Local AI Guide

See the speed before choosing hardware

Agents and automation

Can a local LLM run the house? Exposure scope and safety boundaries

Safety comes less from how smart the model is than from which devices you show it.

A local LLM1 interprets the sentence and proposes an action, but the Assist API2 in Home Assistant performs it. Safety therefore lives in the exposed list and the automation conditions, not in model quality. This guide covers leaving risky devices out and starting with reversible actions.

Contents
Requirements and key details
  • The exposed list decides what can run, and the model chooses only within it.
  • Leave out locks, garage doors and heaters, where a wrong action is dangerous.
  • Start with reversible lights and switches; keep sensors read-only.

The model proposes; Home Assistant executes

The conversation model reads the sentence and decides which device to act on and how. That decision does not flip the switch by itself: Home Assistant executes through the Assist API, only for exposed entities.

Changing the model leaves the risk surface unchanged. Widening the exposed list raises risk even with a weaker model. The knob for safety is the list, not the model.

A waveform and a document on a monitor feed into a padlocked enclosure, while a dashed line toward the cloud is crossed out
The devices a model can pick are limited to the exposed list.

What to leave out, what can stay

Leave locks, garage doors, heaters and fans off the exposed list; a wrong action can injure someone or damage property. Sensors only report values, so exposing them first carries little risk.

A light or switch left on can simply be turned off again. Start with one room's lights and widen as the phrasing works.

Start with actions you can undo

Locking a door and turning on a light fail differently. Undoing the first needs a key or a code; undoing the second needs one more sentence.

Automations can wrap the action. If the model asks for 'turn everything off at night', the automation turns off only the listed devices and leaves the rest alone.

A robot inside a clear enclosure sits beside a router, storage box, and documents
Leave locks and garage doors off the list; start with lights.

Closing the paths out of the house

A local model keeps questions inside. An external API model sends the sentence and device state to that server. Check settings for which one you run.

Voice services connected over Wyoming3 have no authentication. Keep them on a trusted home network and do not open their ports on the router.

Decide the default when it fails

For the moments a model answers oddly or not at all, keep critical actions on sentence triggers. Turning off lights keeps working even with the model down.

Keep a record of what ran. The log shows which sentence moved which device, and if an odd action repeats, remove that device from the list.

A person at a laptop is surrounded by connected document, spreadsheet, folder, image, settings, and robot panels
Keep critical actions working without the model, on their own.

Terminology notes

  1. Large language model — A model trained on large text datasets to learn language patterns and process or generate text in context.

    Back to the text

  2. API — A defined interface that lets other code call a program’s functions.

    Back to the text

  3. Wyoming — The protocol Home Assistant and voice services use to exchange audio and text. It has no authentication or encryption, so it is used on trusted networks only.

    Back to the text

Frequently asked questions

What should you check first in this article?
A local LLM interprets the sentence and proposes an action, but the Assist API in Home Assistant performs it. Safety therefore lives in the exposed list and the automation conditions, not in model quality. This guide covers leaving risky devices out and starting with reversible actions.