What is it?
What it is
A file-based harness with 49 small, trigger-specific skills for coding agents, enforcing a Plan → Act → Verify loop via disk operations. Skills load only when relevant, enabling specialization without framework overhead.
Why it exists
To provide a disciplined, file-driven workflow for agents, preventing premature success and ensuring reproducibility. It emphasizes minimalism, cross-agent compatibility, and zero runtime state.
Who should use it
AI coding agents (Claude Code, Cursor, Codex, Gemini, Windsurf), developers using autonomous LLM workflows, and engineers building multi-session agentic automation.
Who should avoid it
Users looking for a rigid methodology or framework, those wanting a heavy runtime/daemon-based CLI, and developers who prefer manual, non-automated agent management.
How it works
A quick walkthrough in plain English
How loopkit works
Step 1 of 3
You interact with it
Open loopkit, send a request, or connect it to your stack.
Features
Advantages
- No methodology commitment or heavy framework overhead
- Extremely lightweight and portable (plain markdown files)
- Prevents 'premature victory' through adversarial verification
- Avoids vendor lock-in via agent-agnostic skill format
- Enforces discipline without owning the entire process
- Easy to cherry-pick individual skills via npx
Disadvantages
- Harness components (settings.json, hooks) are optimized for Claude-Code
- Requires manual pruning of MEMORY.md to prevent context rot
- Requires user to manage the loop via shell execution
Installation
FAQ
How do I install loopkit into my existing project?
You can use the curl installer: `curl -fsSL https://raw.githubusercontent.com/Archive228/loopkit/main/install.sh | bash`. To prevent overwriting your current `.claude/` directory, use `BACKUP=1` to create a timestamped snapshot or `FORCE=1` to overwrite existing files.
Is loopkit compatible with agents other than Claude Code?
Yes. While the bundled harness (settings, hooks, verifier) is optimized for Claude Code, the 49 skills are plain markdown files with YAML headers. They are compatible with any agent that can read `.claude/skills/`, such as Cursor, Codex, Gemini CLI, Windsurf, or GitHub Copilot CLI.
Can I use individual skills without installing the entire harness?
Yes. You can cherry-pick a single skill using npx: `npx -p claude-loopkit loopkit-add <skill-name>`. This will install only that specific skill into your `./.claude/skills/` directory.
What is the purpose of the 'erifier' subagent?
The verifier is an adversarial subagent (designed to run on Haiku) that reads code diffs to identify 'fake done' shortcuts. It acts as a check to ensure the agent hasn't bypassed necessary steps before claiming a task is complete.
Does loopkit require a specific coding methodology or runtime?
No. loopkit is a harness and a library of skills, not a methodology like BMAD or Prime Radiant. It has no runtime state, no daemon, and no server; it relies on files on disk and a simple 8-line `run.sh` script to drive the Plan → Act → Verify loop.
Featured in Videos
YouTube tutorials and walkthroughs for loopkit
Alternatives
Similar projects ranked by category, topics, and text overlap.
