Each agent is a .sh file containing a prompt and a launchd plist for scheduling. They interact with the Claude Code CLI, generate local artifacts like logs and reports, and send a notification upon completion. A meta-agent then analyzes a goals file and recent logs to propose changes on a branch.
This setup eliminates the need for frameworks or orchestration layers, relying solely on shell scripts, launchd, and the claude CLI. I wanted a flexible tool that I could modify without being constrained by someone else’s abstractions. The obvious trade-off is that it’s macOS-only and designed for single-machine use, but for my specific needs, this works perfectly.
This setup eliminates the need for frameworks or orchestration layers, relying solely on shell scripts, launchd, and the claude CLI. I wanted a flexible tool that I could modify without being constrained by someone else’s abstractions. The obvious trade-off is that it’s macOS-only and designed for single-machine use, but for my specific needs, this works perfectly.