opensam

The combined open-source POC (Proof of Concept) of a pure-mesh WebRTC stack, a mediasoup SFU, and an AI-native voice/video bot — three progressively richer building blocks for next-gen real-time voice/video applications.

The source code lives on the development branch. main is the project landing page; check out development to run any of the projects.

Projects

1. full-mesh — Peer-to-peer WebRTC calling

A real-time multi-user video calling app built on a pure WebRTC mesh, where every participant connects directly to every other participant (no media server in the path). Node.js + WebSocket signalling, React UI, and a coturn TURN relay.

• Video/audio calls for up to 4 users, mute/unmute, camera toggle, screen sharing
• In-call chat, full-call recording (.webm), connection-quality indicator
• JWT-authenticated room access
• Simplest architecture; best for small rooms.

See full-mesh/README.md.

2. mediasoup — SFU-based conferencing

The same experience re-architected on a mediasoup SFU: clients upload their media once and the server forwards it to everyone else, cutting client CPU and bandwidth and scaling far better than a mesh.

• Multi-user conferencing, mute/camera/screen-share, DataChannel chat, client-side recording
• JWT auth, per-peer signalling rate limiting, graceful shutdown
• Redis persistence, Prometheus + Grafana monitoring
• Best for larger rooms.

See mediasoup/README.md.

3. v2ai — AI-native meeting bot

An AI participant that joins a live mediasoup conference, listens, thinks, and speaks in real time. It pulls RTP audio (and VP8 video) off a mediasoup plain transport and runs it through STT → LLM → TTS and vision pipelines.

• Wake-word detection and Silero Voice Activity Detection (VAD)
• Speech-to-Text: Deepgram (primary), Whisper (fallback)
• Context-aware conversation with Anthropic Claude, with PII scrubbing before the LLM
• Meeting summaries, action-item extraction, decision tracking, and meeting Q&A
• Consent-gated screen-share analysis and slide summarization via Claude vision (LLaVA local fallback)
• Multi-provider Text-to-Speech: Cartesia / Deepgram / Kokoro
• OpenTelemetry tracing, Docker-based deployment

See v2ai/README.md.

Getting started

The projects live on the development branch:

git clone git@github.com:Samvyo/opensam.git
cd opensam
git checkout development

Then open the folder for the project you want to run and follow its README:

Project	Folder	Best for
Full-mesh WebRTC	full-mesh	Small peer-to-peer rooms
mediasoup SFU	mediasoup	Larger, scalable conferences
AI meeting bot	v2ai	Adding an AI participant to a conference