Claude Duet wants to run your workday while your phone is off

Claude Duet is a proactive agent, which means it starts conversations instead of only answering them. Every morning it sends a short brief covering the most important items across the apps you’ve connected. After that, it checks those apps every few hours and only pings you when something looks like it needs a decision.

You can also hand it jobs. Anthropic says those range from booking meetings to bigger projects like building dashboards, putting together sales or financial forecasts, and prototyping web apps. You talk to it by text or voice in one long running thread, so a two-word reply and a week-long project live in the same place.

The part that sets it apart from a regular chatbot is where the work happens. Duet runs inside its own virtual machine, a computer in the cloud with its own web browser. Your phone and laptop can be switched off and it carries on. When a task is done, or needs a look from you, you get a notification.

Setup is light for existing Claude users. Duet pulls in the memory, skills and connected apps already linked to your Claude account, and it’s built around Claude Opus 5.5, the model Anthropic credits for its judgment about what can wait.

Duet can link to more than 6,000 apps and services through Anthropic’s connector marketplace. The list includes everyday work tools such as Gmail, Slack, LinkedIn and Canva, plus business software like HubSpot, QuickBooks, Stripe, Gusto and Shopify.

In practice, that opens up chores that usually eat an evening:

• Money admin: spotting overdue invoices in Stripe, drafting reminder emails and preparing a payroll run in Gusto.
• Marketing: pulling together a campaign across platforms in Canva.
• Online stores: managing a Shopify shop and completing approved purchases through Stripe’s Link wallet or Shop Pay, using virtual card numbers that keep your real card hidden.
• Planning: building forecasts, financial models and interactive dashboards from data in several apps.

Anthropic shared stories from early testers. One, the owner of a South Florida restaurant chain with about 110 employees, used it for accounting, customer support and marketing and worked through invoices one location at a time. Another tester’s morning brief flagged a property showing that had been put on her calendar for the wrong day. A third got a prep brief for a client call, plus an offer to move the call because it clashed with her flight.

Duet asks for permission before any step that’s hard to undo. Payments, outgoing messages, deletions and signatures on documents all need your sign-off: it shows you exactly what it plans to do and waits for a yes.

Some actions are off the table even if you approve them. Anthropic says Duet won’t trade securities or move money between your accounts, two areas where a single mistake could be costly.

There are other guardrails worth knowing about:

• No password access: when you connect an account, Duet never directly sees your passwords or sign-in tokens.
• A check on every action: an automatic classifier (a second AI system that screens behavior) approves each step, the same safeguard Anthropic uses for its Claude Code coding tool.
• A leak filter: a new classifier checks each outgoing message. It’s designed to refuse to share credit card numbers or government ID numbers at all, and it only sends health or personal finance details when you’ve clearly told it to.
• Outside testing: internal red teams and an outside security firm tried to break it, and it’s covered by Anthropic’s public bug bounty, which pays researchers who find flaws.

Those protections are aimed at a real risk. An agent that reads email can be fed hidden instructions by a stranger’s message, a trick known as prompt injection. Anthropic says Duet’s defenses come from its research on AI browsing, though it also says it will share more detail in the coming days.

For now, Claude Duet is only open to Pro and Max subscribers in the United States, and you have to join a waitlist. There’s no extra fee during the preview, but there’s a catch that’s easy to miss: Duet uses the same usage limits as the rest of Claude. Every background check of your inbox and every long task counts against your allowance, so a busy agent could leave you less room for regular chats.

Readers in the UK and Canada will have to wait. Anthropic hasn’t said when the preview will open outside the US.

If you do get in, start small. Connect one or two apps you use most, see whether the morning brief flags things you actually care about, and tune it from there. An agent that pings too often becomes noise, and one that filters too hard can miss the email that mattered.

Duet arrives late to a crowded field. Google introduced Gemini Spark at its I/O conference in May, Meta launched its Muse assistant in September, and OpenAI followed with Dots about a week later. Meta said this week that more than 10 million people in the US and Canada have already set up Muse.

Anthropic is pitching Duet differently. Muse and Dots are consumer helpers with friendly cartoon avatars and a heavy focus on shopping. Duet has no mascot and is aimed squarely at work: professionals and small business owners with more on their plate than they can track.

That focus cuts both ways. Anthropic admits Duet is not as strong at personal errands like booking tickets, and that it treats them as less urgent. It also says this early version might misread a thread or misjudge what’s urgent. And unlike a person sitting next to you, you can’t yet watch its browser while it works, so you’re judging the results rather than the process.