skill-mcp

A small, lightweight MCP (Model Context Protocol) server in Python that serves SKILL.md skill documents to a tool-capable LLM.

It exposes two tools:

Layout

Skills are organized as one subfolder per skill:

<skills-dir>/
  code-review/
    SKILL.md
  git-workflow/
    SKILL.md

A skill is a folder that contains a SKILL.md file (a lowercase skill.md is also accepted on case-insensitive file systems). Any other file or folder is ignored.

Related MCP server: Skillz

Install

uv is used to run the server in an isolated virtualenv. One dependency is required: the official mcp Python SDK (v2), which brings its usual transitive deps.

uv sync          # create .venv and install mcp

Run

uv run server.py --skills-dir skills

CLI parameters:

Example:

uv run server.py --host 127.0.0.1 --port 8080 --skills-dir skills

The server speaks MCP over streamable HTTP and serves the endpoint at /mcp.

Connect an MCP client

Point any MCP client at:

http://127.0.0.1:8080/mcp

The server sends CORS headers (Access-Control-Allow-Origin: *) and exposes Mcp-Session-Id, so browser-based clients (e.g. the llama.cpp web UI, which connects cross-origin from its own port) work out of the box. The OPTIONS preflight is answered and the real POST/GET/DELETE calls are granted. curl/Node clients work too — they simply skip the preflight.

Example client config (Claude Desktop / Claude Code style):

{
  "mcpServers": {
    "skill-server": {
      "url": "http://127.0.0.1:8080/mcp"
    }
  }
}

Notes

• Skills are read fresh on every call, so you can add, edit, or remove skill folders while the server is running — no restart needed.

• Unknown skill names return an error listing the available skills.

• The server binds 127.0.0.1 by default (localhost only). Pass --host to reach it from elsewhere; there is no authentication, so keep that in mind.