WebMCP, crawler, DOM-driver, and screenshot agents have different detection guarantees.

Agent Classes

Agentronics models agents by how they interact with the page. Each class comes with a different detection guarantee — knowing which class you're dealing with is what lets you set a sensible trust floor.

Class	Interaction path	Detection confidence
`webmcp`	Structured tools through `document.modelContext`	Exact, once the agent activates or calls a tool
`crawler`	Fetches/renders pages; identifies itself in `navigator.userAgent`	Signature match — high but spoofable
`dom`	Browser automation over DOM/click/type APIs	Beta, heuristic confidence score
`screenshot`	Vision models driving the mouse and keyboard from screenshots	Behaviour score (teleport / centre clicks, even typing)

const identity = await client.detect()
 
if (identity?.class === 'webmcp') {
  // Structured tool client.
}
 
if (identity?.class === 'crawler') {
  // Known AI/search bot — identity.vendor is the bot name (e.g. 'GPTBot').
}

Crawlers (GPTBot, ClaudeBot, PerplexityBot, Googlebot, …) are matched by their User-Agent token. Because a UA is self-reported, the match is high-confidence but not proof (confidence: 0.9), and client-side detection only catches crawlers that execute JavaScript — see Detection for the full signature list and caveats.

Screenshot (vision) agents are spotted by how they interact — see Detection. To be trusted beyond detection they should authenticate, typically with Web Bot Auth.

Agent Classes

Agent Classes

On this page