K-Dense-AI/claude-skills-mcp

What it solves

It allows any AI model or coding assistant (such as Cursor, Gemini, or GPT-5) to use Anthropic's Agent Skills framework. It solves the problem of skill discovery and loading by providing a way to search for and retrieve specialized skills from various sources without causing timeouts in AI editors like Cursor.

How it works

The project uses a two-package architecture consisting of a lightweight frontend proxy and a heavier backend. The frontend starts instantly to avoid timeouts, while the backend handles the heavy lifting: vector embeddings and semantic search using PyTorch and sentence-transformers. It implements a progressive disclosure system where skills are loaded in stages (metadata first, then full content, then files) and can pull skills from GitHub repositories or local directories.

Who it’s for

Developers and AI power users who want to equip their MCP-compatible AI assistants with specialized capabilities, particularly those using scientific skills or official Anthropic skills.

Highlights

• Semantic Search: Uses vector embeddings to intelligently find relevant skills based on a task description.
• Two-Package Design: A lightweight frontend prevents editor timeouts while the backend downloads and indexes skills in the background.
• Multi-Source Loading: Supports loading skills from local directories and GitHub repositories.
• Auto-Updating: Automatically checks skill sources hourly and re-indexes when changes are detected.
• Progressive Disclosure: Efficiently loads skills in levels (metadata $\rightarrow$ content $\rightarrow$ files).