Who's at the controls?
Understand your human and automated traffic.
Evidence-based segments for metrics and experiments.
Reading this page's automation signals…
Use the page naturally. Several clicks/taps and scrolls over time can support “likely human.” Keyboard-only and touch-only activity can also qualify with more evidence.
Checking WebMCP support…
Ask your browser agent to reverify this session using the page’s WebMCP tool. The resulting “AGENT TOOL USED” label reports recent tool execution, not passive agent detection. An agent using only clicks or screenshots may remain unclassified. After 30 seconds of no tool use, the monitor automatically rechecks and expires this evidence.
This page uses the monitor API. State changes appear here automatically, including evidence expiry. Background browser throttling can delay checks.
No state changes yet.
Trusted events can come from automation, too. The classifier uses a bounded 30-second activity window; these totals count raw input. Confidence describes evidence quality, not a probability.
Watch the input, not just the flag.
Use these ordinary buttons in any order, or browse the rest of the page. Mouse analysis combines repeated sparse jumps or uniform straight paths with short presses or exact-center targeting. Speed alone never decides. Touch, keyboard, and pen are excluded from these mouse rules.
UNCLASSIFIED · Experimental inference enabled on this demo. Thresholds are not calibrated against real human traffic.
Optional labels are saved only in your downloaded results. They never feed the classifier.
No target selected. This is an input measurement sandbox, not an access challenge.
Live pointer measurements
Recent coordinates are processed locally in a bounded buffer and never exported. The download includes aggregate counts only. Assistive tools, remote desktops, timer precision, and device differences can resemble automation. Read the standards research and validation plan ↗
Show your working.
No mystery score. No fingerprinting.
Just the signals, and what they mean.
Inspect the live detector result
New scroll. Fresh check.
Reaching this checkpoint reruns detection. Scroll back above it and return to check again.
Scroll this panel into view to reverify.
Change the evidence.
See what holds up.
These are simulated inputs to the real library.
Your live result above stays untouched.
UNKNOWN
Attach evidence. Keep the variant.
Event properties for your own analytics client.
This demo sends nothing to a server.
Current analytics properties
Choose a sample variant, then record an exposure. Later conversions keep that variant and preserve the exposure segment alongside the current segment. This is a local example, not a randomization service.
Behavior can be affected by the treatment itself. Use changing segments for exploratory analysis; keep the overall randomized result and exposure-time segment for comparison.
A little history. No tracking.
The last 12 live checks, kept in memory until you leave this page. The download includes these checks and aggregate event counts.
| CHECK | TRIGGER | RESULT | EVIDENCE | LOCAL TIME |
|---|
Useful segments, honest limits. Likely human and behavioral automation labels are unvalidated heuristics, not verified identities or accuracy scores. Passive visitors stay unclassified. Agents can imitate human input, and people can produce automation-like patterns. Validate against labeled traffic before excluding segments from metrics. A declared agent is self-reported. Keep A/B assignment stable, preserve the segment at exposure, and report the unclassified share.