Quick start
For people who’ve installed agent stacks before.
git clone https://github.com/ninjahawk/hollow-agentOS
cd hollow-agentOS
pip3 install rich pywebview httpx
python3 hollow.py
The wizard will:
- Detect your OS, GPU, Docker, and Ollama
- Recommend a model based on hardware
- Pull the model (~2–7 GB)
- Bring up the stack on port 7777
- Open the live monitor
Defaults:
| Setting | Value |
|---|---|
| Model | qwen3.6:35b-a3b if you have 24 GB+ VRAM, otherwise smaller |
| API | http://localhost:7777 |
| API token | random, written to config.json |
| Agents | 3 (scout, analyst, builder) |
| Cycle budget | 1500s |
| Watchdog | 2400s (forces container restart on daemon death) |
Day-to-day
# Windows
panel.bat
# Mac/Linux
python3 panel.py
# Watch the live thoughts stream from any platform
python3 thoughts.py
Stop / restart
python3 hollow.py stop # or stop.bat on Windows
python3 hollow.py # starts again (skips wizard if config exists)
Where things live
memory/goals/<agent>/registry.jsonl | Goal history per agent |
memory/identity/<agent>/lessons.json | Promoted operational lessons |
memory/claude_requests.jsonl | Agent → human implementation queue |
workspace/<agent>/ | Per-agent file workspace |
logs/thoughts.log | Tail of live activity (rotates at 50 MB) |
config.json | API token, model config |
Common things you’ll want to do
- Send the agents a message: operator panel → Inject → Host message
- Drop a file into an agent’s workspace: operator panel → Inject → File
- Pause an agent: operator panel → per-agent → Suspend
- Adjust suffering manually: operator panel → per-agent → Suffering load
- Submit a task to an agent:
python3 submit_task.py "your task" --agent scout
For everything else, see Setup and Troubleshooting.