The week Claude watched the house — and us

This one predates half the tooling we use now. No MCP, no purpose-built integrations — just Claude wired directly into the Home Assistant APIs, and a job description delivered in plain English: we’re away for a week; watch the house.

The rules were sentences

There was no rules engine and no YAML. The security policy was a conversation:

• If an exterior door opens or unlocks while we’re gone, and especially if a camera then sees a person inside, alert me immediately — Home Assistant push notification, critical priority, the kind that punches through a silenced phone. I want to know the moment someone walks in.
• Check for false positives before you page me. A motion blip isn’t a person; a low battery isn’t a break-in.
• Here’s the layout of the house — the exterior doors, and how people actually flow through them day to day. Judge what you see against how this house normally moves, and flag anomalies, not just events.
• The cat is home alone. Expect cat-shaped sensor noise; the sitter’s visits should look like the pattern I described, and anything else shouldn’t.
• Some lights change on their own — that’s the automations. Know which ones those are, and worry about the ones that shouldn’t.

That last mile — layout, flow, the cat, the automations — is everything. Any alarm system can tell you a door opened. Knowing that door opening at that hour, in that sequence, isn’t how this house behaves takes context. Context is precisely the thing you can hand an LLM in a paragraph and can’t easily encode in an automation rule.

The week passed with the right number of critical alerts: zero. The system proved it could tell the difference between the cat, the automations, the expected visits — and nothing else.

Then we pointed it at ourselves

The other half of the experiment: the Home Assistant app on my phone was reporting latitude and longitude the whole trip, and Claude had a standing request — keep an eye on us. When you see we’ve settled somewhere new, look around and tell us what’s worth knowing.

So as Sara and I moved through the city, the emails and texts arrived on their own: a note about the neighborhood we’d just landed in, something interesting nearby, timed to where we actually were instead of where an itinerary said we’d be. The one I still bring up: a message suggesting a new exhibit at MoMA — because the coordinates showed we were standing half a block away. We hadn’t asked. We went.

It worked because the assistant already had the full picture: it had read through the trip emails, so it knew the itinerary; it had created the calendar events itself; and it kept the family in the loop as plans firmed up — one running thread of awareness across email, calendar, location, and house.

What this one taught us

• Plain English is a real policy language. The entire security posture and the entire concierge behavior were sentences, revised by saying new sentences.
• Critical alerts must be rare to be believed. The false-positive discipline was the feature. An alert channel that cries wolf is worse than none.
• Anomaly detection is context plus flow, not sensors. The exact same door event is routine or alarming depending on the story around it.
• The same loop serves opposite jobs. Vigilance at home, serendipity on the road — one awareness, pointed two directions at once.

Everything we’ve built since — the MCP servers, the wall tablets, the trackers — is this week’s experiment, productized. This was the proof that the assistant model works: not a dashboard you check, but a colleague who’s paying attention.