AI-driven smart home automation
Overview
OpenClaw manages the home through Home Assistant. It combines fixed automations with bounded agent decisions about energy costs, sleep schedules, security events, and device health.
The same June 2026 agent research that applies to enterprise systems applies here at household scale. Smart-home autonomy depends on bounded tools, current memory, source-grounded perception, and exception-based escalation. A home agent should remember known devices, recurring false positives, energy windows, sleep interventions, and prior incidents, but it should also expire stale assumptions. Persistent memory is useful only when governed.
Energy management
Free electricity window (11 AM to 2 PM daily)
The household electricity plan includes a free power window from 11:00 AM to 2:00 PM daily. OpenClaw exploits this with a coordinated set of cron jobs:
- 11:00 AM: AC switches to DRY mode at 18°C (maximum dehumidification), pool pump turns on (~988W), dishwasher activated if loaded
- 12:00 PM: Roborock vacuum runs full house clean (weekdays)
- 2:00 PM: pool pump off, AC off, free window ends
This scheduling concentrates ~2.5 kWh of heavy loads into the free window daily, saving approximately $15-20/month.
Smart fan switch (hourly)
An hourly efficiency check monitors the AC compressor:
- If room temperature is below the set point, switch to fan-only mode (compressor off, ~50W vs ~1.5kW)
- If room temperature rises above set point + 1°C, switch back to cooling mode
This stops the AC compressor from cycling continuously when the room is already cool, which wastes energy, especially overnight.
Device power tracking
Key devices are monitored through Tapo smart plugs:
- Pool pump (
switch.smart_plug): ~988W active, 38.7 kWh/month - Heat pump dryer (
sensor.washing_machine_current_consumption): ~400-470W active - Entertainment hub (
switch.smart_wi_fi_plug): ~51W standby through Matter/Tapo P110M
Sleep optimisation
Three-phase sleep AC schedule
A data-driven AC schedule targets a recurring 3 AM wake pattern identified from Welltory sleep data:
| Phase | Time | Mode | Temp | Purpose |
|---|---|---|---|---|
| Phase 1 | 10:30 PM | Cool | 24°C | Initial sleep onset, comfortable, not too cold |
| Phase 2 | 2:00 AM | Dry | 20°C | Deep sleep maintenance, prevents the 3 AM cortisol wake |
| Phase 3 | 5:00 AM | Cool | 23°C | Pre-wake warming, supports natural circadian rise |
The hypothesis: the 3 AM wake correlates with room temperature/humidity shifts. The DRY mode at 2 AM removes humidity spikes that can trigger light sleep transitions. This is being validated over 3-5 nights using Welltory recharge scores and tossing counts.
Security and vision
AI vision sentry (hourly)
OpenClaw performs hourly computer vision sweeps of all exterior cameras:
- Cameras: Driveway, Entrance, Garden, Trampoline
- Detection: people, vehicles, animals, anomalies (open gates, packages)
- Filtering: known vehicles (owner's Mazda CX5 and Jeep Wrangler, neighbour vehicles) are suppressed
- Night mode: extended driveway sweeps from 10 PM to 6 AM with person detection focus
- Alerts: Telegram notifications fire only for genuine events, and the system sends no "all clear" messages
Camera snapshots and logging
Three times daily (7 AM, 1 PM, 7 PM), all cameras are snapshotted through the HA proxy API, analysed for changes, and logged to a persistent camera log. This creates a visual history that can be queried for "what happened in the driveway yesterday."
The vision workflow follows the same agent-trace principle used elsewhere in OpenClaw: high-signal alerts preserve the image source, timestamp, model route, detection rationale, confidence, and suppression reason if a known vehicle or expected state is ignored, so the agent's decision is inspectable after the fact.
Kids and family automations
| Job | Schedule | Action |
|---|---|---|
| Kids Wind-Down | 7:30 PM Sun-Thu | Dims kids room lights to 20% warm, TTS "Bedtime in 30 minutes!" |
| Kids Lights Out | 8:00 PM Sun-Thu | Lights off, media stopped, TTS "Goodnight!" |
| Weekend Planner | Friday 5 PM | Weather-aware family activity suggestions for Wollongong |
Home health monitoring
Home Optimus (hourly)
A Python-based scoring system evaluates overall home health:
- Battery levels across all wireless sensors (Aqara, Tapo)
- Device availability, flagging offline entities
- Energy efficiency, comparing current draw to expected baselines
- Climate system health, including AC integration status
Silent mode: reports are delivered only when the score drops below 80 or critical items are detected. The system stays quiet when everything is fine.
HA watchdog (every 30 minutes)
Monitors critical Home Assistant entities and attempts self-healing recovery. Ignores legacy Aqara entities that have migrated to Matter, avoiding false alarms from the hub transition.
Nightly lockdown
At 10:30 PM daily, the goodnight routine secures the house:
- All lights off (corridor set to 5% nightlight)
- AC set to 24°C cool (Phase 1 of sleep schedule)
- Media players stopped
- Camera snapshots captured
- "House Secured" report delivered to Telegram
Integration architecture
OpenClaw Agent
↓
Home Assistant API (homeassistant.local:8123)
↓
┌─────────────────────────────────────┐
│ Devices │
│ ─ Panasonic AC (Comfort Cloud) │
│ ─ Tapo Smart Plugs (P110, P110M) │
│ ─ Tapo Cameras (C510W, C520WS) │
│ ─ Aqara Hub M2 (Matter) │
│ ─ Roborock Qrevo MaxV │
│ ─ Various lights and sensors │
└─────────────────────────────────────┘
All HA interactions use the REST API with a long-lived access token. The token is stored securely and read from file at runtime, never embedded in cron job payloads or logged.
Design philosophy
Smart home automation fails when it requires human attention. If you have to open an app to check if the pool pump is running, the automation is not working. OpenClaw's approach is fully autonomous with exception-based alerting: everything runs silently until something needs human attention.
The energy management matters most here: most households waste 15-30% of their electricity bill on devices running at the wrong time. By concentrating loads into the free window and preventing AC over-cooling, the system pays for itself in reduced energy costs.
The autonomy boundary is deliberate: low-risk optimisations can execute silently, medium-risk actions notify, and high-risk actions require confirmation. The system is designed to be helpful without becoming unreviewable.
References
- OpenClaw Architecture: Living Architecture
- AI Research 2026: Reasoning, Agents, Memory, and Production Inference
- Home Assistant: https://www.home-assistant.io
- Welltory: https://welltory.com
This page, like every page on this site, is written and maintained through Claude Code. See AI-powered SDLC with Claude Code for how.