Voxwire
Open source · WebRTC · Gemini Live

Real-time voice calls with an AI that listens.

A self-hostable stack for talking to AI over WebRTC. Pion handles media in Go, Voxwire server orchestrates, and Gemini Live answers, with natural turn-taking and barge-in.

Try the live demo Self-host in 2 minutes
$ docker compose up -d --build
idle
Ready when you are
Microphone access is required
opus · 48kHz00:00

// features

Everything a voice agent needs, nothing it doesn't.

Small, readable services you can fork and swap out piece by piece.

Low latency

Opus over WebRTC with 20ms frames, paced in real time by the Pion SFU.

Natural turn-taking

Server-side voice activity detection. Interrupt the AI mid-sentence and it stops.

Echo mode

No API key yet? The pipeline echoes your voice so you can test the full media path.

Self-hosted

One docker compose command runs the whole stack on your machine.

gRPC contract

A small protobuf interface between media and AI. Swap either side freely.

Model agnostic

Ships with Gemini Live. The AI client is four small methods, so you can bring your own.

// architecture

How a call flows

Media stays in Go, intelligence stays in Python. They talk over bidirectional gRPC streams.

client

Browser

Captures the mic and plays AI audio over a WebRTC peer connection.

go · :50051

Pion SFU

Decodes Opus, resamples 48k to 16k and back, paces playback.

python · :8000

Voxwire server

WebSocket signaling and the audio pipeline between SFU and model.

model

Gemini Live

Streams speech in and speech out, with built-in barge-in.

// quick start

Running locally in three steps

  1. Clone the repoRequires Docker with Compose v2.
  2. Add your keyLeave it blank to try echo mode first.
  3. Start and callOpen localhost:3000 and press Start call.
$ git clone https://github.com/AnjalLLL/voxwire
$ cd voxwire
$ cp .env.example .env
# set GEMINI_API_KEY=... (or leave empty for echo mode)
$ docker compose up -d --build
# open http://localhost:3000