Skip to content
NewsProduct launch

Anthropic launches Claude Duet, an agent that checks your inbox while you sleep

· by Pondero Newsdesk

The short version

Anthropic opened a research preview of Claude Duet on October 21, 2026, a standalone app that works through more than 6,000 connected apps, sends a morning brief, and pauses for approval before it sends a message, pays, signs, or deletes anything.

Anthropic launches Claude Duet, an agent that checks your inbox while you sleep

Claude now messages a user first, works for hours without a laptop or phone turned on, and still stops to ask before it sends a payment or signs a document, per Anthropic's own announcement of Claude Duet, a research preview that opened October 21, 2026.

What Duet actually does

Duet is a standalone app, separate from the main Claude app, available on iOS, Android, and claude.ai, per Anthropic. It runs on Claude Opus 5.5 and is included in Pro and Max plans in the US only, opening through a waitlist; Team and Enterprise are not part of this preview. Each morning it sends a short brief on what needs attention across a user's connected accounts, and through the day it checks those accounts every few hours (not continuously, Anthropic says, specifically to avoid burning through a plan's usage limits faster) and messages only when something needs a decision. Duet runs on its own virtual machine with its own browser, so a task keeps progressing after someone closes their laptop; a notification arrives when the work is done or needs review.

The connection layer is the Claude Marketplace's more than 6,000 connectors and plugins, including Gmail, Slack, Google Calendar, Google Drive, Intuit QuickBooks, Stripe, Shopify, HubSpot, Gusto, Docusign, Notion, Canva, Microsoft 365, and Ramp, per the same announcement. Anthropic lists concrete jobs it hands to those connections: finding missed Slack and Gmail messages and drafting replies, checking QuickBooks or Stripe for unpaid invoices, running payroll in Gusto with hours and bonuses filled in, drafting a job post for LinkedIn after a human edits and approves it, and preparing Docusign agreements for signature. Duet can also spend money on a user's behalf through Stripe's Link digital wallet or Shopify's Shop Pay, paying with a virtual card rather than exposing a real card number, and Anthropic says sign-in support for sites saved in 1Password is "coming soon."

The approval gate is the actual product, not a footnote

Anthropic drew a specific, named line rather than a vague "human in the loop" promise: before Duet sends a message, makes a payment, signs a document, or deletes anything, it shows exactly what it is about to do and waits for approval, per the company. A separate category, including trading securities and moving money between accounts, is described as off-limits even with that approval in place. For a tool being pitched at solo operators and small-business owners rather than engineering teams, that distinction matters more than the headline connector count: the product's entire premise is that someone can hand off real financial and administrative work without reviewing every step, and the places where Anthropic chose not to automate are the places it is telling buyers to still watch closely.

Anthropic also detailed a second, narrower control built specifically for an agent that reads email and messages people on a user's behalf: a dedicated classifier and safety system prompt screens each message and blocks the agent from sharing what it calls the strictest tier of information, including credit card numbers and government identification, unless the account owner explicitly instructed it to send that data. Looser categories, like health details or personal finances, pass through only with an explicit instruction too. Anthropic says the credential-handling design mirrors how Claude Managed Agents work, so Duet never directly sees a user's passwords or sign-in tokens, and that its defenses against prompt injection draw on the company's own browser-use research. Anthropic also said it ran red-team testing and had an external firm penetration test Duet's security infrastructure, and the product falls under Anthropic's existing HackerOne bug bounty program.

Context and reactions

Duet is Anthropic's answer to a format two other labs already shipped. Meta launched Muse on September 8, 2026, and OpenAI unveiled Dots at DevDay roughly three weeks later; both run on a dedicated cloud machine, keep memory of a user's preferences, and surface relevant information without being asked, per Engadget's comparison of the two. The three products diverge on specifics: Muse checks a user's Google Docs and email on roughly a 30-minute heartbeat and can go shopping or make travel plans directly, while each OpenAI Dot runs on GPT-6 Astra and connects to more than 4,000 apps through ChatGPT's plugin ecosystem. Anthropic's version checks in every few hours rather than every 30 minutes, and its connector count (more than 6,000) is the largest of the three as claimed by each vendor.

Anthropic also published its earliest testers by name rather than only citing anonymous feedback, which is a departure from how the company has rolled out past research previews. Boris Cherny, who leads Anthropic's Claude Code team, said he talks through his to-do list with Duet on walks instead of saving that work for his desk. Amin Ebrahimi, who owns Shahs of Kabob in South Florida with about 110 employees, said Duet handled accounting, customer support, and marketing work across his locations without him needing "100 tabs" open at once, per Anthropic's announcement. Anthropic was explicit that the preview still has rough edges: it can misread a thread or miscalibrate what counts as urgent, it was built around work rather than personal errands like booking travel, and it cannot yet show a user its own browser session.

What to watch next

Anthropic has not given a timeline for Team and Enterprise access, for expanding past the US, or for letting Duet share a single app and memory with the main Claude product rather than living as a separate download. The specific test worth tracking is false-positive rate: a proactive agent that pings a user about something that did not actually need attention burns the trust the entire pitch depends on, and Anthropic's own framing (tuning "based on what testers tell us") signals that calibration is still a work in progress rather than a solved problem.

Sources