Anthropic launches Claude Duet, an agent that works without being asked

Anthropic has opened a research preview of Claude Duet, a standalone agent app that checks a user’s connected accounts every few hours, sends a morning brief and messages the user when something needs attention, without waiting for a prompt. Built around Claude Opus 5.5, it runs on a cloud virtual machine with its own browser, so work continues after the user closes a laptop or phone, and it asks before it sends a message, makes a payment, signs a document or deletes anything.

Anthropic’s post opening the preview is dated 21 October 2026. Mashable reported the announcement as made on a Wednesday, while another summary of TechCrunch’s coverage gives Thursday, October 22. Access is a waitlist for US subscribers on Claude’s Pro and Max plans.

What Claude Duet does differently from earlier Claude products

The shift is from reactive to proactive. HackerNoon’s Sebastian Buzdugan, an ML engineer, notes that every earlier Claude product waited for the user to open a chat or a terminal and type; Duet reverses that, and Anthropic’s one-line pitch is an agent that “comes to you with what needs your attention.”

In practice, the loop has three parts, according to coverage of Anthropic’s post:

• A morning brief summarising what matters across connected apps.
• Periodic check-ins every few hours during the day, with a message when something needs attention.
• Delegated tasks such as scheduling, dashboards, forecasts, prototypes and web apps, handed over in a single long thread by text or voice.

Duet polls rather than watching accounts continuously. Anthropic’s stated reason, as HackerNoon reports it, is that constant checking would burn through plan limits much faster. Results come back as files such as a document, a spreadsheet or a slide deck.

Anthropic’s own examples, listed by Mashable, include replying to missed Gmail and Slack messages, running payroll in Gusto, drafting reminders for unpaid invoices found in QuickBooks or Stripe, drafting a job posting and posting it to LinkedIn after approval, building a product-launch campaign in Canva, and setting up and managing a Shopify store.

Claude Duet specs and access at a glance

Item	What the sources say
Model	Claude Opus 5.5
Where it runs	A cloud virtual machine with its own browser
Apps	iOS, Android and claude.ai
Plans	Pro and Max; not yet Team or Enterprise
Region	US only, by waitlist
Usage	Shares the plan’s usage limits; background checks count
Connectors	More than 6,000 connectors and plugins in the Claude Marketplace
Payments	Stripe’s Link wallet and Shopify’s Shop Pay

How Duet reaches a user’s apps and accounts

Duet carries over the memory, skills and connected apps already on a user’s Claude account. Anthropic names Gmail, Slack, QuickBooks, Stripe, Gusto, HubSpot, Docusign, Canva and Shopify among the more than 6,000 connectors and plugins it can use.

Anthropic says Duet never directly sees passwords or sign-in tokens. Purchases can go through Link, Stripe’s wallet: Duet confirms the item, Link asks the user to approve the spend, and payment uses a virtual card. HackerNoon adds that this means the agent never handles the user’s real card number. A richer Shopify editor and sign-in with saved 1Password logins are described as coming soon.

One Hacker News user who asked Duet about itself was told each session is a cloud Linux container with a shell, code tools and a browser; it refused to list its files or instructions, according to HackerNoon. That account comes from the agent’s own answer, not from Anthropic documentation.

The approval rules: what Duet does alone and what it never does

The permission design sorts actions into three tiers, as summarised in coverage of Anthropic’s post:

1. Acts on its own: reading, drafting and preparing.
2. Asks first: sending messages, making payments, signing documents and deleting files. Duet shows the step and waits.
3. Refuses even with approval: trading securities and moving money between accounts.

Two classifiers (small models that screen actions or messages) sit on top, according to HackerNoon. One auto-approves safe actions and is the same kind used in Claude Code’s auto mode; the other watches for data leaks. By default, the agent declines requests for credit card numbers or government identification. Health and personal-finance details go out only when the user has explicitly told it to send them.

HackerNoon also reports that Anthropic says the app was red-teamed, penetration-tested by an outside firm, and is in scope for Anthropic’s public bug bounty. None of these tests has been published as an independent report in the sources.

Who Duet is for, and where it sits among rival agents

Anthropic aims Duet at professionals and small businesses. It distinguishes Duet from personal agents such as Instinct, saying it is not designed for simple errands or booking tickets. Anthropic’s post itself says personal errands such as booking tickets are weaker in this early version.

The launch lands in a crowded field. Coverage lists Meta Muse, OpenAI Dots, Hark, Instinct, Grok Bot and Fo as rivals. Mashable reports that Meta CEO Mark Zuckerberg said Muse had been downloaded 10 million times in less than two months, and that the app has faced controversy, including reportedly creating files on users’ friends and loved ones and sending a Facebook Marketplace buyer to a user’s home. HackerNoon notes that OpenAI, Meta and xAI all shipped similar agents within the same few weeks.

Several of those rivals are also leaning into business use: Meta has a version of Muse tailored for small businesses, and OpenAI’s Dots keeps a business focus through connections to professional tools.

What Claude Duet means for builders

For developers, the more useful part of the launch may be the design pattern rather than the app. HackerNoon describes Duet’s shape as a scheduled agent with memory and a notification channel, and sums up the permission model as: read freely, gate every irreversible write, and hard-block actions that should never be automated.

Plenty of engineers already run homemade versions, Buzdugan writes: Claude Code on a schedule, a few MCP servers (Model Context Protocol servers, which connect a model to outside tools and data) and a message to a phone. His minimal example is a cron job that runs Claude Code in headless mode with a fixed prompt each weekday morning. A DIY setup gives full visibility into the prompt and files and a free choice of model or connector; what it lacks, in his account, is Duet’s approval gates and credential isolation.

Builders weighing the trade-off should note three points from the coverage:

• Shared quota. Background checks count toward the same plan limits, and larger jobs use more because they take more model turns. Hacker News commenters flagged that this eats into the quota used for coding.
• Opacity. Users cannot inspect the virtual machine, its files or the rules it has learned, commenters said, and Duet cannot yet show its browser.
• Lock-in. A persistent agent holding a user’s memory is tied to one vendor and one model family, another point raised in the Hacker News discussion.

Known limits and early reactions

Anthropic lists its own limits. The early version can misread a thread or misjudge what is urgent, and it lives in its own app, separate from the main Claude app. Because it polls, it can miss something that arrives between check-ins, HackerNoon points out.

Early reaction has been loud but anecdotal. According to HackerNoon, the launch reached the top of Hacker News the same evening with 470 points and over 400 comments, most of them about trust, limits and who owns the agent rather than the model. One commenter who tried it said the suggestions after setup were mediocre and that it dumped raw tool-use cards into the chat without explanation. Commenters also noted that many sites block agents or present CAPTCHAs.

No independent benchmark or structured hands-on test appears in the sources. Claims about what Duet does well rest on Anthropic’s descriptions and a handful of user reports.

What to watch next

• Wider access. Whether Anthropic extends Duet beyond the US and to Team and Enterprise plans, which it says are not yet included.
• Promised features. The richer Shopify editor and 1Password sign-in described as coming soon, and whether Duet gains a way to show its browser activity, which it cannot yet do.
• Usage-limit behaviour. How quickly background checks drain Pro and Max quotas in real use.
• Safeguards under pressure. Whether the approval gates and leak classifier hold up as the bug bounty and outside testers probe them.