Selora AI — Home Assistant Integration

Selora AI is a smart-home AI butler for Home Assistant. It connects to an LLM backend — Selora AI Local (our own on-device model), Anthropic Claude, OpenAI, Google Gemini, or Ollama — learns your home's patterns, and proactively generates automations, all while keeping you in full control.

Documentation

Features

Requirements

• Home Assistant 2025.1 or later
• For Selora AI Local: a self-hosted llama-server serving the Selora AI model, reachable from your HA host. No API key.
• For Anthropic Claude: an Anthropic API key
• For OpenAI: an OpenAI API key
• For Google Gemini: a Google AI Studio API key
• For Ollama: a running Ollama server reachable from your HA host

Installation

See the installation guide for detailed instructions.

Selora AI Local

Selora AI Local is our own task-tuned model that runs entirely on your network — nothing leaves your home. Released in v0.10.0, it's the recommended way to use Selora AI when privacy or offline operation matters.

What it is

• Base model: Qwen3 1.7B.
• Four LoRA adapters, each fine-tuned on a specific Home Assistant task and hot-swapped per request:
  • command — execute natural-language commands ("turn off the kitchen light").
  • automation — generate automations.yaml blocks from a prompt.
  • answer — answer questions about your home state.
  • clarification — ask the user to disambiguate when intent is unclear.
• Weights, adapters, and trained system prompts are published at huggingface.co/selorahomes/Selora-AI.

Where the model runs

Selora AI Local talks to a llama-server instance (from llama.cpp) over its OpenAI-compatible HTTP API. You run llama-server yourself — with the base model loaded and the four LoRAs registered as slots — see Running with llama.cpp below.

During setup, the integration probes common locations (localhost and the Home Assistant Supervisor bridge network) for a reachable server and pre-fills the host URL. If none is found, point the integration at http://<host>:8080 during the Selora AI Local config step.

The integration handles LoRA-slot activation (POST /lora-adapters) per request, so the right specialist answers each call.

Running with llama.cpp

Install llama-server from llama.cpp (build from source, or brew install llama.cpp / winget install llama.cpp).

Download the files from huggingface.co/selorahomes/Selora-AI:

• Base model: qwen3_17b_base.Q6_K.gguf.
• LoRA adapters (one per specialist):
  • selora-v047-command.f16.gguf
  • selora-v047-automation.f16.gguf
  • selora-v047-answer.f16.gguf
  • selora-v047-clarification.f16.gguf

Start the server with all four adapters registered but not applied — the integration activates the right one per request:

llama-server \
  --model qwen3_17b_base.Q6_K.gguf \
  --ctx-size 8192 \
  --ubatch-size 1024 \
  --n-gpu-layers 999 \
  --cache-reuse 256 \
  --mlock \
  --jinja \
  --reasoning off \
  --lora-init-without-apply \
  --lora selora-v047-command.f16.gguf,selora-v047-automation.f16.gguf,selora-v047-answer.f16.gguf,selora-v047-clarification.f16.gguf

Verify it's up:

curl http://localhost:8080/v1/models

Expected response:

{"models":[{"name":"selorahomes/Selora-AI","model":"selorahomes/Selora-AI","modified_at":"","size":"","digest":"","type":"model","description":"","tags":[""],"capabilities":["completion"],"parameters":"","details":{"parent_model":"","format":"gguf","family":"","families":[""],"parameter_size":"","quantization_level":""}}],"object":"list","data":[{"id":"selorahomes/Selora-AI","aliases":["selorahomes/Selora-AI"],"tags":[],"object":"model","created":1781658061,"owned_by":"llamacpp","meta":{"vocab_type":2,"n_vocab":151936,"n_ctx":8192,"n_ctx_train":40960,"n_embd":2048,"n_params":2031739904,"size":1667055616}}]}

Point the integration at http://<host>:8080 via Settings → LLM Provider → Selora AI Local → Show Advanced Options → Host. Generation parameters (temperature=0.0, stop tokens, per-intent max_tokens caps, LoRA hot-swap) are managed by the integration — no tuning required on the server side.

Privacy

Languages

Selora AI is localized on two levels:

• Interface — the config flow, entity names, and error messages ship translations for English, French, German, Spanish, Italian, Dutch, Hungarian, Portuguese, Russian, Japanese, Korean, Simplified Chinese, and Traditional Chinese. Home Assistant picks the one matching your configured language and falls back to English otherwise.

• Conversational replies — chat and Assist responses follow your Home Assistant language. Selora AI instructs the model to answer in that language; the following are recognized explicitly:

  English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Swedish, Danish, Norwegian, Finnish, Czech, Russian, Ukrainian, Turkish, Hungarian, Japanese, Korean, Chinese.

  Any other language code falls back to English replies. Entity IDs, service names, and code blocks are always left untouched regardless of language.

Learn More

License

MIT License. See LICENSE for details.