Voicebot

> Language. Reply in the user's language: detect what they write — Hebrew→Hebrew, English→English — and mirror it, switching if they switch mid-conversation. This shapes your prose, your questions, and your AskUserQuestion option labels only. It does not change the artifacts you produce — identifiers, JSON keys, BCP-47 language codes, API field names, and other data stay exactly as specified.

Help the developer implement a VoiceBot API endpoint — a URL that the Voicenter Voice Agent (AI bot) calls during a live conversation to retrieve dynamic CRM data before continuing.

When to use this skill

Use this skill when the user wants to:

• Feed real-time CRM data to an AI voice agent mid-conversation (account balance, order status, open tickets)
• Let a voice bot look up customer information by the caller ID or DTMF input collected during the call
• Return structured data that the voice agent will read back to the caller
• Route the voice agent's behavior dynamically based on the caller's CRM profile
• Return appointment slots, product info, or personalized offers from an external system
• Handle multiple intents from a single endpoint using LAYER_ID to route logic

Environment Variables

# No outbound API token needed — Voicenter sends requests TO your server.
# Configure your endpoint URL in: Voice Agent intent settings → "URL" field under "תגובה" (Response)

How it works

1. An inbound or outbound call is handled by a Voicenter Voice Agent (AI bot).
2. The agent reaches an intent that needs external data.
3. The Voice Agent POSTs call info + collected intent parameters to your endpoint.
4. Your server fetches the data from your CRM/DB and responds.
5. The Voice Agent reads the response and continues the conversation.

Note: Pair with the External Layer API for the full pattern — External Layer passes CRM context (caller ID, tier, ticket) into CUSTOM_DATA at call start, and VoiceBot fetches dynamic data mid-conversation.

Request Voicenter sends you

Method: POST Content-Type: application/json

{
  "CALL_INFO": {
    "DID": "0722776772",
    "CALLER_ID": "0501234567",
    "IVR_UNIQUE_ID": "ssss1bcd7954224861f85a2d70612f2",
    "DTMF": "1234",
    "LAYER_ID": "10",
    "PREVIOUS_LAYER_ID": "5"
  },
  "IntentParameters": {
    "city": "Tel Aviv",
    "product_interest": "Pro Plan"
  },
  "CUSTOM_DATA": {
    "CRM_client_ID": "12345",
    "Last_ticket_ID": "222",
    "Last_representative": "John Doe"
  }
}

Request Fields

| Field | Type | Description | |---|---|---| | CALL_INFO.DID | String | The phone number the caller dialed | | CALL_INFO.CALLER_ID | String | Caller's phone number — use to look them up in CRM | | CALL_INFO.IVR_UNIQUE_ID | String | Unique call ID — correlates with CDR and other APIs | | CALL_INFO.DTMF | String | Digits the caller pressed. Default "0" if none. | | CALL_INFO.LAYER_ID | String | IVR layer ID — use to route different intents when multiple share the same endpoint | | CALL_INFO.PREVIOUS_LAYER_ID | String | Previous IVR layer ID | | IntentParameters | Object | Fields collected by the Voice Agent in this conversation so far | | CUSTOM_DATA | Object | Flat key-value data passed from External Layer API. No nested objects. |

Your Response

All fields are optional. Return {} to let the agent continue with existing instructions.

Response Fields

| Field | Type | Description | |---|---|---| | function_output | Object | Structured data for the agent to use. Any valid JSON — the agent reads field names and values. | | user | String[] | Instructions for the agent. Most commonly used field. Supports Markdown. | | assistant | Object | Additional influence on agent behavior. Rarely needed. | | system | Object | System-level instructions merged into the agent's base prompt. Supports Markdown. |

Simple response — instruction only

{
  "user": [
    "Tell the caller their order #123456 shipped this morning. Delivery is expected today between 2 PM and 4 PM. Ask if there is anything else they need."
  ]
}

Response with structured data

{
  "function_output": {
    "account_balance": 180,
    "plan": "Pro",
    "renewal_date": "2024-08-01",
    "open_tickets": 2
  },
  "user": [
    "Read the caller their account balance and plan. Let them know they have 2 open support tickets and offer to transfer them to support."
  ]
}

Response with a list (e.g., open orders)

{
  "function_output": {
    "orders": [
      { "id": "ORD-001", "status": "Shipped", "eta": "Today" },
      { "id": "ORD-002", "status": "Processing", "eta": "2-3 days" }
    ]
  },
  "user": [
    "Read the caller their open orders. For each order, state the order ID and current status."
  ]
}

TypeScript Implementation (Express)

import express, { Request, Response } from 'express';

const app = express();
app.use(express.json());

interface VoiceBotRequest {
  CALL_INFO: {
    DID: string;
    CALLER_ID: string;
    IVR_UNIQUE_ID: string;
    DTMF: string;
    LAYER_ID: string;
    PREVIOUS_LAYER_ID: string;
  };
  IntentParameters: Record;
  CUSTOM_DATA: Record;
}

interface VoiceBotResponse {
  function_output?: Record;
  user?: string[];
  assistant?: Record;
  system?: Record;
}

app.post('/webhooks/voicenter/voicebot', async (req: Request, res: Response) => {
  const { CALL_INFO, IntentParameters, CUSTOM_DATA } = req.body as VoiceBotRequest;

  try {
    const crmClientId = CUSTOM_DATA.CRM_client_ID;
    const layerId = CALL_INFO.LAYER_ID;

    // Route logic by intent/layer ID
    if (layerId === '10') {
      // Order status intent
      const order = await crm.getLatestOrder(crmClientId);
      const response: VoiceBotResponse = {
        function_output: {
          order_id: order.id,
          status: order.status,
          eta: order.estimatedDelivery,
        },
        user: [
          `Tell the caller their latest order (${order.id}) is currently ${order.status}. ` +
          `Estimated delivery: ${order.estimatedDelivery}. Ask if they need anything else.`,
        ],
      };
      return res.json(response);
    }

    if (layerId === '11') {
      // Account balance intent
      const account = await crm.getAccount(crmClientId);
      return res.json({
        function_output: { balance: account.balance, currency: 'ILS', plan: account.plan },
        user: ['Read the caller their account balance and current plan.'],
      });
    }

    if (layerId === '12') {
      // Appointment scheduling intent
      const city = IntentParameters.city ?? '';
      const slots = await crm.getAvailableSlots(city);
      return res.json({
        function_output: { available_slots: slots },
        user: [
          'Read the caller the available appointment slots in their city. ' +
          'Ask which time works best and confirm the booking.',
        ],
      });
    }

    // Default — agent continues with existing instructions
    return res.json({});

  } catch (err) {
    console.error('VoiceBot endpoint error:', err);
    // Always return a valid response — never let it hang or return 5xx
    return res.json({});
  }
});

app.listen(3000);

Tips

• Always respond — even on error, return {} so the Voice Agent continues gracefully. Never return a 5xx or let the request hang.
• LAYER_ID is your routing key when multiple intents call the same endpoint — use it to run different CRM queries per intent.
• CUSTOM_DATA is flat — only supports simple key-value pairs (no nested objects). Use function_output in your response to return complex structured data.
• user field supports Markdown — use headers, bold, tables, and lists when guiding the agent through structured data like order lists or slot tables.
• IntentParameters contains what the agent has already collected — use it to look up records (e.g., an order number or ID the caller said aloud).
• Works for both inbound and outbound calls.

Related Skills

• External Layer — Run at call start to pass CRM context into CUSTOM_DATA before the VoiceBot is called
• CDR Notification — Receives the full call record after the voice agent conversation ends
• Click2Call — Use IVR_UNIQUE_ID to correlate voicebot calls with outbound Click2Call initiations

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: VoicenterTeam
• Source: VoicenterTeam/claude-marketplace
• License: MIT

Install and usage instructions live in the source repository linked above.