langgraph-voice-call-agent

# LangGraph Voice Call Agent

A real-time voice/call AI agent that lets you talk to a LangGraph agent over LiveKit, similar to "voice mode" experiences in Gemini Voice, OpenAI Realtime API sessions, and Gemini Live. This repo demonstrates adapting any LangGraph agent into a full-duplex, low-latency voice assistant using LiveKit's real-time communication infrastructure.

This backend provides the core voice processing and AI agent functionality, built to work seamlessly with LiveKit's real-time infra and any frontend that supports LiveKit client connections.

## Features

- **Real-time voice interaction** with LangGraph agents
- **Full-duplex communication** with low-latency audio processing
- **Flexible LangGraph integration** - works with any LangGraph agent
- **Comprehensive audio pipeline** including VAD, STT, TTS, and turn detection
- **Thread-based conversation continuity** via participant metadata

## Project Structure

```
langgraph-voice-call-agent/
├── src/                         # Main source code
│   ├── livekit/                 # LiveKit agent implementation
│   │   ├── agent.py             # Main agent entrypoint
│   │   └── adapter/             # LangGraph integration
│   │       └── langgraph.py     # LangGraph adapter for LiveKit
│   └── langgraph/               # LangGraph Agent Sdefinitions
│       └── agent.py             # An example agent
├── compose.yml                  # Docker Compose for local LiveKit server
├── pyproject.toml               # Python project configuration
├── uv.lock                      # uv dependency lock file
└── Makefile                     # Development commands
```

## How it works (high level)

1. **Agent Initialization** → LiveKit agent connects to room and waits for participants
2. **Audio Pipeline Setup** → VAD, STT, TTS, and turn detection models are loaded and configured  
3. **LangGraph Integration** → Connect to LangGraph server
4. **Voice Processing** → Real-time audio is processed through the pipeline:
   -