Give your AI a second sense: on-device human sensing with EDNA

Today, an AI assistant knows exactly what you type — and nothing else. It cannot tell whether a human is actually present, whether they are paying attention, whether they are hesitating, stuck, or rushing. Every adaptation has to be guessed from text alone.

We built EDNA Engine to close that gap: a behavioral-biometric human-sensing layer that runs 100% on-device and turns mouse, touch, typing and motion into a live, machine-readable signal vector for AI systems.

What it measures

EDNA scores 16 behavioral channels in real time: timing distribution of keystrokes and clicks (human intervals are log-normal, bots are uniform), pointer path curvature, speed variation, scroll and reading-pause behavior, device motion (gyroscope proves a physical device in a hand), contact area, click duration, corrections, typing bursts, natural drift, autocorrelation, timing entropy, motion character and event rate.

The output is a live Human Confidence score (0–99.9%) that typically reaches 99%+ after 30–90 seconds of normal usage — presence, attention and hesitation as a first-class signal for your model.

Privacy by architecture, not by policy

Behavioral biometrics usually means shipping intimate data to someone else's cloud. EDNA inverts the model:

• All capture and scoring happens in the user's browser sandbox
• Zero raw biometrics leave the device — no keystroke content, no pointer paths, no motion data
• When proof must be shared, the device emits only non-personalized cryptographic hash patterns — irreversible, unlinkable, revocable
• No server, no API keys, no analytics, no cookies

There is literally nothing to leak, sell or subpoena.

Integration: MCP first

EDNA ships as an MCP server, so any MCP-capable assistant can perceive its user:

{
  "mcpServers": {
    "edna": {
      "command": "npx",
      "args": ["-y", "edna-mcp-server"],
      "env": { "EDNA_LICENSE_KEY": "edna_live_…" }
    }
  }
}

Three tools are exposed to the model:

• edna.get_human_proof() → confidence, classification, consent
• edna.get_signal_vector() → 16 statistical humanness scores
• edna.get_human_context() → tempo trend, per-channel changes, pauses — with explicit uncertainty

Plus a REST API, a JS SDK and an installable PWA for product teams.

Bot classification — and its honest boundary

Signature and cheap-evasion bots are classified against published test vectors. Statistical mimicry — a bot that genuinely reproduces human statistics — is a documented open boundary, with challenge-response planned for v1.1. We would rather document a limit than overclaim.

Try it in 30 seconds

The live console at openos.space scores your own behavior in real time as you move, scroll and type. Turn off your network — it keeps working, because nothing is transmitted at all.

Evaluation and testing are free under the EPCAL-1.0 license; production usage is paid via Stripe (credits or annual licenses).

How would you use a human-context signal in your agents? Presence-aware pacing? Hesitation-triggered assistance? One-human-one-voice for web3? Curious what you'd build.