// the find
AlphaAvatar/AlphaAvatar
A real-time interactive Omni Avatar built on LiveKit, which allows you to seamlessly integrate with any open source Avatar components (real-time model, visual, voice, memory, search, etc.).
AlphaAvatar is a self-hostable personal AI assistant framework built on LiveKit, combining real-time voice/video, memory, persona tracking, RAG, MCP tools, and an optional animated virtual character. It targets developers who want to run a persistent, multimodal AI companion on their own infrastructure rather than paying for a hosted service. At v0.6.2 it is genuinely usable for the core voice+memory flow, but much of the ambitious architecture is still on the roadmap.
The plugin packaging is done right — memory, persona, character, deepresearch, and MCP are each separate installable packages with their own pyproject.toml, so you only pull in what you actually need. The provider abstraction layer tracks token usage per provider (Anthropic, OpenAI, Google, OpenRouter) with dedicated usage extractors, which is exactly what you want before a bill surprises you. LanceDB ships as the default fallback VDB when Qdrant isn't configured, making local self-hosting viable without standing up a separate vector DB service. The face detection + speaker vector fusion for multi-user identity in the Persona plugin is non-trivial work that most similar projects skip entirely.
Four of the seven core plugins — Reflection, Planning, Behavior, and Sandbox — are listed as Planned with no implementation. The architecture's headline promises (self-improvement, long-horizon planning, proactive assistance) are not yet there. The WhatsApp channel is built on Baileys, an unofficial reverse-engineered client that WhatsApp actively bans; this is a fragile dependency you cannot rely on in production. The LiveKit transport is baked in at every layer — there is no abstraction over the real-time layer, so if you need a different WebRTC backend or a purely local mode you are in for a significant rewrite. Documentation for multi-user deployment is absent; the architecture diagram shows a single session per room, and it is unclear how this scales or isolates state across concurrent users.