🧰 skillkit

The toolbox for AI agent skills. Scaffold, lint, pack, and install SKILL.md skills — and serve your whole skill library to Claude Code, Cursor, and any MCP client with a built-in Model Context Protocol server.

Community audits found that the overwhelming majority of published SKILL.md files carry at least one "skill smell" — and a large share leak secrets. skillkit catches those before you ship.

📖 Deep docs: Usage guide — lint rules reference, MCP patterns, CI usage · Architecture · Contributing · Changelog · Security

✨ Features

• 🆕 new — scaffold a spec-compliant skill folder in one command
• 🔍 lint — validate against the Agent Skills spec plus security smells: leaked API keys, vague descriptions, oversized bodies, folder typos, and more (scored 0–100 with a grade)
• 📦 pack — zip a skill for upload to skill-capable platforms
• 📥 install — install from a folder or any git URL into ~/.claude/skills (or your own directory)
• 🗂️ list / remove — manage your installed skill library
• 🔌 mcp — a zero-dependency MCP server (stdio, JSON-RPC 2.0) that exposes list_skills, read_skill, and lint_skill tools to any client
• 🪶 Zero dependencies — pure Python standard library
• 🧪 76 tests, CI on Python 3.9 → 3.12

🚀 Quick start

git clone https://github.com/furkan708/skillkit.git
cd skillkit && pip install .

# 1. create a skill
skillkit new commit-writer -d "Writes conventional commit messages from staged diffs. Use when the user asks to commit changes."

# 2. lint it (spec + security + quality)
skillkit lint commit-writer
#   ✓ no issues found
#   score: 100/100 (grade A) · 0 errors · 0 warnings · 0 notes

# 3. pack it for upload
skillkit pack commit-writer        # → commit-writer.zip

# 4. install it for Claude Code
skillkit install ./commit-writer   # → ~/.claude/skills/commit-writer

🔌 Serve your skills over MCP

Add skillkit to any MCP client config — Claude Code, Cursor, Windsurf, and every other MCP-compatible agent:

{
  "mcpServers": {
    "skillkit": { "command": "skillkit", "args": ["mcp"] }
  }
}

Your agent can now discover and read every installed skill on demand:

🔍 What lint catches

Every run ends with a 0–100 score and a letter grade, so you know when a skill is ready to publish.

📖 CLI reference

skillkit new <name> -d <description> [--dir DIR]   scaffold a skill
skillkit lint <path> [--json] [--strict]           validate a skill
skillkit pack <path> [-o FILE.zip]                 zip for upload
skillkit install <path|git-url> [--agent claude|project] [--dir DIR]
skillkit list [--agent ...] [--dir DIR] [--json]   show installed skills
skillkit remove <name> [--dir DIR]                 uninstall a skill
skillkit mcp [--dir DIR]                           serve skills over MCP

Built on the open Agent Skills specification — the same format supported by 40+ platforms including Claude, OpenAI Codex, and GitHub Copilot.

🧪 Tests

pip install pytest
pytest -v

🗂️ Project Structure

skillkit/
├── skillkit/
│   ├── frontmatter.py   # minimal YAML frontmatter parser
│   ├── model.py         # skill loading (Agent Skills spec)
│   ├── linter.py        # rules, security smells, 0-100 score
│   ├── scaffold.py      # new + pack
│   ├── installer.py     # install / list / remove (folder or git)
│   ├── mcp_server.py    # zero-dependency MCP stdio server
│   └── cli.py           # command-line interface
└── tests/

🗺️ Roadmap

• skillkit search — search community skill registries
• More install targets (Codex, Copilot CLI paths as they standardize)
• skillkit doctor — prompt-injection heuristics (planned, not implemented yet)

📄 License

MIT — see the LICENSE file for details.