Describe It, Part Two: You Can Talk to the House Today

The AI layer I wrote about is connected. You can sit at your desk and ask the house what's on. Here's what it can do, what it deliberately can't, and the network guard that came along with it.

Describe It, Part Two: You Can Talk to the House Today

Last time I wrote about putting an AI layer on a house full of brands: describe what you want instead of memorizing device names and routine names. I said I'd report back.

0:00
/0:00

It's connected.

What works today

Home Assistant has a built-in way for AI assistants to connect to it, and the desktop AI assistant I use already knows how to talk to it. You add the house as a connection, sign in with your own Home Assistant account on Home Assistant's own login page, and that's it. The assistant gets a key that I can revoke at any time, and it only sees the devices I've chosen to share with it.

The first thing I asked it was what was still on in the master bath. It came back with the closet, the overheads, the vanity, and the shower light. No device names, no app, no dashboard. A question, and an answer from the live state of the house.

That's the "describe it, don't name it" idea, working in a chat window. Saying it out loud is next: first through the speaker in the wall panel, then dedicated listeners around the house. That part isn't built yet, and I'll say so when it is.

Getting connected took two fixes, for anyone trying this at home. Home Assistant needs to know its own public web address, or the sign-in can't finish. And the assistant needed to be told which app it is, in a settings field I'd never have guessed. Neither took long once the cause was clear, but both failed with messages that didn't point at the cause.

What it can't do, by design

Last time I said the AI would never unlock a door or open the garage, and that the rule would live in the tool, not in a prompt.

It's stronger than that now. The locks and the garage aren't shared with the assistant at all. It can't unlock the front door because, as far as it knows, there is no front door lock.

The same clean-up took the network gear off the list: the power plugs for the router, the firewall, the camera recorder, and the computers in the rack. About sixty things came off. The assistant gets lights, fans, TVs, speakers, and climate. The things that keep the house running aren't one confused sentence away from being switched off.

"Grounded" now means the internet too

The kids' TVs already had curfews and a grounded switch. Now grounding a kid also cuts the internet on that kid's computers, and so does bedtime.

Two details I cared about. First, it's computers only. Phones and tablets have their own screen-time controls, and the TVs have their own curfews; this covers the things nothing else can manage. Second, "internet off" doesn't mean "everything off." Operating system updates, antivirus updates, and the clock still get through, so a grounded laptop doesn't come back a week later three patches behind.

School laptops go to school

The school Chromebooks have their own network that reaches the school's learning site and Google sign-in, and nothing else. All the time, not just at night. Homework works. Everything else doesn't.

A network guard: the easier button

This is the piece I'm happiest with, and it isn't really about the kids.

Something now watches every Wi-Fi network in the house. If a device the network has never seen tries to join, it's blocked the moment it connects, and I get a notification with two buttons: Allow, or Keep blocked.

That's a security setting first. The usual way to do this is a list of every approved device's hardware address, kept up to date by hand, forever. Nobody keeps that list up to date. This flips it: everything already here is known, anything new has to be approved once, and the list maintains itself.

The side effect parents will like: if a kid learns the password to a different network in the house and tries to move a laptop over to it, the guard knows that laptop belongs on its own network and blocks it on the new one.

It earned its keep on day one, in an unexpected way. It flagged an unknown "iPhone." Claude's first theory was a Wi-Fi extender rewriting addresses. I don't own an extender. So it went and looked: the "iPhone" was a digital photo frame. Those frames invent a brand-new hardware address every time they restart, and the router had simply reused an old name for the address. The guard now knows what the frames look like and lets them onto the one network they belong on, and nowhere else.

Confidence is not accuracy. Same lesson as last time, from the other direction.

For anyone who wants this

None of this needed custom software. It's Home Assistant's own features, a careful pass over what the AI is allowed to see, and a small watcher on the network. I wrote the setup down as a checklist, so it can be done in another house the same way, with the same limits.

I'll report back when I can say it out loud.