⚠️ Warning: Experimental proof‑of‑concept. Potentially unsafe. Treat skills like untrusted code. Review them, sandbox execution, and use at your own risk.
Skillz is a tiny MCP server that exposes Anthropic‑style skills (folders with SKILL.md) as tools for any MCP client, not just Claude. It returns the skill’s instructions and absolute paths to its files, and can optionally run a script bundled with the skill.
Why does it matter
Skills are a promising new format for loading context dynamically and enabling agents to employ the most relevant instructions for a task, without bloating the context on load. They are currently not implemented by any tool other than the Claude suite of tools from Anthropic, limiting their use across the entire agentic AI landscape.
A shim that allows skills to be loaded into agents other than Claude opens up the possibility of using skills as a general technology, in the hope that they will eventually be standardised and supported more widely.
What it does
Recursively finds every
SKILL.mdunder a root directory and registers one tool per skill (tool name = skill slug, e.g.,algorithmic-art).Tool description comes from the skill’s front‑matter
description.Calling a tool with a
taskreturns:instructions(fromSKILL.md) andresources(absolute file paths). Optionally run a helper script viascript+script_payload.
Quick start
Run:
uv run skillz.py /path/to/skillsPoint your MCP client at the server (stdio/HTTP/SSE).
Safety
Experimental PoC; APIs and outputs may change.
Scripts may run arbitrary code. Isolate with containers/VMs; do not run untrusted skills on sensitive machines.
Repo: github.com/intellectronica/skillz
