wand CLAUDE.md

Wand MCP Tool - Claude Context Guide

IMPORTANT: Initial Setup Check

Every Claude instance should perform these steps before proceeding:

1. Verify Wand MCP Server Connection:

   claude mcp list
   

   Expected output should show: wand: ... - ✓ Connected

2. Test Ollama Integration:

   Use mcp__wand__ollama with operation "list_models"
   

   This confirms the Ollama server at http://localhost:11434 is accessible (configure via OLLAMA_BASE_URL env var)

3. If tools are not available:

   • Run the add_to_claude.sh script from the wand directory
   • Restart Claude Desktop to load the MCP configuration
   • Verify connection again with steps 1 and 2

Overview

Wand is a Model Context Protocol (MCP) server that provides 50+ integrations for AI development workflows. This guide explains how to use Wand's Ollama integration to offload heavy AI tasks to local models while maintaining quality control.

Quick Reference for M1 Max 64GB Setup

# Your optimal settings for most tasks
{
    "model": "qwen3:30b-a3b-instruct-2507-q4_K_M",  # Best quality
    "max_tokens": 4096,   # Can go up to 8192+ comfortably
    "temperature": 0.2    # Consistency over creativity
}

# For speed-critical simple tasks
{
    "model": "qwen2.5-coder:14b-instruct",
    "max_tokens": 2048,
    "temperature": 0.3
}

Key Insight: With 64GB RAM, prefer single large requests over decomposition. You can generate complete implementations without hitting memory limits.

Available Wand Tools

When Wand is properly configured, you'll have access to MCP tools prefixed with mcp__wand__. The primary tool for AI code generation is:

mcp__wand__ollama

Local AI model execution through Ollama server (default: http://localhost:11434, configure via OLLAMA_BASE_URL)

Core Operations

1. Code Generation (generate operation)

Use this for heavy code generation tasks:

Tool: mcp__wand__ollama
Operation: generate
Parameters:
  - prompt: The code generation request
  - model: (optional) defaults to "qwen3:30b-a3b-instruct-2507-q4_K_M"
  - temperature: (optional) defaults to 0.7
  - max_tokens: (optional) defaults to 2048