Skyforge MCP Server

⚠️ ALPHA RELEASE - This is an early alpha version. Expect bugs and breaking changes.

🚫 NOT FOR PRODUCTION - This is a development/experimental version. For a production implementation, please contact james@skyforge-labs.com

🔓 NO AUTHENTICATION - This server has no built-in authentication. CORS is wide open (allow_origins=["*"]). Use at your own risk and secure your deployment appropriately.

A Model Context Protocol (MCP) server that connects AI assistants to SkySpark and Haxall building automation systems. Dynamically exposes your SkySpark Axon functions as MCP tools.

Features

• Dynamic Axon Tools - Fetches tool definitions from SkySpark at runtime
• Prompt Support - Expose templated prompts from SkySpark
• Dual Transport - Supports stdio (Claude Desktop) and HTTP/SSE (web clients)
• Type Safety - Full Haystack type system with automatic JSON Schema conversion
• Docker Ready - Simple Docker deployment included

How It Works

The server fetches tools from SkySpark on each list_tools request. This means:

• Add new tools by creating Axon functions in SkySpark
• No server restart needed for schema changes
• SkySpark is your single source of truth

Quick Start

Prerequisites

• SkySpark or Haxall server with API access
• Docker (recommended) OR Python 3.12+ with uv

Quick Setup with Example Tools

For immediate testing, import the included setup.zinc file into your SkySpark project. This provides example MCP tools and the required fetchMcpTools() function.

Docker Setup (Easiest)

1. Clone and configure

   git clone https://github.com/yourusername/skyforge-mcp.git
   cd skyforge-mcp
   
   # Create .env file
   cat > .env << EOF
   SKYSPARK_URI=http://host.docker.internal:8080/api/demo
   SKYSPARK_USERNAME=your_username
   SKYSPARK_PASSWORD=your_password
   EOF
   

2. Start server

   docker-compose up --build
   

   Server runs on http://localhost:8000/mcp

3. Test with MCP Inspector

   npx @modelcontextprotocol/inspector docker exec -it skyspark-mcp-server uv run main.py
   

Local Setup (Development)

1. Install and run

   # Install uv package manager
   curl -LsSf https://astral.sh/uv/install.sh | sh
   
   # Clone and setup
   git clone https://github.com/yourusername/skyforge-mcp.git
   cd skyforge-mcp
   uv sync
   
   # Create .env (same as above)
   
   # Run stdio mode (for Claude Desktop)
   uv run main.py
   
   # OR run HTTP/SSE mode (for web clients)
   uv run uvicorn main:app --host 0.0.0.0 --port 8000
   

Claude Desktop Integration

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):

Docker:

{
  "mcpServers": {
    "skyforge": {
      "command": "docker",
      "args": ["exec", "-i", "skyspark-mcp-server", "uv", "run", "main.py"]
    }
  }
}

Local:

{
  "mcpServers": {
    "skyforge": {
      "command": "uv",
      "args": ["run", "main.py"],
      "cwd": "/path/to/skyforge-mcp",
      "env": {
        "SKYSPARK_URI": "http://localhost:8080/api/demo",
        "SKYSPARK_USERNAME": "your_username",
        "SKYSPARK_PASSWORD": "your_password"
      }
    }
  }
}

Restart Claude Desktop after saving.

Creating SkySpark Tools

In SkySpark, implement fetchMcpTools() to return tool definitions as a grid. Each row should have:

• name - Tool identifier (Str)
• dis - Display name (Str)
• help - Description (Str)
• params - Parameter schema (Dict or List)

Example in SkySpark:

// Return MCP tools grid
fetchMcpTools: () => [
  {
    name: "getSiteEquips",
    dis: "Get Site Equipment", 
    help: "Returns all equipment for a site",
    params: {
      kind: "Dict",
      params: {
        siteId: {
          kind: "Ref",
          help: "Site reference ID",
          required: marker()
        }
      }
    }
  }
].toGrid

// Tool implementation (called via `call()`)
getSiteEquips: (dict) => readAll(equip and siteRef == dict->siteId)

Import the included setup.zinc file into your SkySpark project for example tools and the required fetchMcpTools() function.

The server fetches tools automatically when clients call list_tools.

Configuration

Create .env file:

# For Docker: use host.docker.internal to access host machine
SKYSPARK_URI=http://host.docker.internal:8080/api/demo
# For local development: use localhost
# SKYSPARK_URI=http://localhost:8080/api/demo
SKYSPARK_USERNAME=your_username
SKYSPARK_PASSWORD=your_password

All three variables are required.

Project Structure

skyforge-mcp/
├── app/
│   ├── skyspark/         # SkySpark integration
│   │   ├── client.py     # Phable-based API client
│   │   ├── converters.py # Haystack ↔ JSON Schema conversion
│   │   ├── grid.py       # HGrid wrapper for dual format output
│   │   └── types.py      # Extended Haystack types
│   └── tools/
│       └── axon_tools.py # Hardcoded tool examples
├── main.py              # MCP server entry point
├── docker-compose.yml   # Docker setup
└── Dockerfile           # Container definition

Troubleshooting

Connection errors:

• Docker: Use host.docker.internal instead of localhost in SKYSPARK_URI
• Verify SkySpark URI is accessible: curl http://your-server:8080/api/demo
• Check .env credentials
• Ensure SkySpark API is enabled

No tools appearing:

• Verify fetchMcpTools() function exists in SkySpark
• Check server logs: docker-compose logs or uv run main.py
• Test with MCP Inspector

Docker issues:

docker-compose logs              # View logs
docker-compose restart           # Restart
docker-compose up --build        # Rebuild

Security Notes

⚠️ Important:

• This is NOT for production use - if you are interested in a production implementation, contact james@skyforge-labs.com
• No built-in authentication - secure your network/deployment
• CORS allows all origins - intended for local development
• Store credentials securely (.env files, environment variables)
• For production, add authentication middleware or use VPN/firewall

Credits & License

Built with:

• MCP Python SDK - Model Context Protocol implementation
• Phable - Haystack/SkySpark client library by Rick Jennings
• Project Haystack - Building automation data standard

License: MIT - see LICENSE file

Contributing

Issues and PRs welcome! This is an alpha release - feedback appreciated.

Repository: GitHub