Anthropic opened a research preview of Claude Duet on Wednesday, a standalone app that turns Claude into a proactive personal agent designed to track your work, surface what needs attention and complete tasks while you step away. The launch, covered by Reuters, The New York Times, CNBC and The Verge, positions Anthropic squarely against Meta's Muse and OpenAI's Dots in the fast-crowding always-on agent market.
Access starts with a waitlist. US subscribers on Claude's Pro and Max plans are first in line, and Anthropic says it will email users as their access opens. Claude Duet ships as a native app on iOS and Android and runs on claude.ai, and during the preview it is bundled into existing Pro and Max subscriptions with no separate fee. Usage draws from the same limits as the rest of Claude.
A Morning Brief, Then Quiet Monitoring
Anthropic says it built Claude Duet around Claude Opus 5.5 and designed it to make three judgments continuously: what genuinely needs the user, what can wait, and when it should check first before acting.
Each morning the app delivers a short brief covering the most important items across a user's connected apps. During the day it keeps monitoring those apps and only messages the user when something actually requires attention. The company frames the task range broadly: scheduling meetings, building dashboards that pull data from dozens of tools, producing financial or sales forecasts, prototyping software, and building web apps.
Everything lives in one ongoing thread, and users can type or speak to it. Claude Duet inherits the memory, skills and connected apps already attached to a user's Claude account, so it does not start from zero.
It Runs on Its Own Machine
One structural difference from rival assistants: Claude Duet executes on a virtual machine with its own browser. The user's laptop or phone does not need to be on for work to continue. When a task finishes — or stalls on a decision — the app sends a notification.
The trade-off is metered. Anthropic says the app checks connected accounts every few hours rather than continuously, because nonstop polling would burn through plan limits far faster. Those background checks count toward the same shared usage limits as normal Claude conversations, and heavyweight tasks such as building a financial model or a slide deck consume more model turns than a quick lookup.
Approval Gates and Hard Blocks
Anthropic is emphasizing restraint as a feature. Claude Duet can read, draft and prepare work autonomously, but before it sends a message, makes a payment, signs a document or deletes anything, it shows the user exactly what it intends to do and waits for approval.
A short list of actions is blocked even with explicit approval — including trading securities and moving money between accounts. Users can disconnect any app the agent can access at any time.
The safety architecture borrows from systems Anthropic has already shipped. Credential handling follows the design used for Claude Managed Agents, which the company says means the app never directly sees user passwords or sign-in tokens. Each message passes through a dedicated classifier and a safety system prompt designed to refuse sharing the strictest tier of personal data, such as credit card numbers and government identification. Other sensitive categories, including health and personal financial details, flow only when the user has explicitly instructed the agent to send them.
Anthropic says Claude Duet's defenses against hidden instructions and prompt injection come from its browser-use research, that the app is covered by the auto mode classifier already protecting claude.ai and Claude Code, and that internal red teams plus an external firm penetration-tested the product before release. It also falls under Anthropic's public bug bounty program on HackerOne.
6,000 Connectors, From Gmail to Shopify
The app plugs into more than 6,000 connectors and plugins in the Claude Marketplace. Anthropic's listed use cases sketch out the ambition: finding and drafting replies to missed Gmail and Slack messages, checking QuickBooks or Stripe for unpaid invoices with reminders ready to send, running payroll in Gusto with hours and bonuses filled in, publishing LinkedIn job postings after human edits, routing agreements through Docusign, drafting HubSpot follow-ups from contact history, turning a product launch into a Canva social campaign, and setting up and managing a Shopify store.
Payments run through Stripe's Link wallet: the agent proposes the purchase, Link asks the user to approve it, and the charge executes on a virtual card so the real card number is never shared. Shopify's Shop Pay works for checkout at online stores, and Anthropic says support for signing in with 1Password-stored logins is coming soon, with user permission.
Anthropic's Own Staff Are Testers
Internally, the team leading the project includes Karan Ahuja, who heads Claude Duet, and the company shared usage habits from other leaders: Boris Cherny, who leads the Claude Code team, says he walks through his mental to-do list with the agent; Chelsea Larsson, who leads model and product behavior design, hands it her list before stepping away from her children in the evening and collects the results later.
Early business users described mixed but concrete wins. Jess Provencher, a broker and owner of Pro Homes, said the app flagged a property showing booked on her calendar for the wrong day without being asked. Amin Ebrahimi, owner of South Florida restaurant group Shahs of Kabob, said he has used it for accounting, customer support and marketing across roughly 110 employees. Brand strategist Mya Dalton said it prepared a brief for a client call and, noticing the call sat just before her flight, offered to move it.
Rough Edges Remain
Anthropic is upfront that this is an early version. It can misread a thread or misjudge what counts as urgent. Personal errands such as booking travel tickets are weaker than work tasks, because the product was built around professional workflows first. The agent cannot yet show users its browser session, and for now it lives in a separate app rather than inside the main Claude app — a split Anthropic says early-user feedback will help resolve.
The Agent Platform Race Tightens
The launch lands in a market segment that barely existed eighteen months ago. Meta's Muse and OpenAI's Dots both pursue the same thesis — that the next interface for AI is an agent that knows your context, acts on your behalf and interrupts you only when necessary. CNBC framed Claude Duet directly as Anthropic's answer to both.
For readers tracking the pace of the agent market, AI Buzz Wire follows every major release. The differentiators now look less like raw model quality and more like trust plumbing: how payments are gated, how credentials are isolated, and how reliably an agent knows when to stop and ask. Anthropic's bet is that its approval-first design will travel well with businesses that like the productivity case but fear the failure modes.
Stay Ahead of AI
The AI agent landscape is moving weekly. For more breaking AI news and independent coverage of every major launch, read the latest AI developments on AI Buzz Wire.
Read more AI news →
