iflow-mcp_k-dense-ai-claude-skills-mcp 1.0.6

Lightweight MCP proxy for Claude Skills (auto-downloads backend)

Claude Skills MCP Frontend

Lightweight MCP proxy for Claude Skills that auto-downloads the heavy backend on demand.

Overview

This is the frontend component of the Claude Skills MCP system. It's a lightweight proxy (~15 MB) that:

• Starts instantly (<5 seconds)
• Auto-downloads the backend when first needed
• Acts as MCP server (stdio) for Cursor
• Acts as MCP client (HTTP) for the backend
• Returns tool schemas immediately (no backend wait needed)

Installation

# Via uvx (recommended for Cursor)
uvx claude-skills-mcp

# Via uv tool (persistent install)
uv tool install claude-skills-mcp

# Via pip
pip install claude-skills-mcp

Usage with Cursor

Add to your Cursor MCP settings (~/.cursor/mcp.json):

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

Restart Cursor and the skills will be available!

First Run Behavior

On first run, the frontend will:

1. Start immediately (~5 seconds) ✅ Cursor timeout satisfied!
2. Return tool schemas to Cursor (instant)
3. Download backend in background (~250 MB, 60-120 seconds)
4. When you first use a tool, you'll see "Loading backend..."
5. Once backend ready, all tools work normally

Subsequent runs: Fast! Backend is already installed.

Configuration

The frontend forwards all arguments to the backend:

# Custom configuration
uvx claude-skills-mcp --config my-config.json

# Verbose logging
uvx claude-skills-mcp --verbose

# Custom backend port (advanced)
uvx claude-skills-mcp --port 9000

Remote Backend (Future)

# Connect to hosted backend instead of local
uvx claude-skills-mcp --remote https://skills.k-dense.ai/mcp

Note: Remote backend support coming in v1.1.0

How It Works

Cursor → Frontend (stdio, ~15 MB)
           ↓
         list_tools() → Returns hardcoded schemas INSTANTLY ✅
           ↓
         [Backend downloads in background...]
           ↓
         call_tool() → Proxies to Backend (HTTP)
           ↓
         Backend (HTTP, ~250 MB) → Performs actual search

This architecture solves the Cursor timeout problem by separating:

• Fast startup (frontend, minimal dependencies)
• Heavy processing (backend, downloads async)

Dependencies

Frontend only requires:

• mcp>=1.0.0 (~5 MB)
• httpx>=0.24.0 (~5 MB)

Total: ~15 MB (downloads in <10 seconds)

The backend (claude-skills-mcp-backend) is auto-installed on first use.

Troubleshooting

"Backend not ready" message

On first run, you'll see this message for 30-120 seconds while the backend downloads. This is normal and only happens once.

Backend installation fails

Check:

1. Internet connection
2. Disk space (~500 MB free needed)
3. Python 3.12 installed

Tools not working

Run with verbose logging:

uvx claude-skills-mcp --verbose

Check logs in stderr for backend status.

Development

# Clone the monorepo
git clone https://github.com/K-Dense-AI/claude-skills-mcp.git
cd claude-skills-mcp/packages/frontend

# Install in development mode
uv pip install -e ".[test]"

# Run tests
uv run pytest tests/

Related Packages

• claude-skills-mcp-backend (Backend): Heavy server with vector search
• Main Repository: https://github.com/K-Dense-AI/claude-skills-mcp

License

Apache License 2.0

Copyright 2025 K-Dense AI (https://k-dense.ai)