Claude Skills MCP Backend

Heavy backend server for Claude Skills MCP system with vector search capabilities.

Overview

This is the backend component of the Claude Skills MCP system. It provides:

• Vector-based semantic search using sentence-transformers
• Skill indexing and retrieval
• MCP protocol via Streamable HTTP transport
• Background skill loading from GitHub and local sources

Note: This package is typically auto-installed by the frontend (claude-skills-mcp). You only need to install it manually for:

• Remote deployment (hosting your own backend)
• Development and testing
• Standalone usage without the frontend proxy

Installation

# Via uv (recommended)
uv tool install claude-skills-mcp-backend

# Via uvx (one-time use)
uvx claude-skills-mcp-backend

# Via pip
pip install claude-skills-mcp-backend

Usage

Run Standalone Server

# Default (localhost:8765)
claude-skills-mcp-backend

# Custom port
claude-skills-mcp-backend --port 8080

# For remote access
claude-skills-mcp-backend --host 0.0.0.0 --port 8080

# With custom configuration
claude-skills-mcp-backend --config my-config.json

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

Configuration

# Print example configuration
claude-skills-mcp-backend --example-config > config.json

# Edit config.json to customize skill sources, embedding model, etc.

# Run with custom config
claude-skills-mcp-backend --config config.json

Endpoints

When running, the backend exposes:

• Streamable HTTP MCP: http://localhost:8765/mcp
• Health Check: http://localhost:8765/health

Docker Deployment

Build Image

docker build -t claude-skills-mcp-backend .

Run Container

# For local access
docker run -p 8765:8765 claude-skills-mcp-backend

# For remote access
docker run -p 8080:8765 \
  -e HOST=0.0.0.0 \
  claude-skills-mcp-backend --host 0.0.0.0 --port 8765

Dependencies

This package includes heavy dependencies (~250 MB):

• PyTorch (CPU-only on Linux): ~150-200 MB
• sentence-transformers: ~50 MB
• numpy, httpx, fastapi, uvicorn: ~30 MB

First download may take 60-180 seconds depending on your internet connection.

Performance

• Startup time: 2-5 seconds (with cached dependencies)
• First search: +2-5 seconds (embedding model download, one-time)
• Query time: <1 second after models loaded
• Memory usage: ~500 MB

Development

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

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

# Run tests
uv run pytest tests/

Related Packages

• claude-skills-mcp (Frontend): Lightweight proxy that auto-installs this backend
• 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)