// the take
A Home Assistant custom integration that sends your entity/device state to an AI provider and gets back suggested automation YAML. Supports 10+ providers including local ones (Ollama, LocalAI). Good fit for HA users who want automation ideas but don't want to write YAML from scratch.
- Provider breadth is genuinely useful: supporting Ollama/LocalAI means privacy-conscious users can keep all data local, and the model catalog with per-provider defaults and custom model name override is pragmatic.
- Avoids the openai Python SDK entirely and uses plain HTTP calls, which is the right call for a HA custom component where SDK version conflicts are a real operational headache.
- Suggestion history with accept/decline/dismiss review states and a REST API for the bundled Lovelace card shows thought about the actual UX workflow, not just dumping text into a notification.
- Tests exist for the non-trivial utility modules (endpoint_utils, language_utils, model_catalog, suggestions parsing) and CI runs hassfest and HACS validation, which is above average for a custom component.
- The model catalog lists model IDs like 'gpt-5.4-mini', 'claude-sonnet-4-6', and 'gpt-5.5-pro' that don't exist as of mid-2025 — this looks like aspirational/hallucinated model names baked into code, which will silently produce API errors for users who trust the defaults.
- The core value proposition — quality of YAML suggestions — depends entirely on prompt engineering that isn't versioned or user-visible, and there's no mechanism to detect when returned YAML references non-existent entity IDs or service calls, beyond a simple parse check.
- Test coverage only hits utility functions; coordinator.py and api.py (the actual API call logic and HA integration wiring) have no tests, so regressions in the critical path will only surface in production.
- Sending potentially hundreds of entity states and attributes to cloud APIs on a weekly schedule with no built-in cost estimation or token usage reporting is a footgun for users with large installations; the entity_limit is manual and easy to misconfigure.
We read the README, the directory tree and a key source file in May 2026. Nobody pays to be listed. How we judge →