Claude in Chrome is Anthropic's browser extension that lets Claude read pages, click, type, navigate and fill forms in the user's own Chrome tabs. It started as a research preview for 1,000 Max plan users on Aug 26, 2025, opened to all Max subscribers on Nov 24, 2025 and to every paid plan on Dec 18, 2025, and became generally available on Aug 26, 2026 with an autonomous mode in which a classifier checks each action before it runs. People can drive it from a side panel in Chrome or from Claude desktop and Claude Code on the same machine.
It is by far the most installed agent in this group. The Claude extension's Chrome Web Store listing showed about 20 million users on Oct 31, 2026, and because it runs in whatever sites the person already uses, it turns up across SaaS apps, ecommerce, internal admin portals and anything with a long form. Teams want to know whether a burst of form fills on a real customer's account came from that customer or from Claude acting for them.
How Claude in Chrome shows up on your site
The extension runs inside the person's normal, visible Chrome window on their own computer, with their cookies, local storage and IP. Anthropic does publish user agents and an IP list for its fetchers and crawlers, such as Claude-User and ClaudeBot, but those cover separate traffic and not the extension, which sends the user's ordinary Chrome user agent and signs nothing. The extension drives the page through Chrome's debugger API, and while the agent works it shows a Stop control on the page by injecting an element with id claude-agent-stop-container into document.body. A style element, claude-agent-animation-styles, usually stays in the page after the agent finishes.
| Check | What Claude in Chrome does |
|---|---|
| User agent | The user's own Chrome user agent |
| Web Bot Auth signature | None |
| IP ranges | None, traffic uses the user's own network |
| robots.txt | Not documented for the extension |
Why Claude in Chrome is hard to detect
Every network-level signal belongs to the person. The IP, the cookies and the browser are theirs, so header and IP rules see a returning customer, and the agent can switch on and off in the middle of a session when the person hands it a task. The extension also works in Chrome only, on desktop, which makes the browser itself no clue at all, since that is what most of your visitors already use.
Behavior is where it differs from a person. Researchers at UC Davis found in their FP-Agent study that Claude in Chrome fills fields by default with change events and no keystrokes at all, and that when told to type, its key timings come in under 1 millisecond, which no human hand produces. The injected claude-agent- elements make it easier to spot than most in-browser agents, because a page script can watch for them with a MutationObserver. Element names and input traits can both change with any release of the extension, so they work best as inputs to a score rather than as a fixed rule.
How Verisoul detects Claude in Chrome
Verisoul does not name Claude in Chrome yet, because the extension declares no identity and signs nothing, and we only attach an agent name when we can confirm it. What Verisoul does see is the automation the extension leaves in the browser while the agent acts, including its injected page elements, and the machine-like input, such as fields filled without keystrokes, and both push the session's bot score up while the agent field stays empty.
"agent": nullbot_scoreruns from 0 to 1 and rises as the session looks automated, so 0.97 means the input and page flow look like softwareagentisnullbecause the extension does not declare or sign an identity, so Verisoul does not attach a name- The network looks like a normal visitor, since requests come from the user's own IP, which is why the score leans on device and behavior
Anthropic's Claude desktop app is a different case. When the desktop app browses for a person it declares itself, and Verisoul names it as claude_desktop from automation patterns of an app running on the person's computer. Across sites that use Verisoul, claude_desktop made up about 3% of the personal-agent sessions we named between Oct 29 and Nov 1, 2026. The Agent Atlas entry for the Claude desktop app lists its signals, and the AI agent detection docs explain both the agent field and bot score.
Claude desktop app was 3% 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 in Chrome step by step
- Add the Verisoul SDK to the forms and account pages where agents do the most work, following the AI agent detection docs
- Look for known accounts where
bot_scoreis high andagentis null, which is where Claude in Chrome sessions land - Tag
claude_desktopsessions separately, since that is Anthropic's desktop app browsing with a declared identity - Review payouts, password changes and bulk form submissions when either pattern appears
Should you block Claude in Chrome?
Usually not. Claude in Chrome acts inside a paying user's own logged-in browser, so blocking it mostly means making your own customers do a chore by hand. Browsing, search, checkout and routine form fills can stay open.
Sensitive steps are different. Payouts, password and email changes and large batches of form submissions are where an automated session on a trusted account deserves a review, and sign-up deserves the same care because the extension can create accounts from a residential IP that looks like a new customer.