README.md

LangGraph.js x LiveKit Agents

This is the JavaScript implementation of LangGraph LiveKit Agents, which enables building voice-enabled AI agents using LangGraph and LiveKit.

Initial Setup

1. Install dependencies:

pnpm install

2. Configure environment variables: Create a .env file with your API keys:

LIVEKIT_API_KEY=
LIVEKIT_API_SECRET=
LIVEKIT_URL=
DEEPGRAM_API_KEY=
OPENAI_API_KEY=
GROQ_API_KEY=

3. Run the LangGraph.js dev server and the LiveKit Agents worker:

pnpm run agent
pnpm run voice

Usage

import { LangGraphAdapter } from "./src/runtime.mts";
import { RemoteGraph } from "@langchain/langgraph/remote";

// Can either be a local compiled graph or a deployed graph
const graph = new RemoteGraph({
  graphId: "agent",
  url: "http://localhost:2024",
});

const agent = new pipeline.VoicePipelineAgent(
  vad,
  openai.STT.withGroq({ model: "whisper-large-v3-turbo" }),
  new LangGraphAdapter(graph),
  new openai.TTS()
);

Explanation

The LangGraph LiveKit adapter is a wrapper around llm.LLM, that maps the LangGraph messages stream mode to LiveKit's voice chunks.

The sample provides several key features:

1. Interrupts: Implement human-in-the-loop-style interrupts using LangGraph's custom stream mode. This allows the agent to pause and wait for user input at specific points in the conversation.

2. Manual Voice Speakout: Use the say() method to play static messages or announcements at any point in the conversation.

Here's an example of using these features:

import { typedLiveKit } from "./src/types.mjs";

const liveKit = typedLiveKit(config);

// Using interrupts to get user input
const { content, messages } = liveKit.interrupt("What is your name?");

// Manual voice speakout
liveKit.say("Give me a second to think...");

LangGraphAdapter supports both graphs deployed in LangGraph Platform and standalone graphs running within LiveKit Agents worker, just swap the RemoteGraph with your compiled graph.

import { LangGraphAdapter } from "./src/runtime.mts";
import { StateGraph } from "@langchain/langgraph";

const graph = new StateGraph(...).compile();

const agent = new pipeline.VoicePipelineAgent(
  vad,
  openai.STT.withGroq({ model: "whisper-large-v3-turbo" }),
  new LangGraphAdapter(graph),
  new openai.TTS()
);