Roadmap

Roadmap

Shipped (1.1.0)

• Full-duplex orchestrator with an IDLE / LISTEN / THINK / SPEAK state machine
• Stateful VAD with hysteresis and hangover and a trailing-silence flush
• Streaming STT, LLM, and TTS pipeline with token-to-audio handoff on the first token
• Self-hosted default stack: Groq Llama 4, Whisper.cpp, OpenTTS Coqui XTTS v2
• Pluggable adapters per layer (Groq, SarmaLink-AI, OpenAI, Whisper.cpp, Deepgram, OpenAI Whisper, OpenTTS, Cartesia, ElevenLabs)
• Barge-in handling that aborts the in-flight LLM and TTS streams and resets the adapters
• Function-call passthrough with a server-side tool registry
• Conversation history retained across turns, bounded to the recent window
• End-to-end tests with fixtures, plus per-adapter unit tests
• Lint, typecheck, build, and test in CI
• Next.js web client

Planned

• Production VAD (silero-vad-onnx) behind the existing frameRms seam, keeping the hysteresis and hangover layer on top
• Word-level interim results via the Deepgram streaming WebSocket SDK
• Multi-language hot swap mid-call
• Per-adapter latency dashboards
• mediasoup or LiveKit transport adapter at the edge for SFU fan-out
• Native iOS and Android clients

Will not ship

• Voice cloning marketplace (use the TTS provider directly)
• Hosted SaaS layer (this is open-source infrastructure)
• A LiveKit replacement (this works alongside LiveKit, not against it)

Contribute

Pick from Planned, open an issue, fork, branch, push, PR. Keep changes small and conventional.

I will not merge:

• Framework swaps (Fastify and Next.js stay)
• Sync handlers (everything is async and streaming)
• Adapters for paid-only providers with no free-tier or self-hosted path

Releases: see GitHub Releases.