Questions
These are the questions people ask when they hear that an AI assistant runs a household from a box on a shelf. The answers below are specific, dated, and drawn from the running system rather than from a plan.
It is a private AI operating layer that has run one household. A person talks to it through ordinary messaging apps or an LLM chat app, and it answers from a readable knowledge base of the household's own notes and records rather than from conversational memory. The coordination layer and all the data run on a home server the owner controls. The AI models are rented from external providers and are treated as interchangeable parts.
No. The home server has no GPU and cannot serve a model worth using, so inference is called from three commercial providers. What stays at home is everything else: the orchestration, the memory, the scheduling, the tool access and all of the data.
This split is deliberate. Models improve every few months and are replaceable in a config change, while the knowledge base and the operating discipline around it are the parts that took a year to get right.
A single fanless mini-PC, roughly the size of an Apple TV, with a low-power Intel N100 processor and 16 GB of memory. It runs a hypervisor with two virtual machines: one for the assistant and its scripts, one for home automation. It draws a few watts, has no moving parts, and has run continuously for months at a time.
As of September 2026 it calls Anthropic, OpenAI and Google models, chosen per agent rather than globally. The general assistant runs on a stronger reasoning model; the narrow agents for travel, groceries and chores run on a fast, inexpensive one.
Each agent has a two-deep fallback chain that crosses vendors, so an outage at any single provider degrades answer quality instead of taking the assistant offline.
All of it lives on the home server as plain text files, backed up nightly to the owner's own cloud storage. No third party holds the knowledge base. AI providers see only what a single request sends them, which is the relevant excerpt for that question and not the archive.
Because the knowledge base is ordinary readable text rather than a vendor database, it can be opened, audited, exported or moved to a different AI model without a migration.
Facts are read from source files at the moment they are needed, and answers carry the source they came from. Nothing enters long-term memory automatically: once a week the assistant proposes a short list of candidate facts and a human approves or rejects each one.
That design was chosen after testing the alternative. Letting the system curate its own memory unsupervised made accuracy worse, not better, because it compounded its own errors.
You stop asserting on the wording and start asserting on the things that are deterministic.
A weekly job asks five real household questions whose answers can be derived mechanically from the underlying records, each in a brand-new conversation so that no prior context props up the answer, and checks each reply against the record read live at that moment. Separately, a nightly audit with no AI in it verifies that the configured model is the model actually serving, that no scheduled process has silently stopped, and that no memory file points at something that no longer exists. Both alert only on failure, because a system that reports success trains its owner to ignore it.
The request moves to the next model in that agent's fallback chain, which belongs to a different vendor, and the conversation continues. A watcher notices the switch and messages the owner when the system drops to a backup and again when it recovers.
An exhausted balance or a monthly spending cap looks the same to the system as an outage: the request fails, the next provider picks it up, and the conversation carries on. That is not hypothetical. A spending cap was reached once mid-month, and the assistant kept working on another vendor's model until it was raised. The nightly audit also looks for recent billing and authentication errors, so a quietly failing account surfaces the next morning rather than weeks later.
Separately, a healer process watches for a specific provider error that corrupts a conversation and silently stops replies rather than failing loudly. It resets the affected session and asks the person to resend. Failures that announce themselves are easy; the ones worth building for are the quiet ones.
It works inside WhatsApp and Telegram, in English and Hebrew, for the whole family. It reads and summarises mail and calendar, manages a task board, runs the connected home, sends a daily briefing, maintains the family shopping list, tracks household vendors and payments against receipts, ran the coordination for a three-week family trip, and runs a chores and allowance system the children use directly.
Narrow jobs get their own agent with its own small brief rather than being added to one general assistant. A specialist with a tiny, focused remit is markedly more reliable than a generalist trying to hold everything.
Anything that has to add up is not left to the model. The shopping list, the chores and points ledger, and the household payment ledger are each written by a small purpose-built program that owns the format: the assistant interprets the request, the program records the result. That is what keeps a running total correct on a day when a model is having an off day.
Four controls carry most of the weight.
Not today. It is a working personal system running in one household, not packaged software, and it is honest about the difference: it currently depends on a technically capable owner maintaining it.
The present step is a small number of private conversations with people who want this kind of operating layer built around their own lives.
Etgar Bonar, a commercial executive based in Tel Aviv whose background is revenue leadership at Lokalise, Rapyd, Taboola and Amazon. He built and operates the system himself, choosing to learn AI by running a real one in daily life rather than by studying it. He also built a mobile development workspace so the system can be inspected and changed from a phone.