Claude Duet is Anthropic's proactive AI agent: it contacts you first, works in the background on its own cloud computer and asks for approval before certain consequential actions. This guide covers what it does, who can get it, what it costs and where it falls short - including why you cannot get it in the UK yet, checked against Anthropic's own pages.
Anthropic launched Duet as a research preview on 21 October 2026. You need three things to use it: a paid Claude plan, a Claude account in the US and a waitlist invitation. A UK reader fails the second test, there is no UK date and Anthropic states no price in pounds.
Last checked: 5 November 2026, against the Anthropic announcement, safety page, help articles, pricing pages, terms and privacy policy, plus press reports and named reviewers. Duet is not available here, so nothing was tested on a UK account or hands-on. We update this guide when availability or prices change.
What is Claude Duet?
Duet is an agent, which means you give it a job and it takes the steps itself - our guide to what an AI agent is covers the idea properly. The word proactive describes who starts the conversation: Anthropic's product page says the Claude app waits for you and Duet comes to you, flagging things it thinks need attention.
Duet is a standalone iPhone and Android app, and on a computer it opens on the web at claude.ai; Anthropic lists iOS, Android and the web, and does not list a dedicated desktop app. It runs on its own cloud computer with its own cloud browser, so it carries on with your phone or laptop off. It uses Claude Opus 5.5, it picks its own effort level and you cannot choose the model.
Per the launch announcement it shares your Claude account's memory, skills and connected apps, and Anthropic has not spelled out how the two apps will evolve together. Anthropic says Duet is built for "professionals and small business owners" - someone using a chatbot at home is a different audience, and we come back to that below.
What Claude Duet does: the brief, the checks and the Make tab
Anthropic describes the work in three layers. A morning brief, built from your connected email and calendar, then updated through the day with follow-ups on quiet threads. Background checks: it looks through connected accounts and flags what it judges to need attention. And delegated tasks: it reads, drafts and prepares work on its own, then stops for approval before certain actions - and if you cannot respond at that moment, it parks the request and carries on with other work, bringing the approval back later.
The output side is easy to miss in the coverage: finished work lands in a Make tab as documents, spreadsheets, decks and dashboards, with figures cited back to their sources, and a Tasks tab tracks what is in flight - only you can mark a task complete. It can buy things through Link by Stripe and Shop Pay (available only inside Duet for now), and it uses its own cloud browser for sites with no connector. Anthropic says you can type or record a voice message, though Dan Shipper's review on Every says there is no real voice mode, only transcription.
Is Claude Duet always-on?
Only in a limited sense. Anthropic's product page says "always on", but its announcement says Duet checks connected accounts "every few hours", and no exact interval is published. Those checks draw on your plan limits too. So it is a background worker, not a live monitor - Anthropic itself notes that urgent items can arrive between checks. You can tell Duet what you want to hear about and how often, but that covers its messages to you; no page we read says you can change how often it checks your accounts.
Which apps work with Claude Duet?
Anthropic claims "over 6,000 connectors and plugins" via the Claude Marketplace. It publishes no itemised list against which to check that figure, and press coverage has not reconciled it either, so treat the number as Anthropic's.
The named services skew hard to work tools - Gmail, Google Calendar, Google Drive, Slack (which "works with limits"), QuickBooks, HubSpot, Docusign and Calendly among them, with 1Password sign-in marked "coming soon". The notable gap: Microsoft 365 "isn't available yet" in Duet according to the help page, even though the product page shows a Microsoft 365 logo in its connector strip - so people who run their working life in Outlook are left out for now. Shipper also reports that scheduled runs cannot reach custom connectors. You start each connection and can disconnect it; no list of services aimed at UK users turned up.
Who can get Claude Duet? The US waitlist
Anthropic's help page requires all three of these:
| Requirement | Detail | Position for a UK reader |
|---|---|---|
| Paid plan | Pro, Max or Ultra. Free does not qualify. Team and Enterprise are excluded for now, with support planned but no date | The UK is on Anthropic's supported list for Claude, but a plan alone does not give you Duet |
| US account | "Available in the US only during the research preview" | Not met |
| Waitlist invitation | Sent in waves, with no guaranteed date | A waitlist place would not meet the US account rule |
The invitation is tied to your waitlist email, which must match your Claude account or the Apple or Google account on your phone. Anthropic publishes no waitlist length.
Is Claude Duet available in the UK?
No. The UK is on Claude's supported list in general, but that does not extend to Duet. We found no UK or EU date in the Anthropic material we reviewed, and nothing in it mentions UK or EU regulation of Duet, so we are not printing a cause or an estimate.
Link, Duet's payment route, also needs a US payment method. There is also a terms-of-service question worth watching, without reading a contradiction into it: Anthropic's UK consumer terms (effective 8 October 2025) restrict its consumer services to non-commercial use - "You agree not to use our Services for any commercial or business purposes" - while Duet is pitched at professionals and small business owners. Those terms predate Duet, Anthropic's legal pages separate its consumer services from its commercial offerings, and we found no Duet-specific clarification for UK business users. So it is simply not yet clear which terms would govern a UK business user if Duet launches here - something to re-read on arrival rather than a problem today.
How much does Claude Duet cost?
Anthropic makes no extra charge "during the research preview", but you must already be paying for a plan. There is no end date for the preview, no later price and no promise that Duet carries on in this form. Prices below are as listed on 5 November 2026, in US dollars excluding tax.
| Plan | Listed price | Duet |
|---|---|---|
| Free | $0 | Not included |
| Pro | $20 a month, or $17 a month billed $200 annually | Eligible - US only, by invitation |
| Max | $100 a month | Eligible - US only, by invitation |
| Ultra | $200 a month | Eligible - US only, by invitation |
There is a pound figure after all: the pricing page served to a UK visitor lists Pro at £15 a month billed annually, or £18 monthly, with 20% VAT included - which is what a UK subscriber actually pays for the plan Duet would ride on, if it ever reaches here. You may also see a "$200 Max plan" or a "Max 20x" tier mentioned in older coverage - both are out of date. Max is $100 and the $200 tier is now called Ultra.
How much of your plan limit does Duet use?
We cannot tell you precisely, and neither does Anthropic. Duet draws on your normal limits - the five-hour session limit and the weekly limits - at the same rate as chat, background work counts, and Anthropic warns "you may reach your limits sooner". Engadget warned that sharing limits with Claude Code could drain them fast. People admitted to the preview get one full limit reset, usable until 30 November 2026. We did not find what Duet does when a limit runs out, so we cannot say whether briefs and checks stop.
What can Claude Duet do without asking you?
Anthropic says Duet reads, summarises, drafts and prepares on its own. Its announcement says it waits for your approval before sending a message, making a payment, signing a document or deleting anything, and its help page says it is "designed to" do so. Its safety page says that before sensitive actions the prompt shows the action, the destination and the content, including every recipient. Reading needs no approval at all - which means your connected inbox and calendar are in play from the moment you link them.
The pages do not fully agree. The announcement and help page say "a message", while the safety page, dated 23 October, says "a potentially sensitive message"; Anthropic has not spelled out which messages fall outside that. The same page describes "don't ask me again" rules and says these never cover paying or sending money, deleting an order or creating an account - but does not list what else they can cover. Anthropic also calls approval fatigue "a serious issue" and says the calibration is still changing.
At least three categories are explicitly off-limits even with your approval: trading securities, moving money between accounts and accessing someone else's account. Routines running while you are away have the same safeguards.
A warning on health and financial details
Anthropic says these go out only with explicit approval, then adds that "your own request counts as consent" - so asking Duet to pass something on may be all the approval there is. Those rules cover outbound sharing; we found no page limiting what Duet can read in a connected account. Anthropic's general help tells users to "be mindful about sensitive information", including health records and financial information. That is a reason to keep accounts holding those details away from it.
Is Claude Duet private? What Anthropic can see
Training. No Duet page says whether Duet data trains Anthropic's models, so the general rules apply until Anthropic says otherwise. Those rules: Anthropic's pricing table lists model training as opt-out on every consumer plan, with the toggle under Settings, Privacy, Help Improve our AI models - and Anthropic's connector help page is explicit that it does not train on Gmail, Google Calendar or Google Drive connector data - with the exception that content you copy into a conversation, or that appears in Claude's replies, can be used if training is on. Flagged-for-safety conversations and material you report can be used regardless. Whether a Duet brief built from your inbox counts as connector data or as a reply is exactly the kind of edge Anthropic has not addressed - our inference is that the connector exclusion applies, but it is an inference, not Anthropic's statement.
What it keeps. Duet keeps history and memory until you delete or reset them, and deleting your account wipes what it stored. Duet also searches past Duet chats and non-project Claude chats unless you switch that off. Retrieved connector data is stored with the chat and deleted with it.
Disconnecting is not clean. Three separate stores are in play here: the connected account itself, your chat history and Duet's memory - and they empty separately. Removing an app stops new access, but what Duet has learned stays until you reset memory - Anthropic's help page says so plainly. Your connected apps carry over between Claude and Duet, and switching an app off in one place switches it off in both; chat search runs on the same Search and reference chats setting as Claude. Check what a reset removes from your everyday Claude before you use it.
Passwords and cards. Anthropic says the model never directly sees passwords or sign-in tokens - passwords go into a separate form that encrypts them with a single-use key, straight to the browser. A separate classifier checks outgoing messages, and Duet is instructed never to share sign-in credentials, full card numbers or government ID numbers; Anthropic's own caveat is that the classifier "will occasionally miss something or flag a harmless message".
Other people. An inbox holds other people's messages too. Anthropic's privacy policy says you are responsible for having the necessary rights for what you connect, and that third-party services process what Claude sends them under their own policies. For UK users the controller is Anthropic Ireland, with UK GDPR requests answered within a month - though the policy warns "these rights are limited" and names no Duet specifics.
Where Claude Duet falls short
Anthropic's own pages list most of the gaps, which is to its credit:
Misjudged priorities. It may misread a thread or get urgency wrong. Personal errands. Tasks such as booking tickets "aren't quite as strong" as work tasks - Anthropic's words. Nothing to watch. You cannot view or take over its browser, unlike Meta's Muse, though you can steer a task through the task list. Gaps in reach. It can read but not edit your existing Claude work, it cannot see your computer's files, and some large websites chose to block it before launch. Calls and texts. It cannot phone or text other people, and it can email only from connected accounts. Repeat payments. Link has no recurring payments.
Named reviewers add more, and their findings sort into three themes. The consensus: it is polished and genuinely useful for professional workflows - Claire Vo (How I AI, 26 October) called it the "most polished personal agent launch yet", and Dan Shipper (Every, 21 October, with disclosed early access) "the first personal agent I'd call a pro tool". The weak spots: the proactive side underdelivers - Vo found parallel tasks "routinely took more than 10 minutes", nudges faded after onboarding, triage missed priorities and a sales deck came out off-brand, though a pipeline analysis ran quickly. And the disagreement: nobody agrees where it ranks - Shipper rated its email drafts below OpenAI's dots, while Ejaaz Ahamadeen (Yahoo Finance, 26 October) found Meta Muse stronger for personal use. VentureBeat noted no benchmarks, no task-cost estimates and company-selected customer stories; Engadget said its usefulness depends on how far people trust Anthropic with their digital lives.
What has gone wrong with Claude Duet so far?
We found no reported incident and no independent security research specific to Duet - but it is a fortnight old and limited to invited US users, so a short record proves little, and it should not be read against Muse's longer incident list as a clean bill of health. The safety case so far rests on Anthropic's statements: red teams and an unnamed external firm tested it, a bug bounty runs on HackerOne and a follow-on test of the cloud browser and sign-in form is running now; prompt-injection defences exist - alongside Anthropic's own admission that "no defense is perfect". VentureBeat noted that audit logs and custom retention sit under Enterprise on Anthropic's pricing page, and Duet is not on Enterprise plans; the task list shows what Duet has done, but nothing we read says how complete that record is, so check the account itself.
Who pays if Claude Duet gets a purchase wrong?
Anthropic's consumer terms put the risk on you: "You are responsible for all Inputs you submit to our Services and all Actions", Actions "may not be error free", and other than your legal rights your use of them is "solely at your own risk", with liability capped at the greater of six months' fees or £100 in the UK-resident version. Your statutory rights survive, the terms list exceptions to the cap - and contract wording is the start of the analysis, not the end: statutory consumer rights and payment-scheme protections sit outside any company's terms.
On the payment mechanics, the design is genuinely careful: Link creates a one-time virtual card for each purchase, and Anthropic says Duet "is designed to ask you before each purchase and not submit the Link payment until you approve the maximum amount". Anthropic's Link help page mentions no buyer protection, and the US-centred cover described on Link's own pages should not be assumed to apply to a UK reader. Returns and refunds go to the store, wrong-item or wrong-store mistakes by Duet to Anthropic's own support team and duplicate charges to Link Support.
Claude Duet vs ChatGPT Work, Meta Muse and OpenAI dots
Duet sits with the always-on group, alongside Muse, Gemini Task and dots - none of which has reached UK consumers. What UK users can get is the supervised kind, which you set going yourself (access and prices as read on the makers' pages, 5 November 2026):
| Product | Type | UK consumer access | Listed price |
|---|---|---|---|
| Alexa+ | Assistant with narrow real actions | Request-based Early Access | Free in Early Access, then free with Prime or £19.99 a month |
| ChatGPT Work | Supervised agent mode | Yes - limited on the free desktop plan, expanded on Plus | Plus £20 a month in the UK |
| Claude Cowork | Supervised | Yes, on all paid Claude plans; note it defaults to Auto mode, which does not stop at every step | From $17 a month billed annually; no UK price list published |
| Claude Duet | Always-on, checks every few hours | No - US account + waitlist | No extra charge during the preview, on plans from $17 |
| Gemini Task (formerly Spark) | Always-on | No - Google excludes the UK by name | Google AI Pro at £18.99 a month here does not include it |
| Meta Muse | Always-on | No - no UK date from Meta | Free, $20 or $100 a month where live |
| OpenAI dots | Always-on | Not on personal plans - the only UK route is a business workspace Premium seat at £75 a month (minimum two seats) | Consumer Pro from £89 a month, UK excluded from dots |
On control, our Meta Muse guide notes that Muse lets you watch and take over its browser, which Duet does not. We have seen no independent side-by-side test, so we are not saying any of these does the job better. If you want something today, ChatGPT Work and Claude Cowork both run multi-step jobs and scheduled tasks you set up - and our comparison of personal AI agents for UK users goes through what you can actually buy.
Is Claude Duet worth joining the waitlist?
On Anthropic's own account, Duet suits a US professional or small business owner whose work runs through Gmail and Google Calendar, with plenty of threads to keep track of and nothing too urgent between checks. Shipper's "pro tool" line fits that picture; Vo's findings on triage and nudges are a warning about the same work. It is a poor match for home life - Anthropic says personal errands are the weaker side - and for anyone who needs instant alerts, runs on Microsoft 365, works in a team or wants to watch an agent work.
For a UK reader there is nothing to decide yet: joining the waitlist from here would not meet the US account rule, and a paid Claude plan here does not include Duet, so do not buy one for that reason. For a US subscriber, Duet costs nothing extra during the preview; the bigger cost is the access you hand over, against an evidence base of a handful of early reviews, no independent security work and approval rules Anthropic says are still changing. Treat it as an experiment on a low-stakes account: turn off model training if you do not want eligible conversations used to improve Anthropic's models - but do not treat that toggle as a general privacy or retention control, because storage is a separate question - connect one account with nothing medical or financial in it, skip the "don't ask me again" rules, read the sent folder rather than the summaries and reset memory after disconnecting an app.
Frequently Asked Questions About Claude Duet
How do I get Claude Duet access?
You need a Pro, Max or Ultra plan, a Claude account in the US and a waitlist invitation. Invitations go out in waves with no guaranteed date, and the invite email must match your Claude account or your phone's Apple or Google account.
Can I use Claude Duet in the UK?
No. Anthropic says it is "available in the US only during the research preview", with no date given for the UK or EU. Having a paid Claude plan in the UK does not change that, so do not buy a plan to get Duet.
Does Claude Duet cost extra?
Not during the research preview, though you need a paid plan - listed from $17 a month billed annually, in US dollars and excluding tax. It shares your plan's usage limits and its background work counts against them. Anthropic promises nothing about pricing after the preview.
What can Claude Duet do without asking me?
It reads, summarises, drafts and prepares work on its own. Anthropic says it is designed to ask before sending a message, paying, signing or deleting - though its safety page narrows that to "a potentially sensitive message" and allows "don't ask me again" rules, which never cover payments, deleting an order or creating an account.
Is Claude Duet always on?
Not in the sense of live monitoring. It keeps running in the cloud but checks your accounts "every few hours", with no exact interval published, and Anthropic itself notes that urgent items can arrive between checks.
How safe is Claude Duet?
The safety design is substantial on paper - approval gates, hard blocks on trading and money transfers, one-time virtual cards, a password form the model never sees - but the claims are Anthropic's own, the external test firm is unnamed and no independent security research on Duet has been published. On prompt injection, Anthropic itself says no defence is perfect.
Does Claude Duet train on my data?
No Duet page says either way. Anthropic's general rules: model training is listed as opt-out on consumer plans, with the toggle under Settings, Privacy, Help Improve our AI models, and connector data from Gmail, Calendar and Drive is excluded from training - except content you paste into a chat or that appears in Claude's replies. Flagged conversations can be used regardless of your setting.