Curated public registry of AI agent harnesses for Claude Code
harness-lab is a curated public registry of AI agent harnesses (skills, commands, agents, hooks) for Claude Code. It includes a zero-dependency Node.js CLI for browsing, installing, and managing harnesses from GitHub directly into .claude/ directories.
- Browse harnesses by category (Product, Software Engineering, Design, Harness Engineering)
- Install skills, commands, agents, and hooks with a single command
- Three installation scopes: global, project, or local
- Always fetches the latest versions from GitHub — no local cache
Claude Code has a native plugin system — so why does harness-lab distribute artifacts as standalone .claude/ files instead?
Practical install/uninstall workflow. The CLI lets you browse, install, and remove artifacts from any terminal in seconds (harness-lab install ... / harness-lab uninstall ...). The init command bootstraps slash commands so you can also manage everything from inside Claude Code (/harness-lab:install). No marketplace setup, no manifest files — just files copied in and out of .claude/.
Granularity and cohesion. Harnesses tend to have higher granularity and cohesion than plugins. Each harness groups artifacts around a single, well-defined purpose (e.g. PRD generation), while plugins tend to bundle loosely related capabilities together. This matters because tighter cohesion means less irrelevant context loaded per task — and with harness-lab, each artifact is an independent file you can install or remove individually.
Persistent ghost errors. Plugins can leave behind errors that are difficult to trace and persist even after the plugin is uninstalled or disabled. The screenshot below shows a /doctor output reporting "Plugin ralph-wiggum not found in marketplace" — an error that lingers with no clear way to resolve it. File-based artifacts don't have this problem: uninstalling simply deletes the files, leaving no residual state.
Plugins are great for MCP integrations, LSP servers, and hook-heavy workflows. But for a curated registry of prompt-based artifacts, the file-based approach is simpler and more efficient.
| Harness | Description | Author |
|---|---|---|
| prd-generator | Context engineering PRD generator for AI-driven software development | hgflima |
| agent-md-writer | Generate and maintain CLAUDE.md / AGENTS.md files with Karpathy principles, context7-first docs, DESIGN.md consultation, and parallelized /agent-md:create, /agent-md:update, /agent-md:audit commands |
hgflima |
| rules-generator | Analyze a project's stack and generate tailored .claude/rules/ files | hgflima |
| github-readme-writer | Generate comprehensive, professional README.md files for any project | alfredang |
| product-vision-frame | Define a product's singular purpose and ruthlessly align all decisions to that vision | hgflima |
| heal | Self-healing post-incident harness — diagnoses root causes after a failed Claude Code session, applies approved fixes to skills/rules/CLAUDE.md, and persists structured changelogs to prevent recurrence. Includes /heal:migrate for upgrading from legacy lessons-learned installs |
hgflima |
| product-brief | Interactive wizard for creating structured product briefs across project levels (spike, POC, MVP, feature, product, platform) | hgflima |
| ui-design | Service-agnostic UI/UX specification system — design tokens, screen specs, component definitions, and multi-tool export | hgflima |
| parallel-plan-runner | Execute an implementation plan via Claude Code's experimental Agent Teams — parallelization analysis, teammate spawn, orchestration, and audit | hgflima |
| user-journey-map | Progressive-extraction skill for user journey maps with provenance marks, Mermaid diagrams, validation checklist, and explicit PRD boundary — input for PRD generation | hgflima |
| Category | Technology |
|---|---|
| Language | JavaScript (ESM) |
| Runtime | Node.js >= 18 |
| Dependencies | Zero — Node.js built-ins only |
| Data Source | GitHub Raw + GitHub REST API |
| Distribution | npm / npx |
┌─────────────────────────────────────────────────┐
│ User Layer │
│ │
│ npx harness-lab /harness-lab:install │
│ (Terminal CLI) (Slash Commands) │
└────────────┬──────────────────────┬─────────────┘
│ │
▼ ▼
┌───────────────────────────────────────────────────┐
│ CLI Layer (src/cli/) │
│ │
│ categories · list · install · uninstall · update│
└────────────────────────┬──────────────────────────┘
│
▼
┌──────────────────────────────────────────────────┐
│ Core Layer (src/core/) │
│ │
│ catalog.js installer.js scope.js │
│ (fetch index) (download/remove) (resolve path)│
└────────────────────────┬─────────────────────────┘
│
▼
┌───────────────────────────────────────────────────┐
│ GitHub (Remote Data) │
│ │
│ raw.githubusercontent.com api.github.com │
│ (catalog.json, artifacts) (directory listings)│
└───────────────────────────────────────────────────┘
harness-lab/
├── bin/cli.js — CLI entry point (shebang, command router)
├── src/
│ ├── core/ — Business logic (shared by CLI + slash commands)
│ │ ├── catalog.js — Fetch catalog.json and harness.json from GitHub
│ │ ├── installer.js — Download artifacts to .claude/, remove on uninstall
│ │ └── scope.js — Resolve target path (global/project/local)
│ ├── cli/ — Command handlers (parse args, call core, format output)
│ │ ├── categories.js, list.js, install.js, uninstall.js, update.js
│ └── setup/
│ └── init.js — npx bootstrap (npm install -g + copy slash commands)
├── templates/slash-commands/
│ └── harness-lab/ — 5 .md files → .claude/commands/harness-lab/
├── harnesses/ — The curated harness collection
│ └── <name>/
│ ├── harness.json — Metadata: name, version, categories, artifact list
│ ├── skills/ — Skill files (SKILL.md, prompt.md, references/, scripts/)
│ ├── commands/ — Slash command files (<name>.md)
│ ├── agents/ — Subagent files (<name>.md)
│ ├── rules/ — Rule files (<name>.md)
│ └── hooks/ — Hook scripts and config
├── catalog.json — Registry index (categories + harness entries)
└── package.json
- Node.js >= 18.0.0
npm install -g harness-lab@latestThen install the slash commands for Claude Code:
harness-lab initThe init wizard asks where to place the slash commands (global, project, or local) and copies them into your .claude/commands/ directory.
harness-lab # Run init (install slash commands)
harness-lab categories # List available categories
harness-lab list [category] # List harnesses, optionally by category
harness-lab install <name> # Install a harness (default scope: project)
harness-lab uninstall <name> # Uninstall a harness
harness-lab update [name] # Update one or all harnesses
harness-lab --version # Show version
harness-lab --help # Show helpUse --scope to control where artifacts are installed:
harness-lab install prd-generator --scope global # ~/.claude/
harness-lab install prd-generator --scope project # .claude/ (default)
harness-lab install prd-generator --scope local # .claude/local/After running harness-lab init, these commands are available inside Claude Code:
/harness-lab:categories List available categories
/harness-lab:list List harnesses
/harness-lab:install Install a harness
/harness-lab:uninstall Uninstall a harness
/harness-lab:update Update harnesses
If you prefer not to install globally, use npx to run any command directly:
# Install globally + slash commands
npx harness-lab@latest
# Browse and install
npx harness-lab@latest categories
npx harness-lab@latest list software-engineering
npx harness-lab@latest install prd-generator --scope project
npx harness-lab@latest uninstall prd-generator
npx harness-lab@latest update- Fork the repository
- Create your feature branch (
git checkout -b feat/my-feature) - Commit your changes (
git commit -m 'Add my feature') - Push to the branch (
git push origin feat/my-feature) - Open a Pull Request
For questions and discussions, visit GitHub Issues.
- Anthropic for Claude Code
- alfredang for the github-readme-writer harness
- sethmblack for the product-vision-frame harness
- All contributors who submit harnesses to the registry
If you find this useful, please give it a star!