claude-skills-mcp

Purpose

claude-skills-mcp is a server based on the Model Context Protocol (MCP), designed to let any AI model or programming assistant that supports MCP (including Cursor, Codex, GPT-5, Gemini, etc.) search for and invoke Claude's Agent Skills framework, making these skills available as if they were natively supported. It brings Anthropic's Agent Skills capabilities into the entire AI ecosystem through the MCP protocol.

Main Capabilities

• Semantic Search: Uses vector embeddings and semantic similarity to intelligently discover relevant skills based on task descriptions.
• Progressive Disclosure: Implements the multi-level loading architecture described in Anthropic's Agent Skills engineering blog (metadata → full content → files).
• Multi-Source Loading: Can load skills from GitHub repositories (direct skills or Claude Code plugins) or local directories. Default sources include Anthropic's official skills library, the K-Dense scientific Agent skills library (148+ specialized skills for bioinformatics, cheminformatics, and scientific analysis), and the local ~/.claude/skills directory (if present).
• Automatic Updates: Checks skill sources every hour and automatically reindexes when changes occur.
• Local and Fast: No API key required, with automatic GitHub caching.
• Configurable: Sources, models, and content limits can be customized.

Usage

The project provides three core tools:

• find_helpful_skills: Performs semantic search based on task descriptions to find relevant skills.
• read_skill_document: Retrieves specific files from skills (scripts, data, reference documents).
• list_skills: Views the complete inventory of loaded skills (for exploration or debugging).

Dual-Package Architecture

• Frontend (claude-skills-mcp): Lightweight proxy (about 15 MB), starts instantly (<5 seconds), avoids Cursor timeouts; automatically downloads the backend on first use; used by Cursor as an MCP server (stdio).
• Backend (claude-skills-mcp-backend): Heavyweight server (about 250 MB), uses PyTorch and sentence-transformers for vector search; serves as an MCP server (streamable HTTP); automatically installed by the frontend, and can also be deployed independently. The backend downloads in the background (about 60-120 seconds, one-time), without affecting Cursor responsiveness; it is instantly available for subsequent use. It can also connect to a remotely hosted backend, with no local installation required.

Installation and Configuration

• Cursor Users: Can add it through the Cursor Directory, or add it to the Cursor configuration (~/.cursor/mcp.json).
• Using uvx (Standalone): Run uvx claude-skills-mcp to start the server with the default configuration, loading 160+ skills. Custom configurations are also supported.

Use Cases

Suitable for developers who want to reuse Claude Agent Skills in non-Claude AI programming assistants or models, especially in scenarios requiring specialized skills such as scientific computing, document processing, presentations, and web artifacts.

Limitations and Current Status

The project is no longer hosted or maintained, and new issues and pull requests may not be reviewed. The reason is that Agent Skills have been natively adopted by all major AI platforms (Cursor, Windsurf, Claude Code, Copilot, etc.), and an MCP bridge is no longer needed to deliver skills to coding assistants. The final version, v1.1.0, is still installable from PyPI. If you want to continue development on this basis, you can fork the repository.

License

This project is open source under the Apache License 2.0, with copyright belonging to K-Dense (https://www.k-dense.ai）所有（2025-2026）。