Most AI assistants wait for you to type. Claude Duet, announced by Anthropic on Wednesday, October 21, 2026, is built to do the opposite: it scans your connected accounts, sends a short morning brief, checks in every few hours, and only asks for your attention when something needs it. It is a research preview on a waitlist for US Pro and Max subscribers, runs on Opus 5.5, and is available on the web and in mobile apps (outlets differ on whether Android is included at launch) (Source: AI Weekly).
A proactive agent raises questions a chatbot never did. What can it touch? What happens when it is wrong? Who approves the payment? This explainer pulls together what Anthropic and the early coverage say, including the limits the company itself lists, so you can decide whether to join the waitlist.
Quick summary: Claude Duet was announced on October 21, 2026 as a research preview for US Pro and Max subscribers, by waitlist, on the web and mobile apps (Android availability varies by report), powered by Opus 5.5. It works with more than 6,000 connectors and plugins, sends a morning brief, and asks for explicit approval before sending messages, paying, signing or deleting. Background checks count against the same usage limits as regular Claude chat, with no extra cost during the preview. Anthropic says Duet can misread a thread or misjudge urgency. Team and Enterprise plans are not yet included.
What Claude Duet is
Duet is aimed at professionals and small business owners who have more to track than they can hold in their head. It connects to the apps you already use, flags what needs attention, and handles tasks by message, whether by text or voice, much like messaging a teammate. It keeps everything in one ongoing thread and returns finished files such as documents, sheets or decks (Source: SiliconANGLE).
Setup is meant to be minimal because Duet draws on the memory, skills and connected apps already in your Claude account. It runs in the cloud, uses a virtual machine with its own browser to finish tasks, and does not need access to your local computer (Source: MacRumors). For sites without a connector, it uses a cloud browser (Source: VKTR). MacRumors framed the product as Anthropic's answer to proactive assistants from Meta and OpenAI (Source: MacRumors).
What it does day to day
- Morning brief: scans email and calendar and highlights priorities, per VKTR, and checks connected accounts every few hours.
- Tasks by message: example jobs in AI Weekly's summary include finding unpaid invoices, preparing payroll for review and drafting decks or dashboards.
- Bigger builds: SiliconANGLE lists interactive dashboards, financial or sales forecasts and prototype web apps.
- Connectors: more than 6,000 connectors and plugins from the Claude Marketplace, with examples including Gmail, Slack, QuickBooks, HubSpot, Docusign and Gusto (VKTR). Lists of named integrations differ between outlets, so check the current connector list before you rely on a specific app.
- Purchases: VKTR reports Duet can buy through Stripe's Link and Shopify's Shop Pay, with each purchase approved by you and paid with a one-time card.
The safety rules Anthropic describes
Duet is designed to pause before consequential actions. It asks for approval before sending a message, making a payment, signing a document or deleting a file (Source: MacRumors). According to VKTR, some actions, such as moving money or accessing another person's account, are blocked even with approval, and credentials are handled so that Duet never directly sees passwords or sign-in tokens (Source: VKTR).
An automatic classifier and a dedicated safety prompt check outbound messages for leaks of sensitive information, and VKTR adds that a new classifier is meant to reject requests to share credit card numbers or government IDs. Anthropic also cites defenses against prompt injection drawn from its browser-use research, adversarial testing and bug-bounty coverage, and says it will share more about its safety approach 'soon' (Source: VKTR). Prompt injection means hidden instructions on a web page or in an email that try to hijack an agent, which is why a proactive agent reading your inbox is a more serious safety case than a chatbot answering questions.
| Question | What the coverage says | Source |
|---|---|---|
| Who can join | US Pro and Max subscribers by waitlist, opening in waves; Team and Enterprise not yet included | AI Weekly, VKTR |
| Model | Opus 5.5; no model picker in the preview | VKTR |
| Cost | No extra cost in the preview; shares your plan's usage limits | AI Weekly, VKTR |
| Platforms | Web and mobile apps (outlets differ on Android) | SiliconANGLE, VKTR |
| Before acting | Asks approval to send, pay, sign or delete | MacRumors, AI Weekly |
| Can it call or text? | No phone calls or texts in the preview | VKTR |
| Can you watch it work? | No; you cannot see or take over its browser | SiliconANGLE, VKTR |
| Monitoring | Checks every few hours, not 24/7 | VKTR, MacRumors |
The limits Anthropic admits to
The honest part of this launch is the list of what Duet does not do yet. AI Weekly relays Anthropic's warning that it may misread a thread, misjudge urgency or miss something that arrives between check-ins (Source: AI Weekly). VKTR adds that preview users cannot choose a model, watch Duet work or take over its browser, and that Duet cannot make phone calls or send texts, see all of a user's projects or files, or monitor accounts around the clock. Anthropic advises reviewing drafts and approval requests before you confirm them (Source: VKTR).
SiliconANGLE notes that Duet is designed for work, so personal errands such as booking tickets are not treated as urgent, and that the research preview is meant to reveal use cases and failure modes before Anthropic refines the product (Source: SiliconANGLE).
Usage limits and the free reset
One practical catch: Duet's background checks count toward the same Pro or Max caps as normal chat. VKTR reports that this includes the five-hour session and weekly limits, and that each new user gets one full limit reset available until November 30, 2026. That detail appeared in one outlet, so confirm it in the app before planning around it. A proactive agent that runs several times a day can use up a plan's allowance faster than you expect, so watch your usage meter during the first week.
Before you grant Duet access, make a short list of which accounts it should touch first. Start with read-heavy work like email and calendar triage, and leave payments and signing for later, after you have seen how often it misjudges urgency.
Should you join the waitlist?
If you are a US Pro or Max subscriber with a busy inbox and a dozen connected tools, the waitlist costs nothing and the approval prompts are a reasonable guardrail. If your work involves other people's confidential data, treat the research label seriously: the product is explicitly experimental, and the first few weeks are about finding failure modes. Outside the US, or on a Team or Enterprise plan, you cannot join yet.
FAQ
01When was Claude Duet announced?
On Wednesday, October 21, 2026, with coverage continuing on October 22.
02Who can use it?
US Pro and Max subscribers, by waitlist, in a research preview. Team and Enterprise plans are not yet included.
03Which model powers it?
Opus 5.5. There is no model picker in the preview.
04Does it cost extra?
No extra cost during the preview, but its background checks count toward your normal Pro or Max usage limits.
05Can it spend my money without asking?
According to the coverage, it asks for approval before sending messages, paying, signing or deleting, and some actions such as moving money are blocked even with approval.
A proactive agent is one tool; many people also want to compare how Claude, GPT, Gemini and others handle the same inbox-style task. LumiChats puts 40+ premium AI models side by side in one subscription, so you can test a prompt across them before you hand any single assistant the keys.
