Claude Skills

The Claude Skills MCP Server enables any MCP clients, like Cursor AI, to search, discover, and utilize Claude Skills through semantic search and progressive loading.

Features

• ⚡️ Two-Package Architecture: A lightweight frontend starts instantly to avoid editor timeouts, while the heavier backend downloads in the background.
• 🧠 Semantic Search: Uses vector embeddings for intelligent and relevant skill discovery.
• 📂 Multi-Source Loading: Pulls skills from GitHub repositories (like the official Anthropic and K-Dense AI collections) and local directories.
• 🔌 Zero Configuration: Works right out of the box with a curated set of skills ready to go.
• ⚙️ Highly Configurable: You can customize skill sources, embedding models, and content limits to fit your needs.
• 🔒 Fast and Local: Operates entirely on your local machine with no need for API keys, and it automatically caches skills from GitHub for speed.

How To Use It

For Cursor Users

Add the MCP server through the Cursor Directory. Alternatively, you can manually add it to your ~/.cursor/mcp.json configuration file:

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

Standalone Usage with uvx

If you’re not using Cursor or want to run the server on its own, you can execute it directly from your terminal:

uvx claude-skills-mcp

Custom Configuration

For more advanced setups, you can customize the server’s behavior.

1. Generate the default configuration file:
   bash uvx claude-skills-mcp --example-config > config.json
2. Modify the config.json file to add your own local or GitHub skill sources, change embedding models, or adjust other settings.
3. Run the server with your custom configuration:
   bash uvx claude-skills-mcp --config config.json

Available Tools

• find_helpful_skills: Performs a semantic search to find skills relevant to your task description.
• read_skill_document: Retrieves specific files from a skill, such as scripts, data files, or reference materials.
• list_skills: Shows a complete list of all skills the server has loaded.

FAQs

Q: Why does the first launch take longer than subsequent uses?
A: The initial setup downloads the backend package (~250 MB) with PyTorch and sentence-transformers for vector search. This happens once, and all future sessions use the cached installation.

Q: Can I add my own custom skills or repositories?
A: Yes, create a configuration file that includes additional GitHub repositories or local directories. The system supports any repository following the Claude Skills format or containing Claude Code plugins.

Q: How many skills are available by default?
A: The default configuration loads approximately 90 skills: 15 from Anthropic’s official repository and 78+ from K-Dense AI’s scientific skills collection, plus any skills in your local ~/.claude/skills directory.

Q: Can I run this without uvx?
A: The backend package can be installed separately using pip install claude-skills-mcp-backend for deployment in custom environments or hosted setups.

Q: Does this server send my code or prompts to an external service?
A: No. The entire process runs locally on your machine. It fetches public skills from GitHub and caches them, but the semantic search and all interactions with your AI assistant happen locally. No API keys are required for its core functionality.

Q: How is this different from just using Claude’s native skills?
A: The main difference is portability. Claude’s native skills only work with Claude. This MCP server makes the entire skills framework—including official and custom skills—available to any AI model or assistant that supports the Model Context Protocol.