schema

Imported from jambonz/schema (AGENTS.md). Install upstream with npx skills add jambonz/schema. Copyright stays with the author.

jambonz Developer Guide

This guide covers the jambonz verb model, transport modes, and protocol. For SDK-specific documentation, see the AGENTS.md in the respective SDK repository.

jambonz is an open-source CPaaS (Communications Platform as a Service) for building voice and messaging applications. It handles telephony infrastructure — SIP, carriers, phone numbers, media processing — so you can focus on application logic.

Server Versions

jambonz has two editions: v0.9.x (open source) and v10.x (commercial). Always target the commercial version (v10.x). All verb schemas and features are available.

How jambonz Applications Work

A jambonz application controls phone calls by returning arrays of verbs — JSON instructions that execute sequentially. The runtime processes each verb in order: speak text, play audio, collect input, dial a number, connect to an AI model, etc.

The Webhook Lifecycle

1. An incoming call arrives. jambonz invokes your application's URL with call details (caller, called number, SIP headers, etc.).
2. Your application returns a JSON array of verbs.
3. jambonz executes the verbs in order.
4. When a verb with an actionHook completes (e.g. gather collects speech input), jambonz invokes the actionHook URL with the result.
5. The actionHook response (a new verb array) replaces the remaining verb stack.
6. This continues until the call ends or a hangup verb is executed.

Transport Modes

jambonz supports two transport modes for delivering verb arrays:

• Webhook (HTTP): Your server receives HTTP POST requests with call data and returns JSON verb arrays in the response body. Stateless and simple. Good for IVR menus, call routing, and straightforward flows.
• WebSocket: Your server maintains a persistent websocket connection with jambonz. Verb arrays are sent as JSON messages in both directions. Required for real-time features like LLM conversations, audio streaming, and event-driven flows.

The verb schemas and JSON structure are identical in both modes. The difference is the transport.

When to Use Which

• Webhook: Simple IVR, call routing, voicemail, basic gather-and-respond patterns.
• WebSocket: LLM-powered voice agents, real-time audio streaming, complex conversational flows, anything requiring bidirectional communication, or asynchronous logic, or streaming tts.

IMPORTANT: Any application that uses a speech-to-speech verb (openai_s2s, gptlive_s2s, voicelive_s2s, nova_s2s, google_s2s, deepgram_s2s, ultravox_s2s, qwen_s2s, elevenlabs_s2s, s2s, or agent) MUST use WebSocket transport, not webhooks. These verbs require persistent bidirectional communication for real-time audio and events.

Schema

The complete verb schema is at jambonz-app.schema.json in the package root. This is a JSON Schema (draft 2020-12) that defines the structure of a jambonz application.

Individual verb schemas are in verbs/. Shared component types (synthesizer, recognizer, target, etc.) are in components/. Callback payload schemas are in callbacks/.

MCP Server

AI agents can fetch individual schemas on demand via the jambonz MCP server:

• Remote: https://mcp-server.jambonz.app/mcp (no install needed)
• Local: npx @jambonz/mcp-schema-server (stdio) or npx @jambonz/mcp-schema-server --http (HTTP)

Two tools are available:

1. jambonz_developer_toolkit — Returns this guide plus an index of all available schemas. Call this first.
2. get_jambonz_schema — Fetch the JSON Schema for any verb, component, or callback (e.g. verb:say, component:synthesizer, callback:gather, guide:session-commands).

Core Verbs

Audio & Speech

• say — Speak text using TTS. Supports SSML, streaming, multiple voices.
• play — Play an audio file from a URL.
• gather — Collect speech (STT) and/or DTMF input. The workhorse for interactive menus and voice input.

AI & Real-time

• openai_s2s / google_s2s / deepgram_s2s / ultravox_s2s / qwen_s2s — Connect the caller to a vendor-specific LLM for real-time voice conversation. These are the preferred verbs when the vendor is known. Each handles the full STT→LLM→TTS flow with the vendor pre-set.
• gptlive_s2s — Connect the caller to OpenAI's GPT Live API (limited-access alpha). A different wire protocol from openai_s2s (Realtime): llmOptions carries only a session_update, there is no response_create, and tool calls arrive via a Responses-targeted delegation. See GPT Live S2S specifics below.
• voicelive_s2s — Connect the caller to the Azure Voice Live API. Requires connectOptions.host (your Foundry/Speech resource). Unlike openai_s2s, the session_update uses Voice Live's flat shape and is sent verbatim, so Azure-only settings (azure_semantic_vad, azure_deep_noise_suppression, Azure TTS voices) work.
• nova_s2s — Connect the caller to Amazon Nova Sonic on Amazon Bedrock. auth takes AWS IAM credentials (accessKeyId, secretAccessKey, optional sessionToken) plus region; model defaults to amazon.nova-2-5-sonic. llmOptions (systemPrompt, history, greeting, sessionStart, promptStart) is optional. Nova waits for the caller unless a greeting is set. Requires the mediajam media server.
• elevenlabs_s2s — Connect the caller to an ElevenLabs Conversational AI agent. Unlike other s2s vendors, ElevenLabs requires a pre-configured agent_id (created in the ElevenLabs dashboard) rather than a model and messages. See ElevenLabs S2S specifics below.
• s2s — Generic LLM voice conversation verb. Use only when the vendor is determined at runtime (e.g. from an env var). Requires vendor to be specified.
• agent — Higher-level voice AI agent with integrated turn detection. Mix-and-match STT, LLM, and TTS vendors.
• dialogflow — Connect the caller to a Google Dialogflow agent (ES, CX, or CES).
• stream — Stream raw audio to a websocket endpoint for custom processing.
• transcribe — Real-time call transcription sent to a webhook.

Call Control

• dial — Place an outbound call and bridge it to the current caller.
• conference — Multi-party conference room.
• enqueue / dequeue — Call queuing.
• hangup — End the call.
• redirect — Transfer control to a different webhook.
• pause — Wait for a specified duration.

SIP

• sip:decline — Reject an incoming call with a SIP error.
• sip:request — Send a SIP request within the dialog.
• sip:refer — Transfer the call via SIP REFER.

Utility

• config — Set session-level defaults (TTS vendor/voice, STT vendor, VAD, etc.).
• tag — Attach metadata to the call.
• dtmf — Send DTMF tones.
• dub — Mix auxiliary audio tracks into the call.
• message — Send SMS/MMS.
• alert — Send a SIP 180 with Alert-Info.
• answer — Explicitly answer the call.
• leave — Leave a conference or queue.

Verb Synonyms and Shortcuts

IMPORTANT — Code generation rules:

1. Always use stream, never listen — they are synonyms; stream is the preferred name.
2. Always use the vendor-specific shortcut when the LLM vendor is known — use openai_s2s, gptlive_s2s, voicelive_s2s, nova_s2s, google_s2s, elevenlabs_s2s, deepgram_s2s, qwen_s2s, or ultravox_s2s. Do NOT use llm or s2s with a vendor property when a shortcut exists.
3. Use s2s (not llm) when the vendor is dynamic — e.g. the vendor comes from an env var or runtime config. Both s2s and llm are synonyms, but prefer s2s.
4. Never use llm in generated code — it is a legacy name. Use either a vendor shortcut or s2s.

Common Patterns (Raw JSON)

These are the raw JSON verb arrays that jambonz applications produce. They show the underlying structure for reference.

Simple Greeting and Gather

[
  { "verb": "say", "text": "Welcome. Press 1 for sales, 2 for support." },
  { "verb": "gather", "input": ["digits"], "numDigits": 1, "actionHook": "/menu" }
]

LLM Voice Agent

[
  {
    "verb": "config",
    "synthesizer": { "vendor": "elevenlabs", "voice": "EXAVITQu4vr4xnSDxMaL" },
    "recognizer": { "vendor": "deepgram", "language": "en-US" }
  },
  {
    "verb": "openai_s2s",
    "model": "gpt-4o",
    "llmOptions": {
      "messages": [{ "role": "system", "content": "You are a helpful assistant." }]
    },
    "actionHook": "/llm-done",
    "toolHook": "/tool-call"
  }
]

ElevenLabs S2S Specifics

ElevenLabs works differently from other s2s vendors. Instead of passing a model and system prompt, you create a Conversational AI agent in the ElevenLabs dashboard and pass the agent_id to jambonz. The agent's voice, personality, tools, and LLM configuration are all managed on the ElevenLabs side.

Key differences from other s2s verbs:

• auth requires agent_id (required) and optionally api_key (enables signed WebSocket URLs)
• model is NOT used — the model is configured in the ElevenLabs agent
• llmOptions should be an empty object {} (do NOT pass messages or temperature)
• llmOptions.conversation_initiation_client_data can optionally send data to the agent at conversation start
• Always include eventHook and events: ['all'] — omitting eventHook causes errors on the server

[
  {
    "verb": "elevenlabs_s2s",
    "auth": {
      "agent_id": "your-agent-id",
      "api_key": "your-api-key"
    },
    "llmOptions": {},
    "actionHook": "/s2s-complete",
    "eventHook": "/event",
    "events": ["all"]
  }
]

GPT Live S2S Specifics

OpenAI's GPT Live API (limited-access alpha) is a different wire protocol from the OpenAI Realtime API behind openai_s2s, so it has its own verb and its own llmOptions shape.

Key differences from openai_s2s:

• llmOptions takes only session_update — there is no response_create. The model starts and drives the conversation itself.
• Do NOT put model inside session_update; the model travels in the connection URL (set the verb's model property instead, default gpt-live-1-boulder-alpha).
• Work the model wants done arrives as a delegation, selected by session_update.delegation.type:
  • client — the model asks your app for free-form text context. Answer with an llm:update carrying delegation.context.append (max 500 tokens, exactly one input_text part).
  • responses — the model runs a Responses API turn whose response.* events stream inline. Function calls are answered with delegation.function_call_output.create.
• A responses delegation carries a required nested object whose model is mandatory, and tools go INSIDE it:

  "delegation": {
    "type": "responses",
    "responses": {
      "model": "gpt-5.5",
      "tools": [{ "type": "function", "name": "get_weather", "description": "...", "parameters": {} }]
    }
  }

  Tools placed at delegation.tools are rejected by the server. delegation.responses.model is the Responses-side model, separate from the GPT Live voice model in the verb's model property.
• Tools — mcpServers, handoff and hangup — only work with delegation.type: "responses" and delegation.responses.model set; with a client delegation there is no function-calling protocol to declare them on, and the verb is rejected.
• Mid-call llm:update accepts only session.update, session.context.append, delegation.context.append, delegation.function_call_output.create and session.close. Realtime events (response.create, response.cancel, input_audio_buffer.*, conversation.item.create) are rejected.
• A later session.update is sparse (omitted fields keep their values), but replacing delegation requires the complete delegation object.

[
  {
    "verb": "gptlive_s2s",
    "model": "gpt-live-1-boulder-alpha",
    "auth": { "apiKey": "sk-..." },
    "llmOptions": {
      "session_update": {
        "instructions": "You are a helpful voice assistant. Be concise.",
        "audio": { "output": { "voice": "marin" } },
        "delegation": { "type": "client" }
      }
    },
    "actionHook": "/s2s-complete",
    "eventHook": "/event",
    "events": ["all"]
  }
]

Dial with Fallback

[
  { "verb": "say", "text": "Connecting you now." },
  {
    "verb": "dial",
    "target": [{ "type": "phone", "number": "+15085551212" }],
    "answerOnBridge": true,
    "timeout": 30,
    "actionHook": "/dial-result"
  },
  { "verb": "say", "text": "The agent is unavailable. Goodbye." },
  { "verb": "hangup" }
]

Call Queue

[
  { "verb": "say", "text": "All agents are busy. You are in the queue." },
  {
    "verb": "enqueue",
    "name": "support",
    "waitHook": "/hold-music",
    "actionHook": "/queue-exit"
  }
]

ActionHook Payloads

When a verb completes, jambonz invokes the actionHook URL (webhook) or sends an event (WebSocket) with result data. Every actionHook payload includes these base fields:

Verb-Specific Payload Fields

gather: speech (object with alternatives[].transcript), digits (string), reason (speechDetected, dtmfDetected, timeout)

dial: dial_call_sid, dial_call_status, dial_sip_status, dial_sbc_callid, duration

llm: completion_reason (normal, timeout, error), llm_usage (token counts)

enqueue: queue_result (dequeued, hangup, error)

transcribe: transcription (object with transcript text)

WebSocket Protocol

Message Types (jambonz → app)

Message Types (app → jambonz)

REST API

jambonz provides a REST API for platform management and active call control.

Key resources:

• Calls — Create outbound calls, query active calls, modify in-progress calls (redirect, whisper, mute, hangup)
• Messages — Send SMS/MMS messages

Key Concepts

• Verb: A JSON object with a verb property that tells jambonz what to do. Verbs execute sequentially.
• ActionHook: A webhook URL that jambonz calls when a verb completes. Returns the next verb array. Payload includes call details and verb-specific results.
• Synthesizer: TTS configuration (vendor, voice, language).
• Recognizer: STT configuration (vendor, language, model).
• Target: A call destination (phone number, SIP URI, registered user, Teams user).
• Session: A single phone call. Session-level settings (set via config) persist across verbs.
• Inject Command: Asynchronous mid-call modification (WebSocket). Executes immediately without replacing the verb stack.