Three small loops, one shared brain.
HKOS — the Home Kibbutz Operating System — is the methodology turned into a runnable thing for a single house. Three loops: fish (the first one, being rebuilt as the Seed's bench loop), bees in the yard, a raised bed in the garden. The bees and the bed are on hold while HKOS itself gets built. They write to the same brain. The point is the wiring, not the substrates.
A pilot loop shut down for its rebuild; the apiary and the FarmBot on hold, all the hardware on hand. Honest status below.
Node Zero — Aquaponics
New England · shut down · coming back as the Seed bench loop
The first loop was a deliberately small pilot: a 30-gallon tank in the basement, goldfish and parsley. It's shut down now. Over the next couple of months it comes back as the first bench loop of the Seed: a desk-size tank and grow bed with open sensors and an energy meter on every load, logging to an open dashboard.
- 30-gallon pilot tank — shut down
- Atlas pH probe (I²C 0x63) — on the bench
- DS18B20 water temp (GPIO4) — on the bench
- Raspberry Pi 4 + breadboard
- Seed bench loop — next couple of months
Rebuild as the Seed's bench loop: the soldered sensor stack, metered loads, and one controlled experiment, run and reported.
Build details in the log. Full ops in the manual.
Node One — Bees
Backyard · on hold while HKOS gets built · hardware on hand
The hive went in this spring. The sensor stack — hive weight, internal temperature gradient, acoustic queen detection — will run the same Pi-and-capture pattern as the aquaponics loop. It's on hold while HKOS itself gets built; everything it needs is on hand.
- Hive bodies + frames
- Smoker, suit, hive tool
- Colony — installed spring 2026
- Hive sensing — weight, temp, acoustics (on hold)
Hive sensing — weight, temperature, acoustics — once HKOS is further along.
First reading flips the badge to LIVE.
Node Two — FarmBot
Built · on hold while HKOS gets built · hardware on hand
A FarmBot Genesis, built, with track calibration still to finish. It's on hold while HKOS itself gets built. Later it's the robot-gantry layer of the Seed, judged by what it saves against what it costs to run. Every action and reading will write to the brain on the same capture pattern.
- FarmBot Genesis — built, on hold
- Raised bed lumber
- Soil, seeds, drip line
Finish track calibration, then the Seed's automation layer.
First soil-moisture reading flips the badge to LIVE.
Same pattern, different substrates.
Every node POSTs to the same /capture endpoint with a domain tag — aquaponics, bees, garden. Captures get embedded and stored in a shared vector index, so a temperature trend in the basement is retrievable next to a weight trend on a hive. Domain scoping keeps queries clean; one shared index keeps the cross-substrate questions possible.
The brain is the part that's furthest along: running since late March 2026, about 2,800 captures and growing. The architecture pattern — capture grammar, sensitivity tiers, retrieval surfaces — is written up in the pages below.
- KIRA, a local assistant I talk to from my phone every day (since September 29) — it answers from its own memory
- Sovereign node — holds the brain and KIRA, and gates what leaves
- GPU node — two Intel Arc B70 cards, serving Qwen3.8 27B as KIRA's heavy model since October
- Escalation — a small local model answers what it can and hands off to the bigger local model when its own knowledge isn't enough
- Continuous capture agent working the live corpus (propose-only)
- The node checks itself every 15 minutes; findings become tickets I approve from the phone
- Wall-power metering on the machines
- Signed cross-node crossing — gated → signed → reply
- Faster answers: under 15 seconds, typical, from the phone
- KIRA reads the project boards, calendar and email (read-only)
- KIRA drafts emails and calendar events; nothing goes out until I approve it
- Better retrieval from the brain as it grows
- Sovereignty is the goal, always. KIRA does less than a cloud assistant on purpose, and the work is adding what it can do without crossing that line. A couple of older connections still lean on cloud services; they're being brought home.
The Cortex →
The four-layer stack behind the brain — ground truth, memory, the local model that decides what stays home, and rented reasoning.
The Gossip Protocol →
How two sovereign nodes will trade conclusions without trading data — the vocabulary, the firewall, and the honest status.
Tokens across the gate →
Every model call, live: which model answered, and how many tokens crossed the gate.
HKOS in the Seed →
The computing layer of Gold's Honey's open-source food system: the controller, the memory, and the local reasoning.