本地 AI 與代理應用
Home Assistant 2026.8 Beta Natively Integrates llama.cpp and LiteLLM, Enabling Fully Self-Hosted Home Conversation Agents
Home Assistant adds conversation integrations for llama.cpp and LiteLLM, allowing users to create separate agents and instructions for different models. The update lowers the barrier to connecting local models, but it remains in beta, and end-to-end tool-calling reliability depends on the model, schema, and hardware.
Home Assistant 2026.8 Beta brings two common self-hosted model pathways into its core integrations. The new llama.cpp integration can connect to a local llama.cpp server and also supports OpenAI-compatible endpoints. Users can create multiple conversation agents, each with its own model and system instructions, so voice or text requests do not have to leave their own hardware. The new LiteLLM integration uses a LiteLLM proxy as a unified gateway. Home Assistant creates agents based on the configured models, allowing cloud APIs, self-hosted inference services, and different providers to share a single connection method.
The technical change does not mean Home Assistant itself now runs models. Instead, model routing and agent configuration have become supported core components. llama.cpp handles model loading, quantization, and the OpenAI-compatible API, while LiteLLM handles provider adapters, authentication, and routing. This lets engineers keep sensitive home-state data on the local network, use small local models for simple intents, and route more difficult tasks to another endpoint without maintaining third-party add-ons.
The real risk lies in the action layer. A model that can converse fluently will not necessarily generate the correct tool names and arguments or execute multi-step calls reliably. Once a home agent can control locks, alarms, or appliances, selecting the wrong entity or omitting a confirmation step can have physical consequences. OpenAI compatibility describes only the API shape; it does not guarantee that every model implements tool-calling semantics, streaming formats, context lengths, or error handling identically.
The current release is a beta, and the stable-release page lists August 5, so the functionality may still change. Testers should create a regression suite of real commands and verify the resulting tool calls individually instead of evaluating only the natural-language responses. They should also restrict which entities are exposed, require confirmation for high-risk actions, and measure latency and memory usage on local hardware with long contexts and multiple agents running concurrently.