mac-code-local-ai-agent

What This Skill Does

The mac-code-local-ai-agent is a sophisticated CLI-based AI coding assistant designed specifically for Apple Silicon hardware. It serves as a local alternative to proprietary tools like Claude Code, allowing developers to execute reasoning-heavy coding tasks without incurring subscription costs or sending sensitive code to cloud-based servers. By leveraging local LLM backends such as llama.cpp and MLX, the agent provides high-performance inference—up to 30 tok/s for 35B models—directly on your Mac. It features an intelligent router that classifies user input into three distinct categories: 'search' (utilizing DuckDuckGo for context), 'shell' (executing system commands for environment management), and 'chat' (leveraging LLMs for logic and code generation). It is optimized for memory efficiency, supporting custom MoE Expert Sniper modes that allow running large models even on base 16GB RAM configurations.

Installation

To install this skill, use the ClawHub command: clawhub install openclaw/skills/skills/adisinghstudent/mac-code-local-ai-agent. Once installed, ensure you have the required dependencies by running brew install llama.cpp and pip3 install rich ddgs huggingface-hub mlx-lm. You will need to clone the underlying repository and download the recommended GGUF model files into a local directory, typically ~/models. After setting up the model, start the server using the appropriate backend (llama.cpp or MLX) and launch the main agent interface with python3 agent.py.

Use Cases

This skill is ideal for developers working in secure environments who cannot upload proprietary source code to third-party AI platforms. It is perfect for rapid prototyping, performing complex refactoring tasks using local file system access, and researching libraries via its built-in search tool without leaving the terminal. It excels in scenarios where 64K context windows are required for parsing large documentation or legacy codebases.

Example Prompts

1. "Search for the latest documentation on using Pydantic V2 and create a summary file named pydantic_guide.txt in the current directory."
2. "List all files in the src directory, then search for any Python files that don't have unit tests and list them."
3. "Refactor the authentication logic in auth_utils.py to use the new security middleware pattern and ensure all existing tests pass by running pytest."

Tips & Limitations

Performance is heavily dependent on your Mac's unified memory; ensure you have sufficient RAM available when running 35B models. While the agent supports file writing and shell execution, be cautious when running automated commands in critical directories. Always verify the output of shell commands before allowing the agent to proceed with multi-step workflows. The MLX backend provides superior context management but is currently best suited for 9B models for optimal stability.