Claude Duet Launches as a Proactive AI Agent for Claude Subscribers

What is Claude Duet and who can use it?

Claude Duet is Anthropic’s standalone proactive AI agent app for iOS and Android, with access also available through claude.ai on computers. Anthropic opened Claude Duet as a research preview through a waitlist for U.S. Claude Pro and Max subscribers on October 21, 2026. Anthropic’s Claude Duet announcement describes the service as an agent that can work across connected services rather than responding only to a single prompt.

Claude Duet differs from a conventional chatbot because the agent can take on an assigned task and continue processing it after the user leaves the app. That design is useful for tasks that involve reviewing an inbox, checking a calendar, preparing a morning brief, or assembling information into a dashboard. The research-preview status remains important because Anthropic can change Duet’s availability, behavior, and supported capabilities as it collects feedback.

Claude Duet is not presented as a separate subscription tier. Anthropic says the agent shares a subscriber’s existing Claude usage limits, including limits that apply to background work. That means Pro and Max subscribers should treat Duet tasks as part of their normal Claude usage instead of assuming the agent has unlimited capacity for long-running work.

Claude Duet arrives as more AI companies move from answer-focused chatbots toward systems that can use tools and complete multi-step work. Readers comparing connected AI services should also consider how broader platform access changes privacy exposure, particularly when an agent can review communications or calendar data. The same concern applies to AI tools that can control browser tabs, where a useful automation feature can also increase the amount of sensitive information available to an AI system.

How does Claude Duet work while you are away?

Claude Duet works while you are away by operating on a cloud virtual machine that has its own browser. Anthropic says this setup allows the agent to continue a task after the Claude Duet app is closed, which separates the work from the user’s phone or computer session. Anthropic’s Claude Duet help documentation explains that the agent can perform work in the background and notify the user when review is needed.

Claude Duet’s cloud browser matters because the agent does not need to keep a personal device unlocked and active while it completes a supported task. A subscriber could ask Claude Duet to review information from connected services, organize findings, and return when a decision requires attention. The practical benefit is continuity, but the tradeoff is that users must understand which accounts and connectors they authorize before assigning a task.

Claude Duet runs on Claude Opus 5.5, according to Anthropic. The model choice matters because Claude Duet needs to interpret instructions, navigate connected services, and decide when an action needs user review. Anthropic has not positioned the model as a guarantee that every output or automated action will be correct, so users should examine summaries, drafts, and proposed changes before relying on them for important work.

What tasks can Claude Duet complete?

Claude Duet can review connected inboxes and calendars, prepare morning briefs, draft replies, and create briefs, decks, and dashboards. Anthropic says Claude Duet can also notify users when a task reaches a point that needs review. Those features position Claude Duet as a tool for organizing information and preparing work, rather than as an autonomous replacement for a person who must make the final decision.

Claude Duet also works with a large set of third-party connections. Reuters reported that the service integrates with more than 6,000 connectors in the Claude Marketplace, including Gmail and Slack. The Reuters report indicates that the connector ecosystem is central to Duet’s ability to collect information across services. The practical limitation is that a connector only expands what the agent can reach after the user chooses to connect it.

Claude Duet is likely most useful for recurring tasks with a clear structure. Examples include collecting unread messages for a morning summary, comparing calendar conflicts, drafting a response for review, or turning existing notes into a presentation outline. Tasks that involve unclear facts, sensitive instructions, or irreversible decisions need closer human review because an agent can organize information without fully understanding personal priorities or external consequences.

Claude Duet’s connected-service model also makes permission management important. Users should connect only the inboxes, calendars, and workplace tools needed for a specific task. That approach reduces unnecessary exposure if an agent is assigned to prepare a narrow briefing rather than review an entire digital workspace. Consumers should apply the same restraint to other services that add AI access to personal data, including new Claude-connected home search tools that can combine personal preferences with external services.

What actions require Claude Duet approval?

Claude Duet is designed to ask for approval before sending messages, making payments, signing documents, or deleting material. Anthropic’s approval requirement is the most important protection in the research preview because those actions can create financial, legal, or communication consequences that cannot always be easily reversed. Subscribers should read each approval request instead of treating the prompt as a routine confirmation.

Claude Duet keeps two financial actions off-limits: securities trading and moving money between accounts. Anthropic’s limits reduce the risk that the agent could directly execute particularly sensitive transfers or investment actions. The restriction does not remove the need for caution around other payment-related tasks, especially when a user authorizes the agent to prepare purchases, refunds, invoices, or payment links.

Claude Duet users should begin with tasks where the approval screen is easy to evaluate. Drafting a reply, collecting calendar conflicts, or producing a non-sensitive briefing gives the user a chance to evaluate the agent’s reliability without creating an immediate financial or legal commitment. The most sensible approach is to keep final authority over messages, documents, and payments with the account holder.

Claude Duet’s review requirement also creates a useful stopping point for complex tasks. Stop and handle the work directly if an approval request involves a contract, an unexpected recipient, a payment amount you did not specify, or a deletion that affects shared material. An AI agent can prepare a proposed action, but the person responsible for the account remains responsible for approving it.

How does Claude Duet handle purchases and Stripe tasks?

Claude Duet can support certain purchase and business-payment tasks through Stripe integrations, but users still approve the final transaction. Stripe says Claude Duet users can approve purchases with Stripe Link virtual cards. The virtual-card approach matters because it creates a payment approval step rather than allowing an agent to move money between accounts or trade securities.

Claude Duet can also use Stripe MCP for tasks involving invoices, payment links, products, prices, revenue analysis, and refunds. Stripe’s integration announcement identifies those capabilities and describes the connection between Stripe tools and Anthropic’s agent. The availability of a tool does not mean every business should authorize it for every financial workflow, especially where a mistake could affect customers or accounting records.

Claude Duet users should start with information-gathering and draft work before approving a real purchase or refund. Reviewing revenue data or preparing an invoice draft is easier to verify than authorizing a live payment-related action. Businesses should also confirm internal approval policies before connecting financial systems, since Claude Duet’s product controls do not replace an organization’s own finance procedures.

Claude Duet does not support moving money between accounts, according to Anthropic. That boundary limits one category of harm, but users should still inspect all payment details before approval. The practical rule is straightforward: use Claude Duet to organize, analyze, and prepare, then personally verify any real-world transaction before it is completed.

What cannot Claude Duet do yet?

Claude Duet cannot make phone calls, send SMS messages, access all projects, or access local computer files, according to Anthropic. Claude Duet also does not yet support Microsoft 365. Those limits matter because they define where the agent’s background access stops, and they prevent users from assuming the service can reach every document, device, or workplace account they use.

Claude Duet’s lack of local-file access means the agent is not a replacement for a desktop assistant that can search every folder on a computer. Users who need help with files stored on a personal device should not upload sensitive material merely to work around that limitation. The safer choice is to use supported connections only when the information is necessary for the task.

Claude Duet’s lack of Microsoft 365 support also limits its value for many workplace users. Organizations that depend on Outlook, Teams, OneDrive, Word, Excel, or SharePoint will need to wait for support or use another workflow. The limitation is particularly relevant because workplace AI agents can create data-governance concerns when they gain broad access to company systems, as shown by ongoing scrutiny of consumer-harm questions around AI services.

Claude Duet remains a research preview, so subscribers should expect supported services and limits to change. Users should check Anthropic’s current help documentation before assigning sensitive work, especially if an update adds a new connector or changes the scope of background access. A new capability can improve convenience while also requiring a fresh review of permissions.

How should you use Claude Duet safely in the research preview?

Claude Duet is safest when used first for low-risk tasks that have clear instructions and easy-to-review outputs. Inbox summaries, calendar briefs, draft messages, and document outlines allow subscribers to evaluate the agent without immediately sharing highly sensitive records or approving consequential actions. The research-preview label is a reason to be deliberate, not a reason to avoid every feature.

Claude Duet users should review connected accounts before starting a task. Remove connectors that are not needed, avoid assigning broad requests that collect unnecessary personal information, and inspect the result before forwarding, signing, paying, or deleting anything. This matters because a cloud-based agent can act across connected services, even though Anthropic says it will ask for approval before certain high-impact actions.

1. Start with an inbox, calendar, or briefing task that does not involve financial, medical, legal, or highly confidential information.
2. Connect only the service required to complete that task.
3. Write a specific instruction that identifies the expected output and the actions Claude Duet must not take.
4. Review every draft, summary, and approval request before accepting it.
5. Disconnect services and stop using the task if Claude Duet requests access or approval that does not match your instructions.

Claude Duet users should stop and contact Anthropic support if the agent appears to access an unexpected service, produces an approval request that cannot be explained, or behaves differently from the requested task. Users should also handle urgent financial, legal, medical, or security decisions directly with the relevant professional or service provider. Claude Duet can assist with preparation, but it is not a substitute for professional judgment.

FAQ

Is Claude Duet available to all Claude users?

Claude Duet is not available to all Claude users. Anthropic opened the service as a U.S. research preview through a waitlist for Claude Pro and Max subscribers on October 21, 2026.

Can Claude Duet work after I close the app?

Claude Duet can continue working after the app closes because the agent runs on a cloud virtual machine with its own browser. Anthropic says the service can notify users when a task needs review.

Does Claude Duet send messages without approval?

Claude Duet is designed to ask for approval before sending messages. Anthropic says the same approval expectation applies to payments, document signing, and deleting material.

Can Claude Duet access files on my computer?

Claude Duet cannot access all local computer files. Anthropic also says the research preview cannot access all projects, so users should not assume the agent can search every personal or workplace document.

Can Claude Duet use Microsoft 365?

Claude Duet does not yet support Microsoft 365. Users who rely on Outlook, Teams, OneDrive, Word, Excel, or SharePoint will need to use another workflow until Anthropic adds support.