@mosadd/voice-truthgate-mcp

Voice Truthgate — MCP server

Give any AI agent an honest "is this really my contact, on this call?" check. This MCP server wraps the Voice Truthgate market API as three tools:

tool	what it does
voice_truthgate_enroll	enrol a subject's voiceprint from 1+ clean clips (consent required)
voice_truthgate_verify	verify a clip against a subject → honest verdict + confidence + clone caution
voice_truthgate_list_subjects	list the subjects enrolled under your key

The engine key and the biometric never touch this process — it only forwards your customer X-API-Key.

Configure (Claude Desktop / any MCP client)

{
  "mcpServers": {
    "voice-truthgate": {
      "command": "npx",
      "args": ["-y", "@mosadd/voice-truthgate-mcp"],
      "env": { "VTG_API_KEY": "vtg_live_your_key_here" }
    }
  }
}

Optional VTG_API_URL overrides the endpoint (defaults to the hosted API). Needs Node ≥ 20.

Example agent turn

"Verify the caller against subject acme-ceo — the clip is at /tmp/call.wav."

The agent calls voice_truthgate_verify({ subject_id: "acme-ceo", audio_path: "/tmp/call.wav" }) and gets back e.g.:

{ "verdict": "likely_same_person", "score": 0.81, "confidence_band": "high",
  "synthetic_caution": true, "synthetic_note": "A cloned/synthetic voice can also produce this match ..." }

Honesty rail (read this)

A match is a supporting signal, not proof of a live human — in our own testing a targeted clone of an enrolled person was accepted ~63% of the time. For high-stakes actions (payments, access), combine this with an identity check and a liveness/challenge step. Voiceprints are biometric data; enrol only with the subject's consent and in line with applicable law.