jambonz-agent

A jambonz voice AI agent using the agent verb with multiple STT backends for comparing turn detection strategies.

Callflows

Path	STT Vendor	Turn Detection	Notes
/	Deepgram nova-3-general	Krisp	Multilingual, acoustic end-of-turn via Krisp
/flux	Deepgram Flux	Native (STT)	Acoustic + semantic turn detection built into Deepgram Flux
/aai	AssemblyAI universal-3 pro	Native (STT)	v3 streaming API with native turn detection, language detection enabled

All callflows share the same TTS (Cartesia), LLM (OpenAI), barge-in settings, and system prompt.

Application Environment Variables

These are configurable from the jambonz portal:

Variable	Default	Description
OPENAI_MODEL	gpt-4.1-mini	OpenAI model for the LLM
CARTESIA_VOICE	9626c31c-bec5-4cca-baa8-f8ba9e84c8bc	Cartesia voice ID
SYSTEM_PROMPT	(see source)	System prompt for the voice agent

Setup

npm install
npm run build

Running

# development
npm run dev

# production
PORT=3105 npm start

The PORT environment variable controls the listening port (default: 3000).

pm2

Add to your existing ecosystem.config.js:

{
  name: 'jambonz-agent',
  script: 'dist/app.js',
  cwd: '/home/jambonz/apps/jambonz-agent',
  instances: 1,
  autorestart: true,
  watch: false,
  max_memory_restart: '512M',
  env: {
    NODE_ENV: 'production',
    PORT: 3105
  }
}

jambonz Configuration

Create a jambonz application with the WebSocket URL pointing to the desired callflow path:

• ws://<host>:3105/ for Deepgram + Krisp
• ws://<host>:3105/flux for Deepgram Flux
• ws://<host>:3105/aai for AssemblyAI

Requires jambonz server version 10.1.1+ (the agent verb is commercial-only).