November 2, 2026. Product information rather than a screenshot of a task run for this guide.

Who Can Get Claude Duet Access?

The preview still has an invitation boundary

The current getting-started guide lists three requirements: a US Claude account, a Pro, Max or Ultra plan, and an invitation from the waitlist. It says Team and Enterprise plans are not supported yet. Invitations arrive in waves, without a guaranteed date.

If access is missing after an invitation, check that the invited address matches the accounts used during setup.

Pro Tip: Treat preview eligibility and product availability as separate facts. Save the invitation and confirm that the feature actually opens before planning a business process around it.

What Claude Duet pricing means during the preview

Duet has no extra charge within the eligible plans during the research preview. For subscription context, the official pricing page, checked November 2, lists Pro at USD 20 when billed monthly, Max at USD 100 per month, and Ultra at USD 200 per month. Pro also has an annual option at USD 200 billed up front; the displayed USD 17 monthly equivalent is not a monthly payment option. Applicable taxes are excluded.

Claude Duet vs Cowork: What Is the Difference Now?

The old chat-versus-agent distinction is becoming outdated

Anthropic's current chat and Cowork guidance says those experiences are merging into a single conversation, with rollout occurring gradually. It describes cloud tasks and scheduled work that continue when your computer is off. Work requiring local files or applications still needs Claude Desktop open.

That makes “Duet works in the background; ordinary Claude does not” an unreliable comparison. Readers on different rollout stages may also see different controls. Compare the experience available in your account with the work you want done.

A new interface should solve a problem your current routine leaves unresolved.

How Claude Duet Background Tasks Work

Background work is not continuous monitoring of every event

The launch announcement says Duet checks connected accounts every few hours to manage usage. That is different from an immediate alert whenever something arrives. A request that depends on noticing a last-minute change needs a backup until you understand the actual cadence and behavior in your account.

Background checks also consume the same allowance as other Claude work. A quiet conversation can still involve activity behind the scenes. When comparing the effort of a quick answer with a research task, count the work required to gather inputs, use tools and produce files, not just the number of messages you typed.

For your first routine, define what deserves a notification. “Keep me informed” gives little guidance about urgency. An upcoming meeting missing essential preparation, a draft waiting for your decision and an ordinary newsletter should not all compete for the same attention.

Follow one business thread before connecting your whole day

Consider a proposed meeting-preparation task. The result could be a short note that combines an authorized calendar event, relevant correspondence and selected public company information. This is an evaluation example, not a Duet task performed for this article.

Define what belongs in that note: the meeting objective, recent commitments, unresolved questions and links supporting external claims. A useful result should tell you what still needs checking. It should not confidently merge two people with similar names or treat an old proposal as an accepted agreement.

Ask for the brief early enough to review it. You can then assess whether the agent found the right thread and reduced preparation work. That is a much clearer result than asking it to “manage relationships” and judging whatever it returns.

Pro Tip: Choose a task whose manual version you already understand. You will spot missing context faster, and you will have a realistic reference for how much review the new process needs.

What Are the Current Limits?

A cloud browser does not give you the same controls as your browser

The current help page says you cannot watch or take over Duet's browser during the preview. It also lists a fixed model, Claude Opus 5.5, rather than a model choice. Those constraints matter if your task requires visual inspection during execution or a specific model configuration.

Other documented limits include no phone calls or text messages to other people, incomplete access to Claude projects and local files, limits on some connectors, and no Microsoft 365 support yet. Check the relevant task path before granting account access. “Works with connected apps” does not establish that every tool or operation you already use elsewhere in Claude works in Duet.

A limitation is most useful when attached to a decision. If a result depends on a local workbook, choose a workflow that can reach that workbook. If an urgent phone call is the actual job, an email draft will not complete it. Narrow the task or use another process rather than quietly changing what success means.

What Happens When Duet Needs Approval?

Check the proposed action in its business context

Anthropic's safety explanation describes approvals for consequential actions, separate checks and credential isolation. It also explains that some actions remain unavailable even with consent and that safeguards can make mistakes.

For an owner, the practical issue is what the approval commits you to. Check the final recipient and message when sending correspondence. For a purchase, review the actual item, merchant and total. An earlier instruction to investigate a problem should not substitute for inspecting a later commitment.

The same distinction applies to status. An agent preparing an email has not sent it. A checkout proposal is not a completed order. After you authorize an external action, use the destination service or its confirmation record to establish what happened.

Anthropic's official safety explanation, checked November 2, 2026. The page describes product safeguards; it is not an independent security audit.

Disconnecting an app and removing learned context differ

The safety page says disconnecting a service stops new access, while information already learned remains until memory is reset. That is worth understanding before connecting work accounts. Removing the connection and deciding what context to retain are separate operations.

For a shared business environment, identify who is authorized to connect the data and what may be used for the task. A customer's confidential message should not become supporting material in an unrelated team update simply because an agent can retrieve it. Keep the task's audience and purpose explicit.

How to Judge Whether Duet Is Useful for You

Review the work and the interruptions together

Keep a short record of your first task: what you asked for, what came back, what needed correction and how long review took. For proactive behavior, also note whether a notification arrived at the right time and whether it changed what you did next.

An agent that produces more material can still add work if you must reconstruct every source. A brief that exposes a missing fact may be more useful than a complete-looking document that hides the gap.

Pro Tip: Include ignored notifications in your notes. Useful proactivity includes filtering your attention, so the number of alerts alone says little about its value.

Continue with the same task definition before expanding the scope. Changing the inputs, connected services and expected output together makes it hard to tell what improved. No accuracy score or savings estimate is established by this guide; those would need actual task observations.

When repeated website collection becomes a separate requirement

If the work repeatedly needs the same public-page fields, evaluate that collection job on its own. BrowserAct creates reusable website-data Bots from natural-language requests and supports structured outputs. It can be assessed as a separate input stage; this guide verifies neither a native Duet connector nor a tested combined workflow.

The live-web evidence guide covers source records and reviewable changes. Use a dedicated collection process only when the repeated data requirement justifies it.

What Is Claude Duet Best Used For?

Claude Duet is most worth evaluating when important work is slipping between apps and you want help noticing and following it through. A defined report you already know how to request may fit your existing Claude workflow. The difference should show up in what gets done and what deserves your attention.

Confirm your access, choose one reviewable task and decide what a useful interruption looks like. Then judge the actual result before expanding the role you give the agent.