finds.dev

goruck/home-generative-agent

AI agent for Home Assistant — talk to your home, create automations in plain language, analyze cameras with face recognition, and get proactive anomaly alerts. Cloud LLMs or fully local.

View on GitHubHomepage ↗

★ 372PythonMITupdated Nov 2026

// the take

A Home Assistant custom integration that replaces simple prompt-passthrough conversation agents with a real LangGraph agent: it calls tools to control entities, write and register automations from plain-English requests, analyze camera feeds with face recognition, and run a separate rule-based anomaly/security engine (Sentinel). It's for Home Assistant OS/Supervised users who want an LLM-driven assistant for their home and are comfortable running a Postgres instance and possibly local model infrastructure.

Automations generated from chat are checked against real entity IDs and service calls before being written, so the model can't install an automation that references a guessed entity or a nonexistent state. Sentinel keeps the actual security decisions (unlock detection, unknown-person alerts, network/device audits) in deterministic rule code with the LLM only advising, which is the right split for anything that can affect physical access. It genuinely supports both cloud (OpenAI, Gemini, Anthropic) and fully local (Ollama, any OpenAI-compatible server) model backends, including local STT/TTS, so privacy-sensitive users aren't locked into a cloud dependency. The project has a real test suite and CI across lint/test/HA-compatibility, and redacts MAC/IP addresses before anything reaches a language model.

The feature set demands a lot of infrastructure for a home user: a dedicated PostgreSQL+pgvector add-on that only runs on HA OS/Supervised (not HA Container/Core), an optional external face-service for face recognition, and optionally a GPU box for local model serving plus STT/TTS. Running it well locally is a real sysadmin project, not a toggle. It's maintained by a single person, which is a meaningful bus-factor risk for something wired into door locks, alarms, and automation execution. The surface area is large and growing fast (dozens of Sentinel rules, a discovery engine, network/radio audits, multiple config flows) — more moving parts generated quickly increases the chance that an edge case in automation writing or trigger validation does something unintended to a physical device. Any cloud-model usage means ongoing per-call API costs on top of the home automation setup.

We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
372
Forks
68
Language
Python
License
MIT
Last commit
2 Nov 2026

// nearby

If home-generative-agent isn't quite it.

All repos →