@livekit/agents-plugin-phonic

Realtime voice AI integration for Phonic with LiveKit Agents.

Usage

// SPDX-FileCopyrightText: 2026 LiveKit, Inc.
//
// SPDX-License-Identifier: Apache-2.0
import { type JobContext, ServerOptions, cli, defineAgent, llm, voice } from '@livekit/agents';
import * as phonic from '@livekit/agents-plugin-phonic';
import { fileURLToPath } from 'node:url';
import { z } from 'zod';

const toggleLight = llm.tool({
  description: 'Toggle a light on or off. Available lights are A05, A06, A07, and A08.',
  parameters: z.object({
    light_id: z.string().describe('The ID of the light to toggle'),
    state: z.enum(['on', 'off']).describe('Whether to turn the light on or off'),
  }),
  execute: async ({ light_id, state }) => {
    console.log(`Turning ${state} light ${light_id}`);
    await new Promise((resolve) => setTimeout(resolve, 1_000));
    return `Light ${light_id} turned ${state}`;
  },
});

export default defineAgent({
  entry: async (ctx: JobContext) => {
    const agent = new voice.Agent({
      instructions: 'You are a helpful voice AI assistant named Alex.',
      tools: {
        toggle_light: toggleLight,
      },
    });

    const session = new voice.AgentSession({
      // Uses PHONIC_API_KEY environment variable when apiKey is not provided
      llm: new phonic.realtime.RealtimeModel({
        voice: 'sabrina',
        audioSpeed: 1.2,
      }),
    });

    await session.start({
      agent,
      room: ctx.room,
    });

    await ctx.connect();

    await session.generateReply({
      instructions: 'Greet the user, asking about their day.',
    });
  },
});

cli.runApp(new ServerOptions({ agent: fileURLToPath(import.meta.url) }));
Copy

Configuration

Set the PHONIC_API_KEY environment variable, or pass apiKey directly to RealtimeModel. All other options are optional.

Per-tool configuration

configsForTools takes one entry per tool you want to customize. Each entry is keyed by the tool name; every other field is optional and falls back to the plugin default when omitted. Tools with no entry keep the defaults.

new phonic.realtime.RealtimeModel({
  configsForTools: [
    { name: 'transfer_call', forbid_speech_after_tool_call: true },
    { name: 'submit_form', forbid_tool_call_after_speech: true },
  ],
});
Copy

The plugin always sends tool calls with wait_for_speech_before_tool_call on and allow_tool_chaining off; these are not configurable per tool.

Deprecated: the top-level forbidSpeechAfterToolCall: string[] option still works but is deprecated — it now folds each listed tool into configsForTools as forbid_speech_after_tool_call: true (an explicit configsForTools entry wins) and logs a warning. Prefer configsForTools.

Built-in tools

Phonic's built-in tools — choose_not_to_respond, keypad_input, natural_conversation_ending — are enabled by listing their names in phonicTools, alongside any Webhook tools. To configure one, add a configsForTools entry keyed by the same name (respond_after_sec for choose_not_to_respond; speech_before_tool_call for the other two). A built-in listed without a config uses its Phonic-side defaults.

new phonic.realtime.RealtimeModel({
  phonicTools: ['choose_not_to_respond', 'keypad_input'],
  configsForTools: [{ name: 'choose_not_to_respond', respond_after_sec: 5 }],
});
Copy

If you already have an agent set up on the Phonic platform, you can use the phonicAgent option to specify the agent name. As a note, configuration options you set in the LiveKit Agents SDK will override the agent settings set on the Phonic platform. This means the system prompt you have set on the Phonic platform will be ignored in favor of the instructions field set on the LiveKit voice.Agent. Likewise, options explicitly set in the RealtimeModel constructor will override the Phonic agent's settings.

If you have Webhook tools set up on the Phonic platform, you can use phonicTools to make them available to your agent, together with Phonic's built-in tools. Custom function tools you define on the LiveKit voice.Agent are also supported and run over the websocket.