Delia

A Model Context Protocol (MCP) server that cultivates your local LLM garden. Plant a seed, let Delia pick the right vine, and harvest a fresh melon.

Delia - from Greek Δηλία, "from Delos" (the sacred island). Also, she grows watermelons.

Features

• Smart Vine Selection: Routes seeds to the right vine - quick (7B), coder (14B+), moe (30B+), or thinking
• Multi-Garden Support: Ollama, llama.cpp, and Gemini gardens with automatic failover
• Context-Aware Routing: Handles large seeds with appropriate context windows
• Circuit Breaker: Drought protection with graceful recovery
• Parallel Processing: Tends multiple seeds simultaneously
• Authentication: Optional greenhouse access control
• Usage Tracking: Per-gardener quotas and harvest monitoring
• Dashboard: Real-time garden status with watermelon-themed activity feed

Requirements

Hardware

Component	Minimum	Recommended	Large Models
GPU	4GB VRAM	12GB VRAM	24GB+ VRAM
RAM	8GB	16GB	32GB+
Storage	10GB	30GB	50GB+

Software

• Python 3.11+
• uv package manager
• One or more backends:
  • Ollama (recommended)
  • llama.cpp
  • Google Gemini API (optional cloud fallback)

Quick Start

# Clone and install
git clone https://github.com/zbrdc/delia.git
cd delia
uv sync

# Pull models (examples - choose based on your hardware)
ollama pull qwen3:14b           # General purpose
ollama pull qwen2.5-coder:14b   # Code specialized
ollama pull qwen3:30b-a3b       # Complex reasoning

# Run server
uv run python mcp_server.py

Integration

Delia works with AI coding assistants via MCP. Choose your tool:

VS Code / GitHub Copilot

Add to ~/.config/Code/User/mcp.json:

{
  "servers": {
    "delia": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/delia", "python", "mcp_server.py"],
      "type": "stdio"
    }
  }
}

Reload VS Code to activate.

Claude Code

Create ~/.claude/mcp.json:

{
  "mcpServers": {
    "delia": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/delia", "python", "mcp_server.py"]
    }
  }
}

Then run claude and use @delia to delegate tasks.

Gemini CLI

Option 1: HTTP Mode (Recommended)

# Start server
uv run python mcp_server.py --transport sse --port 8200

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "delia": {
      "url": "http://localhost:8200/sse",
      "transport": "sse"
    }
  }
}

Option 2: STDIO Mode

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "delia": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/delia", "python", "mcp_server.py"]
    }
  }
}

GitHub Copilot CLI

Create ~/.copilot-cli/mcp.json:

{
  "servers": {
    "delia": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/delia", "python", "mcp_server.py"]
    }
  }
}

Configuration

Backend Configuration

Edit settings.json in the project root:

{
  "backends": [
    {
      "id": "ollama-local",
      "name": "Ollama Local",
      "provider": "ollama",
      "type": "local",
      "url": "http://localhost:11434",
      "enabled": true,
      "priority": 1,
      "models": {
        "quick": "qwen3:14b",
        "coder": "qwen2.5-coder:14b",
        "moe": "qwen3:30b-a3b",
        "thinking": "deepseek-r1:14b"
      }
    }
  ],
  "routing": {
    "prefer_local": true,
    "fallback_enabled": true
  }
}

Gemini Cloud Backend (Optional)

Add Gemini as a cloud fallback:

# Install dependency
uv add google-generativeai

# Set API key
export GEMINI_API_KEY="your-key-from-aistudio.google.com"

Add to settings.json:

{
  "id": "gemini-cloud",
  "name": "Gemini Cloud",
  "provider": "gemini",
  "type": "remote",
  "url": "https://generativelanguage.googleapis.com",
  "enabled": true,
  "priority": 10,
  "models": {
    "quick": "gemini-2.0-flash",
    "coder": "gemini-2.0-flash",
    "moe": "gemini-2.0-flash"
  }
}

Authentication (Optional)

For HTTP mode with multiple users:

# Quick setup
python setup_auth.py

# Or manually
export DELIA_AUTH_ENABLED=true
export DELIA_JWT_SECRET="your-secure-secret"

Supports username/password and Microsoft 365 OAuth.

Transport Modes

# STDIO (default) - for VS Code, Claude Code, Copilot CLI
uv run python mcp_server.py

# HTTP/SSE - for Gemini CLI, web clients, remote access
uv run python mcp_server.py --transport sse --port 8200

# View all options
uv run python mcp_server.py --help

Tools

Delia provides these MCP tools:

Tool	Description
delegate	Execute tasks with automatic model selection
think	Extended reasoning for complex problems
batch	Process multiple tasks in parallel
health	Check backend status and statistics
models	List available models and tiers
switch_backend	Switch between backends at runtime
switch_model	Change model for a tier
get_model_info	Get model specifications

Vine Selection

Delia picks the right vine for every seed:

Vine	Size	Best For
Quick	7B-14B	Summaries, simple questions
Coder	14B-30B	Generation, review, debugging
MoE	30B+	Architecture, critique, analysis
Thinking	Specialized	Extended reasoning, research

Override with hints in your prompt: "use the large model" or "quick answer".

Troubleshooting

Server won't start

# Check Ollama is running
curl http://localhost:11434/api/tags

# Test server import
uv run python -c "import mcp_server; print('OK')"

MCP not connecting

• Verify path in config points to correct directory
• Reload VS Code / restart Claude Code
• Check logs: ~/.cache/delia/live_logs.json

"Unknown" responses

• Backend not running or unreachable
• Check settings.json configuration
• Run curl http://localhost:11434/health

Slow responses

• Try smaller models
• Check system resources (nvidia-smi, htop)
• Reduce context size in settings.json

Performance

Typical harvest times (modern hardware):

• Quick vine: 2-5 seconds
• Coder vine: 5-15 seconds
• MoE/Thinking vines: 30-60 seconds

License

BSD 3-Clause

Acknowledgments

• Ollama - Local LLM runtime
• MCP Python SDK - Protocol implementation
• Qwen - Base models