local-ai-stack（本地 AI 堆栈）

SKILL.md — Local AI Stack

Purpose

Transform any Mac into a powerful offline AI workstation. Installs Ollama (local model runner) + OpenCode (terminal coding agent) with the best pre-selected models. Fully offline — no API costs, no internet required.

What You Get

• Ollama — Local model runner (14GB models, ~$0 to run)
• OpenCode — Terminal coding agent with free built-in models
• 4 curated models — qwen2.5-coder, mistral, gemma3, llama3.2
• Bi-weekly auto-updates — New models pulled automatically
• OpenClaw integration — Works with your existing agent

Requirements

• macOS (Apple Silicon recommended)
• 24GB+ RAM (for larger models)
• 50GB+ free disk space
• Homebrew installed

Installation

Step 1: Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

Or download from: https://ollama.com/download

Step 2: Pull Models

ollama pull qwen2.5-coder    # Best for coding
ollama pull mistral          # Fast tasks
ollama pull gemma3          # Reasoning
ollama pull llama3.2        # General purpose

Step 3: Install OpenCode

brew install opencode

Step 4: Configure OpenCode

# Test free built-in model
opencode run "Hello" --model opencode/big-pickle

Usage

Ollama Commands

# Run a local model
ollama run qwen2.5-coder "Write a Python function..."

# List installed models
ollama list

# Pull latest model version
ollama pull qwen2.5-coder

# Remove a model
ollama rm mistral

OpenCode Commands

# Interactive coding session
opencode

# Single command
opencode run "Write a React component" --model opencode/big-pickle

# List available models
opencode models

# Help
opencode --help

Model Selection Guide

When to Use Local vs Cloud

Use Local When:

• Offline (no internet)
• Privacy-sensitive work
• Quick coding tasks
• Cost-sensitive (zero API fees)
• Simple to medium complexity tasks

Use Cloud When:

• Complex multi-step reasoning
• Web search required
• Long creative writing
• Image generation
• Advanced AI capabilities

Bi-Weekly Auto-Update

Add to cron for automatic model updates:

# Edit crontab
crontab -e

# Add this line (1st and 15th of each month at 9 AM)
0 9 1,15 * * /path/to/update-models.sh

Troubleshooting

Ollama won't start

# Check if running
ps aux | grep ollama

# Start manually
ollama serve

# Check logs
cat ~/.ollama/ollama.log

Model runs out of memory

• Close other apps
• Use smaller model (llama3.2 instead of qwen2.5-coder)
• Check available RAM: top | head -20

OpenCode not found

# Find installation
which opencode

# Reinstall if needed
brew reinstall opencode

Files

• Models stored: ~/.ollama/models/
• Config: ~/.ollama/config.json
• Logs: ~/.ollama/ollama.log

License

Ollama: MIT OpenCode: MIT

Author

Built with ❤️ for the OpenClaw community

Notes

• Models load into RAM when used, unload when idle
• Only one model runs at a time by default
• For best performance, use Apple Silicon Mac with 24GB+ RAM