Claude Duet connects 6,000 services

The headline is not the chatbot. It is the permissions layer.

Anthropic’s new Claude Duet is being introduced as a proactive AI agent, but the more important announcement is architectural: Duet is designed to sit across a user’s connected work stack rather than wait inside a blank chat box. Anthropic says the research preview is available by waitlist, starting with U.S. Claude Pro and Max subscribers, and runs as a standalone app on iOS, Android, and the web through claude.ai [1]. That makes the launch narrower than a full enterprise deployment, but the product’s design is clearly aimed at the enterprise problem: who can safely let an agent see, summarize, draft, and act across business systems?

Duet is built around Claude Opus 5.5, and Anthropic frames its differentiator as “judgment”: the ability to decide what needs attention, what can wait, and when a human should approve the next step [1]. Instead of asking the user to return to a chat window, Duet sends a morning brief across connected apps, keeps tabs on those apps during the day, and messages the user when something appears to need a decision [1]. Reuters described the product as a standalone AI agent app for professionals and small-business owners, launched as rivals including OpenAI and Meta push assistants that complete work autonomously on users’ behalf [2].

Six thousand connectors change the competitive question

The scale number is the story: Anthropic says Claude Duet can use more than 6,000 connectors and plugins in the Claude Marketplace [1]. The named examples are not decorative. Anthropic lists Gmail and Slack for missed messages and replies, QuickBooks or Stripe for unpaid invoices, Gusto for payroll preparation, LinkedIn for approved job postings, Docusign for agreements, HubSpot for follow-ups, Canva for launch campaigns, and Shopify for store setup and management [1]. Reuters also highlighted Gmail, Slack, Stripe Link, and Shopify’s Shop Pay as part of the integration story [2].

That breadth shifts the agent race away from pure model demos. If every major lab can show a model writing emails, building dashboards, and summarizing meetings, the next battle is over permissions, connector quality, auditability, and administrative trust. A 6,000-service catalog means the agent is not merely “smart”; it is present in the places where work already happens. For IT leaders, that also means the surface area expands from a chatbot policy to thousands of possible app permissions. Jarvis, in other words, has discovered that enterprise productivity requires 6,000 permission slips.

Always-on does not mean always acting

Claude Duet is described as proactive, but Anthropic is careful to separate autonomy from unchecked execution. The agent can read, draft, and prepare work on its own, yet it must show the user what it intends to do and wait for approval before sending a message, making a payment, signing a document, or deleting anything [1]. Some actions are blocked even with approval, including trading securities and moving money between accounts [1].

This approval layer is central to Anthropic’s positioning. EconoTimes reported that Duet pauses for approval before messages, payments, signatures, or deletions, and that securities trading and inter-account money movement are prohibited even if the user approves [3]. Anthropic also says Duet’s credential handling follows the design used for Claude Managed Agents, so the agent does not directly see passwords or sign-in tokens; it also cites an auto-mode classifier, prompt-injection defenses, adversarial testing, outside penetration testing, and bug-bounty coverage [1].

The message to businesses is clear: Anthropic wants Duet to feel less like a rogue bot and more like a junior operator working behind guardrails. But the guardrails are also an admission that the hardest part of agentic software is not generating a draft. It is knowing when not to act.

The cloud computer is the product interface

Duet also changes the physical model of AI work. Anthropic says it runs on a virtual machine with its own browser, so work can continue even when a user’s laptop or phone is off [1]. Users can hand it tasks through text or voice in one ongoing thread; Duet then decides which connected apps and steps are needed, works in the background, and notifies the user when the task is complete or needs review [1].

MacRumors described the product as Anthropic’s answer to Meta’s Muse and OpenAI’s Dots, noting that Duet can connect to apps, suggest what needs attention, complete tasks in the cloud, and request approval before actions such as sending messages, making payments, signing documents, or deleting files [4]. Reuters similarly positioned Duet inside the broader agent race, noting that OpenAI launched “dots” agents last month and Meta’s Muse had become a major competitor in autonomous AI services [2].

The result is a new interface expectation. The chat window becomes less like the workplace itself and more like the dispatch channel for a background worker. If Duet works, the user does not “use AI” in discrete sessions; the user supervises an AI colleague that periodically reports back.

The preview has real limits

Anthropic is not presenting Duet as finished. The company says the agent checks connected accounts every few hours rather than nonstop because constant checking would consume plan limits faster [1]. Its background checks count against the same Claude usage limits as the rest of the user’s plan, and larger tasks such as building financial models or decks require more model turns and background work [1]. EconoTimes emphasized the same trade-off: Duet’s background activity draws on existing Pro and Max plan capacity [3].

Anthropic also warns that Duet may misread a thread, miscalibrate urgency, or miss something that arrives between check-ins [1]. The company says personal errands such as booking tickets are not as strong because the core functionality was built around work, and users cannot yet see the agent’s browser [1]. At launch, the preview starts with U.S. Pro and Max users rather than Team or Enterprise plans [1].

Those caveats matter. A proactive agent can fail in a different way from a chatbot: by interrupting too much, missing the important item, drafting from the wrong context, or making a routine approval feel harmless until it is not. The real test is not whether Duet can perform impressive individual tasks. It is whether users and administrators can tune its judgment enough to trust it across ordinary workdays.

Why this matters for enterprise AI

Claude Duet’s launch suggests that the next phase of AI competition will be less about who has the friendliest chat interface and more about who can operate safely across messy, permissioned workflows. Anthropic’s choice of examples — invoices, payroll, contracts, CRM follow-ups, storefronts, and approved purchases — points directly at operational work, not just knowledge work [1].

For companies, that raises a practical checklist. Which connectors should be allowed? Which departments can authorize payments or signatures? Which data categories should never leave their source system? How should logs be reviewed? What happens when a user connects a personal account alongside a work app? These are not edge cases; they are the operating model for agentic software.

Claude Duet is still a waitlisted research preview, not a final enterprise platform. But by connecting more than 6,000 services and placing approval gates at sensitive moments, Anthropic has made the market’s direction visible. The agent race is no longer only about the smartest answer. It is about the safest handoff between AI initiative and human authority.

Sources from the last 72 hours

1. [1]Introducing Claude Duet | Claude by AnthropicOct 21, 2026, 2:00 AM
2. [2]Anthropic unveils Claude Duet AI agent app | MarketScreener AustraliaOct 21, 2026, 10:18 PM
3. [3]Anthropic Opens Claude Duet Preview for U.S. Subscribers | EconoTimesOct 22, 2026, 7:48 AM
4. [4]Anthropic Launches Claude Duet, a Proactive AI AgentOct 21, 2026, 10:52 PM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.