Detection

Detection labels traffic; it never blocks it. The SDK uses five sources, from strongest to weakest:

init() runs detection automatically (autoDetect, on by default) and keeps the result in client.agentContext, so every tool call is attributed without you passing identity around. Call client.detect() to re-run it.

const identity = await client.detect()
// → { class: 'dom', trust: 'detected', confidence: 0.72, signals: { webdriver: {…}, automationGlobals: {…} } }
// or null — a human, or an agent that hasn't shown itself

WebMCP agents

document.modelContext exists for every visitor of a WebMCP-capable browser (Chrome and Edge origin trials, ChatGPT Desktop, …), so its presence alone says nothing about an agent. The SDK reports a WebMCP agent once one acts: a toolactivated event, or a call to a tool you published with publishToWebMcp(). From then on the session is attributed to that agent (unless a stronger, declared or verified identity is already in place).

webMcpSupport() tells you whether the page can serve tools at all: 'document', 'navigator' (the old alias, removed in Chrome 152), 'legacy' (old polyfills) or null.

Vision-model agents

Computer-use agents (ChatGPT agent, Claude computer use, Gemini computer use, …) look at screenshots and drive a real browser with synthetic mouse and keyboard input. They often pass DOM checks, but they interact differently from people:

• teleport clicks — a click with no pointer movement before it;
• centre clicks — clicks within a pixel or two of an element's exact centre, again and again;
• metronome typing — key intervals with almost no variance, or text inserted with no key presses;
• no hover, no wheel — clicks and scroll jumps with no mousemove or wheel events between;
• VM-shaped viewport — exact sandbox sizes such as 1024×768 or 1280×800 (a weak signal on its own).

The monitor scores these over a short window and reports a screenshot agent only when several agree after at least three interactions. It reads event timing and coordinates — never key values or text — and ignores synthetic (isTrusted: false) events, which are DOM automation instead.

Agentronics.init({
  publishableKey,
  visionDetection: { threshold: 0.7, minInteractions: 4 }, // or false to turn it off
})
 
client.vision?.score()   // live 0–1 score
client.vision?.signals() // { clicks, teleportClicks, centreClicks, keyIntervalCv, … }

DOM / headless

A weighted bag of signals: navigator.webdriver, ChromeDriver's cdc_* globals, Playwright / Puppeteer / Selenium / PhantomJS / Nightmare globals, a headless User-Agent, empty navigator.languages or plugins on desktop, a Chrome UA without window.chrome. Raise dom.threshold (default 0.4) to be stricter.

Crawlers

detectCrawler() matches navigator.userAgent against AI crawlers (GPTBot, OAI-SearchBot, ChatGPT-User, Claude-User, Claude-SearchBot, ClaudeBot, PerplexityBot, Perplexity-User, Google-Extended, Applebot-Extended, Amazonbot, Bytespider, CCBot, Meta-ExternalAgent, cohere-ai) and search crawlers (Googlebot, Bingbot, DuckDuckBot, YandexBot, Baiduspider, Applebot). In the browser it only sees crawlers that run JavaScript and can't tell a real GPTBot from a spoofed one — verify crawlers on the server with their published IP ranges.

Tuning

await client.detect({
  webmcp: { pollMs: 500 },     // wait for WebMCP activity; 0 = check once; false = skip
  crawler: { confidence: 0.9 }, // false = skip
  dom: { threshold: 0.4 },     // false = skip
})

What we deliberately don't do

• No canvas, audio or font fingerprinting, and no reading what visitors type.
• No retries to defeat anti-detection tooling. An agent that hides well browses as a human — and is still welcome.
• Detection never blocks. Use authorization rules to decide what each kind of agent may call.