Skillz: Use Claude Skills in Codex, Copilot, or Any Other Agent via MCP

⚠️ Experimental proof‑of‑concept. Potentially unsafe. Treat skills like untrusted code and run in sandboxes/containers. Use at your own risk.

skillz.py is a single‑file MCP server that exposes Anthropic‑style skills (directories with a SKILL.md that starts with YAML front‑matter) to any MCP client. It recursively discovers skills, registers one tool per skill, returns authored instructions and absolute file paths, and can optionally run a helper script from the skill in a temporary workspace.

• Python 3.12 or newer (managed automatically when using uv)
• uv package manager (the script metadata declares runtime dependencies)

1. Populate a directory with skills following Anthropic’s format (SKILL.md + optional resources).

2. Run the server with the directory path:

   uv run skillz.py /path/to/skills --verbose

   By default the server listens over stdio. Pass --transport http or --transport sse with --host, --port, and --path if your client needs a network transport.

3. Use --list-skills to validate parsing without starting the transport:

   uv run skillz.py /path/to/skills --list-skills

• Recursively walks the skills root and loads every SKILL.md (nesting supported).
• One MCP tool is registered per skill. Tool name = the slug of name (e.g., algorithmic-art).
• Tool description = the description from front‑matter (no extra metadata included).

Note: Skillz returns absolute paths but does not register MCP resources for file reads. Clients should read files directly from disk or provide their own file‑access flow.

• This code is experimental, untested, and should be treated as unsafe.
• Script execution happens outside any sandbox besides a temporary directory. Use only with trusted skill content and within controlled environments.
• Review and harden before exposing to real users or sensitive workflows.

Released under the MIT License (see LICENSE).