Custom Tools

A Custom Tool is a raw API definition. Unlike FlowApps (which are pre-built plugins), Custom Tools allow you to connect to any HTTP endpoint, giving you total control over the request, response parsing, and latency handling.

The Execution Flow

Creating a Tool

Navigate to Business Dashboard -> Tools -> Add Tool.

General Information

Name: A unique identifier (e.g., check_inventory).
Description: Crucial. This tells the AI when to use this tool.
- Bad: "API V2 Endpoint."
- Good: "Use this tool when the user asks about the stock availability of a specific product item."

Input Schema

Define the Variables the agent must collect from the user to run this tool.

Key: The variable name used in request templating (e.g., product_name).
Type: String, Number, Boolean, or Date.
Description: Instructions for the AI. (e.g., "The item name. Infer from context.").
Required: If checked, the Agent will ask follow-up questions to get this data if it's missing.

Request Configuration

Define the HTTP Request. Use Scriban Templating to inject the Schema variables.

Method: GET, POST, PUT, DELETE, PATCH.
URL: https://api.mystore.com/products/{{ product_name }}.
Headers: Add Authentication (e.g., Authorization: Bearer sk_...).

Body: Construct your JSON payload.

{
  "item": "{{ product_name }}",
  "quantity": 1
}

Response Handling (Javascript Sandbox)

APIs return raw JSON. You must parse this into something human-readable.

Available Variables:

response.status: HTTP Code (e.g., 200).
response.data: The JSON body.

Code Example:

// API returns: { "stock": 50, "status": "ok" }

if (response.status === 200) {
  // Return a clean string/value
  return response.data.stock;
} else {
  return "Error";
}

Latency Optimization

Static Response (Skip the Brain)

By default, the tool result is sent back to the LLM, which "thinks" about it and generates a natural response. This adds latency (Time-to-First-Token).

You can bypass the LLM and send the result directly to the Text-to-Speech (TTS) engine.

Enable Static Response: Toggle ON.
Template: Define exactly what should be spoken. Use {{ result }} to inject the output of your JS Sandbox.
- Example: We currently have {{ result }} units in stock.
Benefit: Saves ~500ms-1s of latency. The agent speaks immediately after the API returns.

Audio Triggers

Tools take time to run. To prevent "Dead Air" silence while the HTTP request is pending, you can play a sound.

Sound File: Upload an MP3 (e.g., keyboard typing sounds, or a voice saying "Let me check that for you...").
Behavior: The audio loops until the HTTP request completes.

Variable Mapping (Logic)

In the Script Builder, when you click the Custom Tool node, you can map the result to a Script Variable.

Output Variable: Select a variable (e.g., inventory_count).
Usage: You can now use a Condition Node to check if inventory_count < 5 -> Transfer to Agent.

Hybrid Power

This mapping allows you to use the AI for Intent (triggering the tool), but use strict Logic (Condition Nodes) to handle the Result.

Roadmap

Tool Debugger

Coming Soon. A built-in testing interface (top-right of the dashboard) to mock tool executions, inspect JSON payloads, and verify JS parsing logic without placing a real phone call.

Tool Debugger

On this page