Claude Duet Explained: Anthropic's Proactive Opus 5.5 Agent

Claude Duet is Anthropic's new standalone app that turns Claude into a proactive agent: it messages you first with what needs attention, keeps grinding on tasks after you close the app, and asks for your sign-off before it sends a message, pays a bill, or signs anything. It opened as a research preview on October 21, 2026, runs on Claude Opus 5.5, and for now it's a US-only, waitlist-gated perk for Pro and Max subscribers.

If that sounds a lot like Meta Muse or OpenAI's Dots, that's because it is the same category of product, three different labs, three weeks apart. Meta launched Muse on September 8. OpenAI shipped Dots at DevDay on September 29. Anthropic landed Duet on October 21. Nobody's calling this a coincidence: the "agent that works while you're not looking" is clearly the thing every major lab decided to ship this quarter.

What does Claude Duet actually do?

Duet lives in one continuous thread per task, not a fresh chat every time. You talk to it with voice or text, and it keeps a running memory of what it's doing for you. The pitch, straight from Anthropic: it comes to you with what needs your attention, completes tasks you hand it, keeps working when you step away, and checks with you before it acts on your behalf.

In practice that breaks down into three things:

• Morning briefings. Duet scans your connected apps overnight and gives you a priority rundown: the email that needs a reply, the invoice that's overdue, the calendar conflict nobody noticed.
• Background execution. Ask it to build a financial dashboard, draft a forecast, or stand up a small web app, and it runs the job on its own cloud virtual machine with its own browser, not your laptop's CPU.
• Quiet monitoring. It watches your connected tools throughout the day and only pings you when something actually needs a human decision, instead of narrating every step.

It's built for professionals and small business owners specifically, not "book me a dinner reservation" territory. That's a real split from Meta Muse, which leans into consumer errands: grocery lists from recipe reels, party invites, travel booking. Duet wants your QuickBooks, your Stripe account, your Docusign queue.

Why does it matter that it's built on Opus 5.5?

Opus 5.5 shipped back in September 2026 as the first model in Anthropic's 5.5 family, and Anthropic says it cuts typical workload costs by around 40% versus Opus 5 while running over 30% faster. That matters a lot for an agent product specifically, because Duet isn't a single request-response exchange. It's a model running continuously in the background, burning tokens on every browser click and every tool call while you're not even watching.

A slower or pricier model would make an always-on agent either too laggy to feel "proactive" or too expensive for Anthropic to bundle into a $20/month Pro plan. Opus 5.5 is the model that makes the economics of "an agent that works all day" actually survivable. There's no model picker in Duet, either: you get Opus 5.5, full stop, no dropping to a cheaper model to save your usage limit.

What can Claude Duet connect to?

Anthropic says Duet reaches more than 6,000 tools through the Claude Marketplace, with the headline integrations being Gmail, Google Calendar, Google Drive, Slack, QuickBooks, Stripe, HubSpot, Docusign, Calendly, and Shopify. Commerce goes through Stripe Link and Shop Pay for actual purchases.

For anything not covered by a connector, Duet falls back to its own browser: it can navigate a site manually the way you would, just slower and with more patience than you have at 11pm.

What it can't do matters just as much. Per Anthropic's own support documentation, Duet:

• Can't move money between accounts or trade securities, full stop: those are hard blocks, not just approval gates
• Can't text or call anyone on your behalf
• Can't touch your local files or existing Claude Projects
• Can't monitor background browser tabs you have open yourself

How does the approval system work?

Before Duet sends a message, authorizes a payment, signs a document, or deletes anything, it stops and asks. You can defer that approval and let the rest of the task keep running in the meantime, which is a nice touch: it doesn't block the whole pipeline waiting on you to glance at your phone.

Anthropic has also bolted on a dedicated classifier meant to catch sensitive-data leaks, plus prompt injection defenses pulled from its browser-use research, external red-teaming, penetration testing, and bug bounty coverage. Translation: they know an agent with your Gmail, Stripe, and Docusign hooked up is a juicy target for prompt injection attacks hiding in a webpage or an email, and they're trying to get ahead of it before a researcher finds the hole for them.

Who actually gets access right now?

This is the part that's easy to overstate. Duet is a research preview, not a GA launch:

• Pro ($20/month, or $17/month billed yearly) and Max (from $100/month) subscribers only
• US-based accounts only
• Invite comes through a waitlist, distributed in waves, and joining doesn't guarantee a date
• Team and Enterprise plans aren't supported yet, though Anthropic says that's coming
• Background Duet activity shares your plan's existing usage limits, so running it hard can eat into the quota you'd otherwise spend on regular Claude chats
• One free usage-limit reset is available through November 30, 2026, which tells you Anthropic expects people to blow through their limits testing this

So if you're on the Free plan, this isn't for you yet. And even paying US subscribers need an invite. That's a narrower rollout than OpenAI took with Dots, which went to Pro and Business Premium users in "eligible markets" rather than one country.

How does Claude Duet compare to Meta Muse and ChatGPT Dots?

All three are the same basic idea (a cloud-hosted agent with its own virtual machine and browser that keeps working after you close the app), but the target user and the access model differ enough to matter.

The smart glasses angle is Muse's real differentiator: Meta can push it through hardware nobody else has. Dots' edge is that permissions carry across ChatGPT, Codex, and ChatGPT Work, so if you've already connected an app anywhere in OpenAI's ecosystem, a Dot inherits it instantly. Duet's edge is narrower but sharper: it's explicitly aimed at the QuickBooks-and-Docusign crowd instead of trying to be everything to everyone.

Pricing-wise, OpenAI bundles a Dot into Pro/Business Premium at no extra cost with a temporarily extended usage allowance, and conversations with a Dot don't count against normal ChatGPT limits. Anthropic went the other way: Duet usage shares your existing Claude quota, which is the less generous call during a free preview.

Is Claude Duet any good so far?

Early reaction on Hacker News was split in a way that feels honest rather than hyped. One commenter described using Claude's agent capabilities to buy a car and work through a house purchase in Kyoto with zero Japanese, genuinely impressive if true, and a real signal that Opus 5.5's agentic chops are a step up from what came before.

But the same thread had plenty of "why am I retyping information Anthropic already has" complaints about the waitlist signup flow, and jabs about search still matching individual words instead of phrases years after that stopped being a hard problem. There's also a real, not-paranoid concern buried in there about handing tax documents and account credentials to a remote agent with browser access: prompt injection on agents with live browsers isn't theoretical, it's a documented attack class, and Anthropic's own blog post name-checks defending against it as a specific engineering problem they had to solve.

My honest take: the proactive-briefing-plus-approval-gate model is the right shape for this category. Nobody wants an agent that silently fires off emails on your behalf, and nobody wants to approve every single click either. Checking in only on payments, messages, and deletions is a sane middle ground. What I'd want to see before trusting it with QuickBooks is a longer track record on the prompt injection defenses holding up against actual red team attempts from outside Anthropic, not just the ones Anthropic ran on itself.

If you're already deep in Claude for coding work, it's worth reading how Opus 5.5 performs in that context before judging Duet purely on the agent demo: see our breakdown of the Opus 5 launch and the rest of the Claude 5.5 family for the model side of this. And if you want the fuller picture on what Duet is actually racing against, we covered Meta's Muse and OpenAI's agent push separately.

Watch the announcement