Claude Duet is Anthropic's proactive agent app. It launched on Oct 21, 2026 as a waitlist research preview for US Claude Pro and Max subscribers, on iOS, Android and claude.ai. Duet watches the work apps a user connects, sends a morning brief and finishes tasks in the background on its own cloud computer, such as scheduling meetings, chasing unpaid invoices, running payroll or setting up a Shopify store. It asks for approval before it sends, pays, signs or deletes anything.
It is still early. Across sites that use Verisoul, claude_duet made up under 1% of the personal-agent sessions we named between Oct 29 and Nov 1, 2026. Anthropic's other browsing agent shows up more often: the Claude desktop app's browser made up 3% in the same window, and Verisoul names it separately as claude_desktop.
Duet matters anyway because of how it is built. It signs in, buys with one-time virtual cards and does business admin, which puts it on exactly the pages fraud teams care about, and its waitlist will open up.
How Claude Duet shows up on your site
Duet's browser runs in a dedicated cloud virtual machine kept separate from the model, which controls it only through open, click and type commands. Cookies stay on the browser machine, and users can't watch or take over the browser yet. Anthropic's safety, privacy and security article says the browser sends a distinct User-Agent and that Anthropic publishes the IP addresses Duet browses from, though the article doesn't link the list.
The exact string has been seen in the wild. A GitHub issue on Adobe's RUM collector from Oct 30, 2026 recorded Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko; compatible; Claude-Agent/1.0) Chrome/156.0.0.0 Safari/537.36 in production traffic, and generic parsers had been reading it as plain Linux desktop Chrome.
| Check | What Claude Duet does |
|---|---|
| User agent | Linux Chrome with a Claude-Agent/1.0 token |
| Web Bot Auth signature | None documented by Anthropic |
| IP ranges | Anthropic says it publishes them, list not linked yet |
| robots.txt | Aims to follow site rules for agents, details unpublished |
Why Claude Duet is easy to detect, with one catch
Duet declares itself, which already puts it ahead of most personal agents. A plain string match on Claude-Agent/ catches honest Duet traffic, and Anthropic also contacted several large sites before launch and let them opt out, which shows it expects sites to make decisions about it.
The catch is that a user agent is a claim, while a signature is proof. Without a documented Web Bot Auth signature, anyone can paste Claude-Agent/1.0 into a script, so a match should be confirmed against the source network and the session's behavior before you trust it. Keep Duet separate from ClaudeBot, Claude-User and Claude-SearchBot as well, since those are Anthropic's crawlers and fetchers with their own rules.
How Verisoul detects Claude Duet
Verisoul treats Duet as self-declared and then checks the declaration. The network signal confirms the source IP belongs to the operator's own network, and the behavioral signal confirms automated clicks, typing and page flow instead of a person's. A Claude-Agent string arriving from the wrong network or with human-looking input is treated as a spoof, and device signals are added when automation flags show in the browser.
"agent": {
"id": "claude_duet",
"name": "Claude Duet",
"provider": "Anthropic",
"type": "ai_assistant",
"detection_method": "self_declared",
"detection_signals": [
"network",
"behavioral"
]
}idisclaude_duet, the name to put in your rulestypeisai_assistant, an agent acting for one persondetection_methodisself_declared, meaning Duet announced itself and Verisoul verified that announcementdetection_signalslistsnetworkandbehavioral, the two checks that confirmed it
See the Agent Atlas entry for Claude Duet for the sample response and the AI agent detection docs for the full list of methods.
Claude Duet was <1% of the personal-agent sessions Verisoul named, Oct 29 to Nov 1, 2026
Share of sessions with a named personal agent on sites that use Verisoul, %. Oct 29 to Nov 1, 2026
How to detect Claude Duet step by step
- Install Verisoul on sign-up, login and checkout so every session returns an agent field next to the bot score
- Look for
agent.idequal toclaude_duet, and keepclaude_desktopsessions in a separate bucket - Write a rule per id in your decision engine and treat a
Claude-Agentuser agent that Verisoul doesn't confirm as suspicious - Review Duet on account creation, payouts and account changes, where an agent's speed matters most
Should you block Claude Duet?
Duet acts for a paying Claude subscriber and asks before it pays or sends anything, so most teams allow it to browse and check out. It signs in through an encrypted password form on the user's device that the model never sees, and it pays with Link one-time virtual cards or Shop Pay, which limits how far a mistake can go.
Add review on sign-up, payouts and account changes, and treat a Claude-Agent user agent from the wrong network as a spoof. If you would rather keep Duet off your site entirely, a rule on the claude_duet id does that without touching Anthropic's crawlers.