MCP Directory

Claude Skills MCP Server

Semantic search and progressive loading of Claude Agent Skills for any MCP-compatible AI assistant.

Unverified
stdio (local)
No auth
Python

Add to your client

Copy the config for your MCP client and paste it into its config file.

Install / run
uvx claude-skills-mcp

Paste into ~/Library/Application Support/Claude/claude_desktop_config.json

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

Requires `uv` (the Python package runner). Install it from https://docs.astral.sh/uv/ if `uvx` is not found.

Step-by-step guides: Add to Claude Desktop · Add to Cursor · Add to Windsurf

Before you start

  • uv / uvx installed (Astral) to run the package
  • Python 3.12
  • Internet access on first run to download the backend (~250 MB) and fetch skill sources from GitHub
  • An MCP-compatible client such as Cursor

About Claude Skills MCP Server

An MCP server that lets any MCP-compatible AI assistant discover and retrieve Claude Agent Skills via semantic vector search. It exposes three tools for searching, reading, and listing skills, and uses a frontend/backend split so the MCP client (e.g. Cursor) gets an instant-starting stdio proxy while the heavy vector-search backend downloads and runs in the background. Ships with ~90 curated skills by default and supports adding skills from GitHub repos or local directories.

Tools & capabilities (3)

find_helpful_skills

Semantic search for relevant skills based on a task description, using vector embeddings and similarity matching.

read_skill_document

Retrieve specific files (scripts, data, references) from a skill, supporting progressive disclosure of skill content.

list_skills

View the complete inventory of all loaded skills, intended for exploration and debugging.

When to use it

  • Give a non-Claude AI assistant (Cursor, Codex, GPT-5, Gemini) access to Anthropic's Agent Skills framework
  • Semantically discover the most relevant skill for a given coding or analysis task instead of manually browsing a skills list
  • Pull specific reference files or scripts from a skill on demand via progressive disclosure
  • Surface specialized scientific skills (bioinformatics, cheminformatics, scientific analysis) from K-Dense AI's collection inside an MCP client
  • Load and search custom skills from your own GitHub repositories or a local ~/.claude/skills directory

Security notes

No API keys required; runs locally over stdio. On first use the lightweight frontend auto-downloads and starts a ~250 MB backend (PyTorch + sentence-transformers) from PyPI in the background, which can take ~60-120 seconds one time. Skills are fetched from public GitHub repositories and cached locally. The README states the server is no longer hosted or maintained.

Claude Skills MCP Server FAQ

Do I need an API key?

No. The server runs locally and requires no API keys. Skills are fetched from public GitHub repositories and cached locally.

Why is there a delay on first use?

The lightweight frontend (~15 MB) starts in under 5 seconds and immediately exposes the tools, then auto-downloads the ~250 MB backend (with PyTorch and sentence-transformers) in the background, which takes about 60-120 seconds the first time. Subsequent uses are instant.

Which skills are loaded by default?

About 90 skills: Anthropic's official skills (~15) and K-Dense AI's scientific skills (78+). You can also configure additional GitHub repositories or local directories such as ~/.claude/skills.

How do I customize the configuration?

Run `uvx claude-skills-mcp --example-config > config.json`, edit the file, then run `uvx claude-skills-mcp --config config.json`.

Is this server still maintained?

No. The README states the server is no longer hosted or maintained, because major AI platforms now support Agent Skills natively, removing the need for an MCP bridge. It remains available as an open-source project.

Alternatives to Claude Skills MCP Server

Compare all alternatives →

Official MCP reference server that fetches a URL and returns its content as clean Markdown, with chunking.

Verified
stdio (local)
No auth
Python
2 tools
Updated 12 days agoRepo

Official Firecrawl MCP server — scrape, crawl, map, search, and structured extraction for any LLM client.

Verified
stdio (local)
API key
TypeScript
8 tools
Updated 10 days agoRepo

Exa's neural web search and crawling MCP server — runs locally via npx or as a hosted remote endpoint.

Verified
stdio (local)
API key
TypeScript
6 tools
Updated 11 days agoRepo