| name | visio-sentinel |
| description | Use when the user asks about home security camera events, unknown person alerts, camera status, or wants to manage known persons, configure alert thresholds, or start/stop the visio-sentinel vision AI daemon. Use when the user says "who is at the door", "show me recent security events", "add person to security system", or "is the security camera running".
|
| version | 1.0.0 |
| author | arananet |
| license | MIT |
| metadata | {"hermes":{"tags":["smart-home","security","camera","vision","ai","ollama","local-first"],"related_skills":["home-assistant","telegram-notifications","local-ai"],"config":[{"key":"skills.config.visio_sentinel_dir","description":"Absolute path to the visio-sentinel installation directory","default":"~/visio-sentinel","prompt":"Where is visio-sentinel installed? (e.g. ~/visio-sentinel)"},{"key":"skills.config.visio_env_file","description":"Path to the .env file used by the daemon","default":"~/visio-sentinel/.env","prompt":"Path to visio-sentinel .env file?"}]}} |
visio-sentinel
visio-sentinel is a local-first, offline home security system.
It fine-tunes a small vision model (Moondream2) on your family's photos, runs it
via Ollama, and streams alerts to Telegram (or Home Assistant) when an unknown
person is detected.
Installation directory: ${HERMES_SKILL_DIR}/../../../../
Use the configured path: skills.config.visio_sentinel_dir
Available CLI commands
All commands run via visio_ctl.py from the installation directory.
python visio_ctl.py status
python visio_ctl.py events [N]
python visio_ctl.py persons
python visio_ctl.py add-person <name>
python visio_ctl.py remove-person <name>
python visio_ctl.py tail
python visio_ctl.py start
python visio_ctl.py stop
Configuration
All runtime config lives in .env. Key variables:
| Variable | Purpose | Default |
|---|
CAMERA_SOURCE | USB index (0,1,…) or RTSP URL | 0 |
FRAME_SAMPLE_INTERVAL | Seconds between samples | 2 |
MOTION_PIXEL_THRESHOLD | Pixel delta to trigger emit | 500000 |
CONFIDENCE_THRESHOLD | Below this → treat as unknown | 0.80 |
UNKNOWN_ALERT_HOUR_START | Night window start (24h) | 22 |
UNKNOWN_ALERT_HOUR_END | Night window end (24h) | 6 |
TELEGRAM_BOT_TOKEN | Bot token from @BotFather | — |
TELEGRAM_CHAT_ID | Your chat or group ID | — |
HA_ENABLED | Enable Home Assistant webhook | false |
HA_WEBHOOK_URL | Home Assistant webhook URL | — |
SNAPSHOT_DIR | Where to save alert snapshots | ./snapshots |
LOG_PATH | Structured JSON event log | ./logs/events.jsonl |
Alert rules are in agent/skills/home_security.md — edit Known Persons and
Thresholds there to manage family members and confidence settings.
Common tasks
Check if someone unknown was detected today
python visio_ctl.py events 50 | grep -i unknown
Add a new family member
python visio_ctl.py add-person alice
python dataset/prepare.py --source videos/alice.mp4 --label alice
python dataset/augment.py
python training/finetune.py
Restart daemon after config change
python visio_ctl.py stop
python visio_ctl.py start
Check confidence on a specific frame (one-shot)
python inference/analyzer.py
Event log format
Every classified frame appends one JSON line to LOG_PATH:
{"ts": "2026-05-07T14:32:01Z", "person": "edu", "known": true, "confidence": 0.94, "priority": "INFO", "error": null}
Priority levels: INFO (known person), WARNING (unknown, daytime),
CRITICAL (unknown, night window), ERROR (parse/inference failure).
Troubleshooting
- Daemon not running:
python visio_ctl.py start — check .env is present
- Always unknown: lower
CONFIDENCE_THRESHOLD or add more training data
- No Telegram alerts: verify
TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID in .env
- RTSP won't open: check URL format
rtsp://user:pass@ip:554/stream
- Metal OOM on Mac: reduce
training.batch_size in training/config.yaml
