I had 4 old laptops collecting dust and was frustrated paying for AI coding tools. So I built Ralph Loops — a system that coordinates multiple AI agents to work on code overnight while I sleep.
How it works:
- Write tasks in markdown, commit to Git
- Run `start-night.sh` before bed
- Agents (old laptops on Tailscale) claim tasks, run Claude/Gemini CLI, push results
- Manager agent reviews work, auto-creates fix tasks for failures
- Morning: run `morning-review.sh` to see what happened
What's actually working:
- 95% overnight execution success rate
- Git-based coordination (no central server)
- Automatic retry/fix cycle for failed tasks
- Heartbeat monitoring so I know agent status
What's NOT magic:
- You still write the task specs
- AI makes mistakes — the manager catches ~70% automatically
- Complex features need multiple task iterations
- This is a "batch processing" model, not real-time pair programming
Cost: ~$15/month electricity vs $500/month for Devin. Trade-off is you need the hardware and patience to set it up.
Code is MIT licensed. Happy to answer questions about the architecture or failure modes (there were many).
