Roadmap

Kith's development is planned in three phases. v0.1 is shipped and published on npm.

v0.1 — Shipped

The foundation. Everything needed to ship a voice companion with natural-sounding TTS.

@kithjs/core — 5 adapter contracts, 12-event normalized bus, policy hooks
@kithjs/runtime-pipecat — Python Pipecat sidecar with auto-respawn, barge-in, full event emission
@kithjs/runtime-livekit — Mock-mode portability stub (proves adapter contract works across runtimes)
@kithjs/voice-router — Sentence-aware chunking, 4 default slang dicts, pronunciation overrides, emoji-to-emotion, VoiceCharacter profiles, ElevenLabs v3 laugh tags
@kithjs/observability — Span tracing, duplicate-send guards, reconnect metrics
@kithjs/avatar-events — Placeholder (events live in core; renderer utilities in v0.2)
Reference demo — Working web app with procedural avatar

Production-grade voice quality and real avatar integration.

Feature	Description
LiveKit production adapter	Full WebRTC integration replacing mock mode
Multi-provider TTS	Cartesia + OpenAI TTS with automatic fallback
VRM avatar utilities	Three.js VRM loading, phoneme-to-viseme mapping, blend shapes
Mic input / STT	Full duplex — user speaks, Kith transcribes, agent responds
React Native support	LiveKit RN SDK path for mobile companions

Scale and platform expansion.

Feature	Description
Telephony	SIP/PSTN via LiveKit for phone-based companions
Multi-agent orchestration	Multiple companions in the same voice session
Pluggable memory backends	First-party adapters for common vector stores
Streaming LLM integration	`voice.streamText(tokenIterator)` for real-time LLM output
Voice cloning pipeline	Character voice training workflow

These are intentionally out of scope for Kith at any version:

Replacing your agent runtime. Kith sits beside your agent (LangGraph, custom orchestrator, etc.), not inside it.
Reimplementing memory/RAG. The MemoryAdapter is pass-through. You own your storage.
Owning the avatar renderer. Kith emits events; your app renders however it wants (Three.js, Unity, 2D canvas, whatever).

v0.1 API surface is stable. Issues and PRs welcome, especially for: