Laya for Home Assistant: a fully local Assist conversation agent

# Laya for Home Assistant

A fast, fully local conversation agent for [Home Assistant Assist](https://www.home-assistant.io/voice_control/) powered by the open-weight [Laya](https://huggingface.co/collections/laya) decision engine.

> [!NOTE]
> Looking for a cloud-accelerated version? See its sibling project [**home-assistant-jev**](https://github.com/allenporter/home-assistant-jev), powered by TypeSafe AI's hosted Jev API. Both integrations share the same 5-stage speculative decision pipeline architecture.

Laya is a lightweight, non-autoregressive decision model that evaluates structured choices in a single forward pass. It integrates directly as a native **Conversation Agent** in Home Assistant Assist voice pipelines—dispatching routine device control commands directly with zero cloud reliance, while seamlessly escalating compound or open-ended conversational queries to a configured fallback LLM.

Instead of generating free-form text token-by-token and parsing the output back into code, Laya scores the probability of registered Home Assistant intents and exposed entities directly in a single pass.

## Highlights

- **Native Assist Voice Agent**: Selectable as a first-class conversation agent in Assist pipelines (Voice PE, satellite microphones, dashboard Assist).
- **100% Local & Private**: Runs entirely in-process on local hardware (CPU, CUDA, or Apple Silicon). No external API keys, cloud subscriptions, or audio/device state leaving your local network.
- **Fast Single-Pass Execution**: Avoids the multi-second latency and overhead of autoregressive generative token generation.
- **Zero Hallucinated Targets**: Candidate targets are constrained to entities and areas currently exposed to Assist (`async_should_expose`).
- **Speculative Fan-Out**: Evaluates intent, area,