Build · founder · 5 min read

Claude Duet Builds Dashboards and Web Apps While You're Away. Here's What Founders Should Check First.

Anthropic's Claude Duet is a proactive agent app on Opus 5.5. What it can build, what it can touch, and the limits to know before you join the waitlist.

Published October 26, 2026 ·

Claude AI agents Anthropic automation prototyping agent safety

On October 21, Anthropic announced Claude Duet, a research-preview app where Claude doesn’t wait for you to open a chat. It reads the apps you connect, sends you a morning brief, and pings you when something needs a decision. It runs on Claude Opus 5.5.

This isn’t a vibe coding tool in the Lovable or Bolt sense. But it can build dashboards, forecasts, prototypes and web apps when you ask, and it does that work on its own machine while your phone is in your pocket. That puts it in your orbit if you’re a founder or PM who builds things with AI.

What it actually is

Think of it as an assistant with its own cloud computer. You hand it a task by text or voice. It works in a virtual machine with a browser, so nothing needs to stay open on your end. Reports describe connections to thousands of connectors and plug-ins, including Gmail, Slack, Microsoft 365, HubSpot, Stripe, QuickBooks, Docusign and Shopify.

The business-ops examples Anthropic leans on are things like chasing unpaid invoices, preparing payroll and getting contracts ready to send. The build examples are dashboards, decks, spreadsheets and small web apps.

The access limits are real

• It’s a research preview behind a waitlist.
• US-only, and only Claude Pro and Max subscribers. Team and Enterprise aren’t included yet, with no timeline.
• Europe is excluded for now.
• It ships as an iOS and Android app, plus the web at claude.ai.

If you’re a solo founder on a Pro or Max plan in the US, you can try it. If you run a team plan, you’re waiting.

What it won’t do without you

The guardrails matter more than the feature list here. Duet asks for your approval before it sends a message, makes a payment, signs a document or deletes data. Reports say securities trades and transfers between accounts are blocked entirely, even with approval. Anthropic cites isolated credentials, filters for actions you didn’t request, and a bug bounty.

That’s a sensible design. It’s also an approval queue, and approval queues have a known failure mode: people click yes. If Duet sends you twelve “approve this email” prompts a day, you’ll start skimming. Decide up front which categories you will never rubber-stamp, such as anything involving money or contracts.

Where it can go wrong

Even Anthropic’s own framing admits an early version that can misjudge tasks. The specific risks coverage flagged:

• Misread threads. An agent summarizing a long email chain can get the gist wrong, and then act on the wrong gist.
• Gaps between check-ins. A proactive agent only knows what happened when it last looked.
• Prompt injection. An email or web page that contains instructions aimed at the agent is a real attack path. We’ve covered this in our prompt injection guide.
• Limited visibility. You can’t yet watch its browser as it works, so you’re reviewing results, not the process.

The cost nobody puts in the headline

There’s no extra charge. Duet draws from your existing plan allowance, including background checks you never see. Larger jobs, like building a financial model or a deck, use more. Reports say Pro subscribers in particular may hit limits faster. If you already burn through your Claude allowance on coding, adding a proactive agent that works while you sleep may eat the same budget. Our pricing calendar is worth a look before you commit workflows to it.

Should you join the waitlist?

Yes, if your pain is operational: invoices, scheduling, reporting, inbox triage. Those are low-glamour tasks where a decent agent with an approval step beats doing it yourself.

No, or at least not yet, if you want it as your app builder. For building real products, a dedicated tool like Lovable, Replit or Claude Code gives you version history, deployment and control over the code. Duet’s built apps are a side effect of an assistant doing errands, not a development environment.

How to try it safely

1. Connect the minimum number of apps. Start with read-only access where the connector allows it.
2. Keep payment and contract tools disconnected for the first two weeks.
3. Review every approval prompt for the first week, then audit a sample.
4. Watch your plan usage. If background work is draining your allowance, narrow its schedule.
5. Treat anything it builds as a prototype. Don’t ship it to customers without a human review.

The pattern across this year’s agent launches is the same: the capability arrives first and the controls catch up later. Duet’s approval design is better than most. Your job is to make sure it stays meaningful.