On October 21, 2026, Anthropic opened research-preview access to Claude Duet, a proactive AI agent that identifies work-related matters requiring attention, carries out tasks in connected apps, and keeps working after the app is closed. The new service uses Claude Opus 5.5, has its own cloud computer, and inherits the memory and integrations of the regular Claude experience. At launch, however, it is available only through a waitlist for Pro and Max users in the United States.
Key Takeaways
- Claude Duet was introduced on October 21, 2026, as a separate app for proactive work with AI.
- The agent runs on Claude Opus 5.5 and uses its own cloud computer with a browser.
- Memory, skills, and connected services from a Claude account are available in Duet without having to set them up again.
- The agent can prepare morning briefings, analyze email, and create dashboards, financial models, and web apps.
- User approval is required to send messages, make payments, sign documents, and delete data.
- The research preview is limited to the United States and Pro and Max subscriptions; background activity counts toward shared Claude usage limits.
How Claude Duet Differs from Regular Claude
A conventional AI chat mainly responds to prompts: the user opens a conversation, describes a task, and waits for a result. Claude Duet is built around a different interaction model: the agent can notice an event in connected services on its own and suggest the next step.
According to Anthropic, Duet produces a short briefing every morning based on email, calendars, and other available sources. Throughout the day, it checks for changes and flags matters it considers important.
For example, the agent might notice an unanswered email, find an unpaid invoice, or prepare materials for a scheduled meeting. At the same time, it is not supposed to perform sensitive actions on its own without the account owner’s approval.
This is not a new language model but a separate product built on Claude Opus 5.5. Anthropic uses the model’s capabilities to plan tasks, select tools, and carry out multi-step work.
How Duet Keeps Working After You Close the App
Claude Duet’s key technical feature is its own cloud-based workspace. The agent uses a virtual computer and browser, so it does not depend on whether the app is open on the user’s phone.
If a task requires working across several websites or services, Duet can perform successive steps in its own environment. For sites without a dedicated connector, it uses the browser.
Anthropic says it supports more than 6,000 connectors and plugins. Examples include Gmail, Google Calendar, Google Drive, Slack, Shopify, QuickBooks, HubSpot, and DocuSign.
However, ecosystem support does not mean that every feature of every service is available without restrictions. For example, Anthropic specifically warns that some integrations have limited functionality and that Microsoft 365 is not supported in the current preview.
Background Work Does Not Mean Instant Monitoring
Duet checks connected accounts at set intervals. An important event may therefore occur between checks and not trigger an immediate notification.
This is a fundamental limitation for scenarios that demand a real-time response, such as monitoring critical outages, time-sensitive financial operations, or day-to-day management of advertising campaigns.
For such tasks, Duet should not yet be viewed as a replacement for dedicated alerting and automation systems.
Memory and Connectors Carry Over from Regular Claude
Anthropic cites quick setup as one of the benefits. Claude Duet uses an existing account, including its memory, skills, and previously connected apps.
The agent can also search for context in previous Claude conversations if the relevant feature is enabled. This means users do not have to repeat project descriptions or working arrangements with every new task.
The main interaction takes place in one ongoing conversation. Separate channels can be created for different topics while sharing the same memory.
However, automatic use of earlier context has a downside. If the agent misinterprets an old agreement or outdated information, its subsequent suggestions may be affected.
It is also important to remember that disconnecting a connector stops new access to that service but does not necessarily erase information the agent has already retained in memory from past conversations.
What Tasks Anthropic Suggests Delegating to the Agent
Claude Duet is primarily aimed at professional work and small businesses. Anthropic demonstrates scenarios in which the agent combines information from several apps and delivers a finished result.
- Communications. Reviewing incoming messages, drafting replies, and reminding users about unresolved conversations.
- Planning. Preparing for meetings, finding relevant materials, and managing calendars.
- Analytics. Creating dashboards, reports, financial models, and sales forecasts.
- Document work. Preparing research materials, presentations, and working files.
- Development. Prototyping and building web apps.
- Business operations. Checking outstanding invoices and drafting client reminders.
Larger outputs appear in the Make tab as files that can be viewed and edited. Ongoing tasks are shown in the Tasks tab.
At the same time, the agent finishing its work does not mean the result is automatically accepted. According to Anthropic’s documentation, only the user can ultimately mark a task as complete.
What Claude Duet Could Offer Marketers and Online Businesses
For marketing teams, the most interesting use case is not generating an individual text but carrying out a sequence of operations across different data sources.
For example, the agent could potentially be tasked with collecting campaign results from connected systems, comparing them with a plan, preparing a variance table, and producing a brief report for the team.
For agencies, this could also mean preparing client briefs, organizing research, updating presentations, and keeping track of work deadlines.
However, Anthropic does not claim that Claude Duet automatically supports the entire management workflow for Meta Ads, Google Ads, or TikTok Ads. Such capabilities depend on available integrations, user permissions, and each platform’s restrictions.
Similarly, the agent does not guarantee accurate financial forecasts or advertising analytics. Data, formulas, and conclusions should be checked before decisions are made.
Overall, Duet’s development continues the shift in AI from answering questions to carrying out longer workflows. This requires not only powerful language models but also tools, app access, and mechanisms for controlling actions.
Which Actions Duet Will Not Perform Without Approval
Anthropic has introduced separate controls for operations with potentially significant consequences.
According to the official documentation, Claude Duet must ask permission before sending messages, making payments, signing documents, and deleting data.
Some operations are prohibited outright, even with user consent. Examples include trading securities, transferring money between accounts, and accessing other people’s accounts.
Anthropic also describes protection against prompt injection: hidden instructions in emails, documents, or web pages that might try to alter the agent’s behavior.
The company says Duet does not gain direct access to passwords or login tokens. Separate verification mechanisms are used for sensitive data, and important actions require user approval.
Still, the developer itself acknowledges that no safeguard is foolproof. Granting the agent access to corporate email, financial services, and documents therefore requires carefully considered permissions.
What Does Not Yet Work in Claude Duet
The research preview comes with several limitations worth considering before entrusting the agent with real business processes.
- Users cannot manually choose another model: Duet uses Opus 5.5.
- Users cannot yet watch the agent’s browser activity or take over its controls.
- The agent does not have access to all local files on a computer or to all Claude projects.
- Duet cannot call other people or send SMS messages.
- Support for certain connectors is limited; Microsoft 365 is not yet available.
- The agent may misjudge the urgency of a message or misunderstand its context.
- Anthropic describes personal errands, such as booking tickets, as a less mature use case.
Therefore, Claude Duet should not yet be treated as a fully autonomous digital employee that can be trusted with any task without oversight.
Claude Duet Availability, Pricing, and Usage Limits
As of October 26, 2026, Claude Duet is available only as a research preview. Getting an invitation requires a Claude Pro or Max subscription, a US-based account, and registration on the waitlist.
Team and Enterprise users do not yet have access. Anthropic plans to expand availability but has not announced specific dates.
The service operates as a separate iOS and Android app and through Claude’s web interface. It is not a new mode inside the main Claude mobile app.
During the preview, there is no additional charge beyond Pro or Max. However, Duet uses the same limits as regular Claude, including five-hour and weekly limits.
Morning briefings, periodic service checks, and complex multi-step tasks also consume the available usage allowance. As a result, users may hit their limits sooner than if they were only using the chat.
After access is granted, Anthropic offers a one-time full reset of usage limits. According to the current documentation, this option is available until November 30, 2026.
Official access to Claude Duet has not yet been announced for users in Ukraine. There is also no launch date for countries outside the United States.
You can join the waitlist on the official Claude Duet page, but Anthropic determines whether an account meets the access requirements.
Claude Duet Competes with OpenAI Dots and Meta Muse
Anthropic is entering the market for proactive AI agents, where OpenAI Dots and Meta Muse are already active. The products share an underlying idea: instead of waiting for a separate prompt for every action, they use connected data and context to work over longer periods.
Claude Duet focuses on professional tasks, a cloud-based workspace, and use of an already configured Claude account.
It is still difficult to compare the agents’ quality: public demos are no substitute for independent tests on identical tasks, and availability and integration options differ.
For readers just getting acquainted with the Anthropic ecosystem, UAGEEK has a separate review of Claude’s capabilities for working with text, documents, code, and analytics.
Conclusion: The Key Innovation Is Not Autonomy but Persistent Work Context
Claude Duet combines three capabilities: long-term context from an existing account, work with connected services, and cloud-based task execution after the app is closed.
For businesses, this could reduce the number of manual steps between email, calendars, spreadsheets, and other work tools. But its practical value will depend on the agent’s accuracy, the quality of integrations, the usage-limit cost of background work, and the reliability of controls for sensitive actions.
For now, Anthropic is testing the product with a limited audience in the United States. Claude Duet should therefore be regarded as a promising but not yet fully proven automation tool, rather than a ready-made replacement for a team or business processes.
Sources for Verification
Sources for Verification
Anthropic — Introducing Claude Duet. Official announcement dated October 21, 2026, covering the agent’s capabilities, the Opus 5.5 model, background tasks, security, and launch conditions.
Official Claude Duet announcement
Claude Help Center — Get started with Claude Duet. Access requirements, setup, connectors, browser limitations, permissions, usage limits, and the one-time quota reset.
Claude Duet documentation
Anthropic — Claude Duet product page. Official product page describing use cases, more than 6,000 connectors, supported platforms, and the waitlist.
Claude Duet features and registration
9to5Mac — Anthropic takes on ChatGPT Dots and Meta Muse with Claude Duet. Independent coverage of the launch, workflows, integrations, and user approvals.
9to5Mac’s Claude Duet coverage
