Claude Skills MCP Server

Claude Skills MCP Server is an MCP server that enables intelligent search and retrieval of Claude Agent Skills using vector embeddings and semantic similarity. It implements a progressive disclosure architecture so AI applications can discover and load skills in stages (metadata → full content → files) while remaining fast and local. The server can load skills from multiple sources, including Official Anthropic Skills, K-Dense AI Scientific Skills, and local directories, providing a zero-configuration experience out of the box for Cursor or standalone usage. The architecture is split into a lightweight frontend and a heavy backend, enabling instant startup and background backend download, with no API keys required and the ability to connect to remote hosted backends if desired.

Use this MCP server to search, discover, and retrieve Claude Agent Skills from multiple sources with fast, local vector search. The server exposes three tools for working with Claude Agent Skills: find_helpful_skills (semantic search based on a task description), read_skill_document (retrieve specific files from a skill), and list_skills (enumerate all loaded skills for exploration/debugging). The frontend-backend two-package architecture provides a lightweight startup with an intensive backend that handles vector search. You can run it via Cursor or standalone with uvx. Example usage from the docs:

{
  "mcpServers": {
    "claude-skills": {
      "command": "uvx",
      "args": ["claude-skills-mcp"]
    }
  }
}

uvx claude-skills-mcp

<h1 class="text-2xl font-semibold mt-5 mb-3">1. Print the default configuration</h1>
uvx claude-skills-mcp --example-config > config.json
<h1 class="text-2xl font-semibold mt-5 mb-3">2. Edit config.json to your needs</h1>
<h1 class="text-2xl font-semibold mt-5 mb-3">3. Run with your custom configuration</h1>
uvx claude-skills-mcp --config config.json

This MCP server loads a curated set of skills by default (e.g., Anthropic Official Skills and K-Dense AI Scientific Skills) and can also load from a local directory (~/.claude/skills if it exists), enabling fast, private skill discovery without external API calls.