Anthropic has opened a research preview of Claude Duet, a separate app where Claude works as a proactive agent instead of waiting for you to open a chat. It sends a morning brief on what needs your attention, takes tasks you hand it by message, keeps working on its own cloud computer while you are away, and asks for approval before it sends, pays, signs, or deletes anything.
Duet runs on Claude Opus 5.5. Access is by waitlist for Claude Pro and Max subscribers in the US, on iOS, Android, and claude.ai on the web. Team and Enterprise plans are not included yet.
It was the most-discussed AI launch of the last day. Anthropic’s announcement post on X passed 4.6 million views and 39,000 likes, and two Hacker News threads on Duet together drew more than 380 points and 290 comments.
Key points
- What it is: an always-on Claude agent in its own app, with one ongoing thread you message by text or voice.
- Model: Claude Opus 5.5.
- Who can get it: US Pro and Max subscribers, via waitlist. No Team or Enterprise yet.
- Price: no extra fee during the preview, but it shares your plan’s usage limits, and its background checks count against them.
- Guardrails: approval before sending, paying, signing, or deleting. Trading securities and moving money between accounts are blocked even if you approve.
- Status: research preview. Anthropic says it can still misread a thread or misjudge what is urgent.
What shipped
Duet is not a new model. It is a new way of running Claude: the agent reaches out to you first, rather than the other way round.
- Morning brief: every morning, Duet sends a short summary of the most important items across your connected apps.
- Background monitoring: it checks your connected accounts every few hours, then stays quiet unless something needs you. Anthropic says it does not check nonstop because that would use up plan limits much faster. You can tell it what you want to hear about and how often.
- Task handoff: you describe a job in a short message and it works out which apps and steps to use. Anthropic’s examples range from scheduling meetings to building dashboards, sales or financial forecasts, prototypes, and web apps. It returns finished work, such as a doc, sheet, or deck, for review.
- Its own computer: Duet gets a virtual machine with its own browser, so your laptop or phone does not need to stay on while it works.
- Your existing Claude context: memory, skills, and connected apps carry over from your Claude account, so there is little setup.
- Connectors: it can use the 6,000+ connectors and plugins in the Claude Marketplace. Anthropic names Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, Docusign, HubSpot, Canva, and Shopify.
- Purchases: it can buy things through Stripe’s Link wallet, which pays with a virtual card after you approve the spend, and it can check out with Shop Pay. Signing in to sites with logins saved in 1Password is listed as coming soon.
Anthropic says it built Duet for professionals and small business owners with more on their plates than they can track. The product page leans hard on that pitch, with sample briefs about contract redlines, hiring deadlines, permit fees, and refunds.
How the safety model works
An agent that reads your email and sends messages for you has two jobs: do the work, and not leak anything on the way. Anthropic’s answer stacks several layers it already uses elsewhere.
- Approvals: before Duet sends a message, makes a payment, signs a document, or deletes anything, it shows you exactly what it is about to do and waits.
- Hard limits: some actions are off-limits even with approval, including trading securities and moving money between accounts.
- Credentials: login handling follows the design Anthropic uses for Claude Managed Agents, so Duet never directly sees your passwords or sign-in tokens.
- Action checks: actions are approved by the same auto mode classifier that guards Claude and Claude Code against steps you did not ask for.
- Prompt injection: defenses come from Anthropic’s browser-use research.
- New leak classifier: a dedicated classifier and safety system prompt check each message. It is designed to refuse sharing the strictest tier of data, such as credit card numbers and government ID. Health or personal finance details only go out when you have explicitly told Duet to send them.
- Testing: Anthropic says internal red teams and an external firm penetration-tested Duet, and it is covered by Anthropic’s public bug bounty.
Anthropic says it will share more on Duet’s safety design in the coming days. That matters, because nothing above has been independently tested yet. 9to6AI has not used Duet. Everything in this section comes from Anthropic.
The limits Anthropic admits
The launch post is unusually direct about what does not work yet:
- It may misread a thread or miscalibrate what is urgent, and it can miss something that arrives between check-ins.
- It was built around work, so personal errands such as booking tickets are weaker.
- It cannot show you its browser yet.
- It lives in its own app, separate from the main Claude app, for now.
The usage question is the one builders should watch most closely. Duet shares limits with the rest of Claude, and background checks count. Anthropic says bigger jobs, like a financial model or a deck, need more model turns and background work. If you already hit your Max limits with Claude Code, an always-on agent drawing from the same pool is a real trade-off.
How it compares with OpenAI’s dots
Duet lands a few weeks after OpenAI launched dots, its always-on ChatGPT agents. The two products share the same core idea: a frontier model with its own cloud computer, connected apps, memory, and a habit of coming to you with work.
| Claude Duet | OpenAI dots | |
|---|---|---|
| Model | Claude Opus 5.5 | GPT-6 Astra |
| Connected apps | 6,000+ connectors and plugins | 4,000+ apps via plugins |
| Plans | Pro and Max (US, waitlist) | Pro and Business Premium, Enterprise beta |
| Where you talk to it | Standalone iOS/Android app, claude.ai | ChatGPT, Slack, Teams |
| Usage | Shares plan limits; background checks count | Conversations don’t count toward ChatGPT limits; tasks in Codex or ChatGPT Work do |
The differences are about posture. OpenAI pitches dots as agents that own a responsibility and work toward goals around the clock, and lets them live inside Slack and Teams. Anthropic pitches Duet on judgment and discretion: fewer pings, check-ins every few hours, and a long list of things it will not do without you. Which approach suits you depends on whether you want a teammate that pushes work forward or an assistant that mostly protects your attention.
What it means for builders
- If you are on Pro or Max in the US: join the waitlist, but connect the minimum set of apps first. Start with read-and-draft jobs (inbox triage, meeting prep, invoice reminders) before anything that sends or pays.
- If you build connectors or plugins: Duet turns Claude Marketplace listings into something that runs in the background on a user’s behalf. Clear permission scopes and predictable actions will matter more than ever.
- If you run a small business: the strongest fit Anthropic shows is admin work spread across many tools, such as chasing unpaid invoices or preparing payroll for review. Treat its drafts as drafts until you trust its judgment on your data.
- If you are outside the US or on Team/Enterprise: there is no access yet, and Anthropic has not given a timeline.
What to watch
- Anthropic’s promised deeper write-up on Duet’s safety design, and whether outside researchers find gaps in the new leak classifier.
- How fast background checks eat into Pro and Max usage once real users arrive.
- How Duet and the main Claude app end up working together, which Anthropic says early users will shape.
- Availability for Team and Enterprise plans and for users outside the US.
- The 1Password sign-in integration, which some Hacker News commenters already see as the riskiest part of the plan.
