Claude Code plugin for agents that ship while you sleep

What it does

This is a Claude Code plugin marketplace that packages two skill sets: a codebase harness to make repositories safe for autonomous agent modification, and a loop engine that lets agents wake up, do work, and log what they learned to a shared file-based memory. Instead of prompting task-by-task, you design loops that read their own history and keep going. Multiple loops—support, SEO, product—can read and write the same folders so knowledge compounds without human hand-holding.

The interesting bit

The shared brain is just a directory structure. The plugin bootstraps ARCHITECTURE.md, LOG.md, and domain folders that act as a filesystem-based memory bus for agent swarms. It is a surprisingly low-tech glue layer between Claude Code and your repo, but that simplicity is what lets parallel loops share state without a database.

Key highlights

• Installs as a Claude Code marketplace plugin, not a standalone framework
• setup-codebase-harness prepares repos with local dev stacks, e2e gates, and per-agent isolated cloud boxes via crabbox-setup for parallel execution
• new-loop bootstraps a file-based knowledge base and scaffolds a loop that reads CLAUDE.md, does work, appends to LOG.md, and ships via a verified PR flow
• The pr skill spawns a fresh sub-agent to prove a feature works before opening the pull request
• Requires git and a working build/test setup; degrades gracefully if Codex review is unavailable

Caveats

• Hard-locked to Claude Code; useless if you use a different agent stack
• Parallel cloud isolation needs extra Crabbox/Daytona tooling beyond the default install
• The README is laced with CTAs for a paid community, so expect upsell noise

Verdict

Worth a look if you are already living in Claude Code and want overnight agents that can actually verify and ship their own changes. Everyone else should wait for a more agnostic standard to emerge.