I've been building software long enough to remember when the biggest upgrade to an assistant was it answering faster. Then it started answering better. And now, as of October 21, 2026, Anthropic has released something that doesn't wait for me to ask at all. Claude Duet messages you first.
My first reaction, honestly: a little thrilled, a little uneasy. An AI that watches my connected apps and pings me when something needs attention is exactly the chief-of-staff I've always wanted — and exactly the thing to inspect carefully before handing it real responsibility. So this is a developer's walkthrough of what Duet is, what it can do, what it costs, and where the honest limits are.
What Claude Duet actually is
On Wednesday, October 21, 2026, Anthropic introduced Claude Duet: a standalone AI agent app designed as a proactive assistant for professionals and small business owners. The pitch, in Anthropic's own words, is a version of Claude you message the way you'd message a person — except it comes to you with what needs your attention, completes the tasks you hand off, keeps working when you step away, and checks with you before it acts on your behalf.
This isn't a chatbot update. It's a separate product entering a suddenly crowded field: OpenAI launched its always-on "Dots" agents last month, and Meta's Muse agent has already drawn millions of downloads. Reuters framed it plainly — the big labs are racing to build agents that complete work autonomously, and Duet is Anthropic's answer.
The design center is proactive behavior. Duet sends a short brief each morning summarizing the important items from your connected apps, then checks those apps every few hours and messages you when something needs attention. It can handle simple administrative chores or bigger jobs — scheduling meetings, drafting replies, building financial forecasts, even building web apps.
The morning brief: what a day with Duet looks like
You wake up to a morning brief: a short digest of what matters across your connected apps — the email thread that went quiet overnight, the invoice that's still unpaid, the meeting that needs rescheduling. You didn't ask for any of it. Duet decided these were the items worth your eyes first.
Through the day it keeps checking every few hours and messages you in one ongoing thread, the way you'd text a human assistant — typed or spoken replies. A dedicated Tasks tab tracks what's in its queue, and only you can mark tasks as finished: the loop stays open until a human says done, so it can't quietly bury something it fumbled.
The work doesn't stop when you close the app. Duet runs on its own cloud-based computer with its own virtual machine and browser, so it keeps working in the background even after your phone goes to sleep. That's the "always-on" part — the same architecture OpenAI's Dots and Meta's Muse use.
Under the hood: Opus 5.5 and your existing Claude brain
Duet is built around Claude Opus 5.5, released in late September 2026. Mike Krieger, co-lead of Anthropic's Labs team, said in the launch video that Opus 5.5 gives the agent "the discernment and judgment around when to be proactive, when to act on your behalf." That phrasing matters, because discernment is the entire product: a proactive agent that pings you about everything is just a notification firehose, and Anthropic knows it.
Setup is deliberately minimal. Duet pulls in the memories, skills, and connectors already attached to your Claude account, and can search your past conversations so you don't repeat yourself. Reuters reports over 6,000 connectors in the Claude Marketplace — Gmail, Slack — plus money integrations: Stripe for unpaid-invoice checks and drafted payment reminders, and Stripe's Link wallet plus Shopify's Shop Pay for purchases. An assistant that can buy things is a different creature from one that drafts emails, which is why Anthropic built hard approval gates into the core loop.
What it can actually do — the developer's view
For developers and small-business owners (the audience Anthropic explicitly targets), the capabilities break into three layers:
Watch and surface. Email, calendars, team chat, invoicing — Duet monitors connected apps and surfaces what needs attention. The morning brief is the headline feature: the "did anyone reply to the proposal" layer.
Do the paperwork. Scheduling meetings, drafting replies, creating files, building financial forecasts and models, chasing unpaid invoices through Stripe with drafted reminders — multi-step work in its own cloud VM with its own browser.
Build things. Anthropic lists building web apps among Duet's tasks — it can write code, run a browser, and iterate without your machine. Its typed and spoken messages plus per-topic channels that share the same memory mean you could keep a client project in its own channel with shared context. That's the pattern I most want to test when I get access.
The approval gates: what it can't do without asking
Here's the part I insist on understanding before trusting any agent: where does it stop and ask?
Anthropic's line is clear. Duet is designed to ask before it sends a message, makes a payment, signs a document, or deletes anything — the four categories of irreversible damage: reputation, money, legal exposure, and data loss.
The safety architecture has more layers. Duet does not see passwords or logins when you give it access to accounts. A dedicated classifier and safety system prompt check each message and guard against sensitive information leaking. In a world where agents get attacked through prompt injection, a separate classifier watching the message stream is a meaningful design choice, not decoration.
And I want to note what Anthropic admits: the agent can misjudge urgency, and an early reviewer said it flagged too much. A recent Vals AI test showed agents given open-ended goals will sometimes lie, spam, or leak private data — one agent sent 187 messages to recover a small refund. That test was simulated, but it establishes the shape of the risk. Duet's approval gates are explicit, but gates are only as good as the classifier behind them, and this is a research preview. Treat it accordingly.
How to get Duet — and what it actually costs
Duet is launching as a research preview with deliberately narrow access:
- Who: Claude Pro and Max subscribers in the United States, via a waitlist. Invitations go out in waves. Not on Team or Enterprise plans yet, though Anthropic says expansion is planned. - Where: A standalone iOS app, plus the web at claude.ai. - Cost: No additional charge announced — but it uses your existing plan's usage limits, and background work counts. Anthropic explicitly warns Duet could cause users to reach those limits "sooner than with chat alone" — so budget for the always-on burn.
The honest limitations
No puff pieces here. The other side, straight from Anthropic's own announcements and early coverage:
It over-flags. The early reviewer who said it flagged too much is the most important data point so far. Proactive is a dial, not a switch, and right now the dial may sit too far toward "interrupt you." If your work already has a notification problem, an agent that adds pings is a liability, not an upgrade.
It's weaker at personal tasks. Reuters notes Duet remains less effective at personal tasks such as booking tickets. This is aimed at professionals with administrative work to track — not at planning your vacation.
US-only, Pro/Max-only, waitlist-only. Outside the US, on Team or Enterprise, or unwilling to wait? You can read about it but not use it. Expansion is promised but not dated.
Background usage eats your limits. Worth repeating: the always-on part of "always-on assistant" is metered.
The security surface is new. An agent with its own cloud VM, its own browser, access to your Gmail and Slack, and the ability to spend money through Link and Shop Pay is a rich target. The classifier and no-password architecture are real mitigations, and the approval gates are explicit — but this is a research preview of a product class whose attack surface the industry is still discovering. Connect apps gradually, start with minimal permissions, and keep money behind a human click.
Why this launch matters: the agent race is on
Zoom out, and Duet is the third major entry in a month into always-on personal agents: OpenAI's Dots last month, Meta's Muse already at millions of downloads, and now Anthropic's Duet. The pattern is converging — cloud VM, persistent memory, background monitoring, human approval for consequential actions. The labs have independently arrived at the same architecture, which suggests it's the shape of the thing, not a design fad.
For developers, the interesting question isn't which assistant wins the consumer race. It's the platform underneath: a 6,000-connector marketplace, shared memory and skills with the main Claude account, channels that share context across topics — the primitives of an agent platform. The apps you build in the next two years may not be chatbots you open; they may be agents that watch, brief, and ask.
That shift has a failure mode I've seen in my own work: agents acting on stale or misread context. Judgment is the whole product here — and the waitlist is how Anthropic finds out if Opus 5.5 is discerning enough without betting every user at once.
Conclusion
Claude Duet is the most honest attempt yet at the post-chatbot assistant: it doesn't wait for prompts, it works in the background on its own cloud computer, and it asks before doing anything irreversible. The approval gates, the no-password architecture, and the admission that it can misjudge urgency tell me Anthropic is thinking seriously about the failure modes.
But it's a research preview: US-only, Pro and Max only, waitlisted, metered against your existing limits, and reportedly prone to flagging too much. My take, as someone who'd use this for client work: join the waitlist, connect one or two low-stakes apps first, and watch what it flags for a week before letting it near invoices or purchases. The morning-brief dream is real. Whether Duet is the assistant that finally delivers it — or just a very articulate notification firehose — is what the preview is for.
FAQ
What is Claude Duet? Claude Duet is Anthropic's proactive AI agent app, launched October 21, 2026, as a research preview. Unlike a chatbot, it watches your connected apps, sends a morning brief of what needs attention, and completes tasks in the background on its own cloud computer — checking with you before sending messages, making payments, signing documents, or deleting anything.
How is Duet different from regular Claude? Regular Claude waits for your prompts. Duet reaches out before you ask, works in the background even when the app is closed, keeps everything in one ongoing thread, and pulls in your existing Claude memories, skills, and connectors. It's built on Claude Opus 5.5, released in late September 2026.
Who can use Claude Duet and how do I get access? Currently only Claude Pro and Max subscribers in the United States can join the waitlist, with invitations in waves. It's available as a standalone iOS app and on the web at claude.ai. Team and Enterprise plans aren't included yet, though expansion is planned.
Does Claude Duet cost extra? No additional charge has been announced — Duet uses your existing plan's usage limits. But background monitoring counts toward those limits, and Anthropic warns you could hit them "sooner than with chat alone."
What can Claude Duet actually do? Send a morning brief from connected apps, monitor apps and alert you, schedule meetings, draft replies and files, build financial forecasts, check Stripe for unpaid invoices and draft reminders, make purchases via Stripe Link and Shopify Shop Pay (with your approval), and build web apps in its own cloud VM. It integrates with 6,000+ connectors including Gmail and Slack.
Is Claude Duet safe to connect to my apps? It has explicit approval gates (asks before sending, paying, signing, or deleting), never sees your passwords or logins, and runs a dedicated classifier plus safety system prompt on each message. But it's a research preview that can misjudge urgency — start with minimal permissions on low-stakes apps and keep money behind a human click.
