Theo's house

TL;DR: A persistent agent for two first-time parents and our real home, built before our son Theo arrives in December. Muse Spark Open decides through tools, remembers across days and speaks first, on real WhatsApp, an Echo, plugs and a sensor I built on Muse Gadgets SDK. Theo is simulated. The real home can't be tested from outside, so https://simulator.bygmv.com runs the same agent with mock events and connections. The video shows the real home. It isn't a baby tracker. Our home has run on it since Nov 3, on our own server, as a rehearsal with a simulated Theo until he is born. LangGraph and PostgreSQL carry its memory across days and restarts. It wakes itself, reads the baby log and the house, plans the night around his feeds and who is on duty, re-plans when a parent says "let me sleep", and speaks first. A schedule only wakes it up. Muse chooses what happens next. Ceci and I get its messages on WhatsApp, through Twilio's sandbox, and we answer there. The kitchen Echo answered me out loud from his log, and it moved the night to me when asked. It has run a full night on our real home, 22:00 to 07:00, with no crash. On Nov 4 my humidity sensor read 76.5 % and woke the agent. Nobody told it what to do. It turned the dehumidifier on and told Ceci, with nobody in the room. I had turned that plug off first so we could film it. SAM 3.1 says "hand" when both hands leave the baby: 33 of 51 within 2.5 s, 0 false alarms in 29 moments, on generated and rendered recordings. Perception Encoder Audio caught 323 of 348 recorded cries. A cry played aloud near the real camera reached WhatsApp as one message, one test, not a rate. The car seat coach never says the seat is safe. The clip you see is my real seat as the camera of Meta's simulated glasses. Waits are cut, voices are added and my answers were scripted. SAM 3.1 didn't see the belt guides there, so the parent confirms the belt by looking. Red flags go through a gate in code before the agent. All 56 emergency lines on record passed with no miss, 3 of 3. The last held-out round scored 36 of 48 on 16 new cases, with real Muse on a mock home and mock channel. In the simulator you can also try the first-month book. Muse writes it from your visit's simulated log and code checks every number. Theo isn't born yet and nobody relies on this yet, so every agent score here is on a mock home and channel. I don't own the glasses yet. All of it is labelled. Built before Oct 26: the Recipe Sender MCP, plus bare-bones prototypes of a hand alarm with SAM 3.1 and of hearing cries. Built during the event: the persistent agent and its memory, night planning, the WhatsApp and Echo connections, control of the real home, the emergency gate, the car seat coach, the sensor, the Muse connector setup for Recipe Sender, the first-month book, the simulator and every test round. I built it so we get the best first month together.