nicedreamzapp

claude-code-local

AIAIMLXLocal-LLMApple-SiliconClaude-Code

// summary

Claude Code Local provides a suite of high-performance AI models that run entirely on Apple Silicon hardware without requiring cloud connectivity. The project features a native MLX server that enables local execution of Claude Code, browser automation, and voice interaction while ensuring complete data privacy. By eliminating outbound network calls and telemetry, it offers a secure, air-gapped environment for handling sensitive professional tasks.

// technical analysis

Claude Code Local is a privacy-focused, local-first ambient computing stack designed to run powerful AI models entirely on Apple Silicon hardware without any cloud dependency. By leveraging the MLX framework and native Apple GPU acceleration, it provides a secure environment for handling sensitive data, such as legal or financial documents, by ensuring zero outbound network traffic. The project addresses the limitations of local models—specifically tool-call reliability and performance—by implementing a custom Python-based MLX server that optimizes model inference and recovers garbled tool outputs, effectively bridging the gap between local execution and enterprise-grade AI coding tools.

// key highlights

Enables 100% local AI execution on Apple Silicon, ensuring that sensitive code and data never leave the user's machine.

Supports a roster of high-performance models including Gemma 4 31B, Llama 3.3 70B, and Qwen 3.5 122B, optimized for different hardware constraints.

Features a custom MLX server that achieves high throughput, reaching up to 65 tokens per second with specific model configurations.

Includes a robust tool-call recovery mechanism that fixes common formatting errors in local model outputs to prevent infinite loops during automated tasks.

Provides four distinct operational modes, including autonomous browser control, hands-free voice interaction, and iMessage integration for a complete local-first workflow.

Maintains strict privacy by eliminating telemetry and phone-home behavior, allowing for air-gapped operation in high-security environments.

// use cases

Run Claude Code locally with Gemma, Llama, or Qwen models for private, subscription-free development.

Execute autonomous browser agents via Chrome DevTools for local web navigation and interaction.

Enable hands-free voice control with on-device speech-to-text and voice cloning capabilities.

// getting started

To begin, ensure you are using a Mac with Apple Silicon and sufficient RAM for your chosen model. Explore the project by navigating to the 'launchers/' directory to find double-clickable command files for different modes, or use the provided scripts to start the MLX server with your preferred model configuration.