claude-code-localRun Claude Code Locally on Apple Silicon

Bringing a powerful coding agent like Claude Code to run entirely on your local machine sounds like a dream for many, but few projects actually deliver. claude-code-local is one of those rare efforts, meticulously building an Anthropic API-compatible local service on Apple Silicon, powered by MLX, ensuring your Claude Code requests never leave your device.

The project's name pretty much spells out its mission. It functions as a local API endpoint, allowing Claude Code to make its usual requests, but instead of hitting a cloud server, the traffic is routed to a local model. The repository lists supported models including Qwen 3.5 122B, Llama 3.3 70B, and Gemma 4 31B. The project claims a reference speed of 65 tokens/second with Qwen 3.5 122B.

It's crucial to put that speed figure into context. Local inference performance is heavily dependent on your specific Apple Silicon chip, the amount of unified memory, and the quantization method used. That 65 tok/s is a benchmark from the project's specific setup, and your mileage will almost certainly vary. A more pressing concern for many will be memory: models of the 122B scale demand a significant amount of unified memory, making them impractical for entry-level Macs. Your ability to run these larger models hinges entirely on your machine's memory capacity.

For those genuinely facing compliance pressures, the ability to keep data on-device often outweighs raw generation speed.

Why Bother with Local AI?

Entrusting your code to cloud-based AI has become commonplace, but for certain teams, this isn't a matter of preference; it's a hard red line. Projects under strict NDAs, processing sensitive medical data, or analyzing legal documents — in these scenarios, merely knowing which server your prompts traverse can introduce significant compliance risks.

This is precisely where claude-code-local shines: data never leaves your device. The project description explicitly states it's 'offline' and 'airgap-ready,' meaning it's not just about minimizing data transfer, but designed to allow full workflows even in disconnected environments. The trade-off in speed for this feature is a worthwhile exchange, enabling the secure use of Claude Code in highly sensitive contexts.

API Compatibility: The Real Game Changer

Running a coding agent locally isn't just about getting it to execute; it's about seamless integration. claude-code-local tackles this by reimplementing the Anthropic API locally, essentially providing Claude Code with a local exit point. This means you don't need to alter your existing workflows or switch tools; you're simply swapping the cloud backend for your own machine.

Of course, there are clear trade-offs. Local models typically don't match the capabilities of their larger cloud counterparts, inference speed is capped by your hardware, and this entire solution is currently exclusive to Apple Silicon. Looking at the repository, the project has garnered around 3.1k stars and over 600 forks. It's implemented in Python, leverages MLX natively, and appears to be in its relatively early stages of development.

Who Should Pay Attention?

• Developers working in legal firms, hospitals, or under strict NDA environments where code cannot leave the internal network.
• Claude Code users with high-memory Apple Silicon machines who want to reduce their reliance on cloud APIs.
• Developers interested in MLX inference and local model orchestration; this repository offers an excellent reference implementation.

Considerations Before Diving In

First and foremost, confirm your hardware. Apple Silicon is a strict requirement, and your unified memory capacity directly dictates the size of the model you can run. Don't expect a low-spec machine to handle a 122B model.

Secondly, always refer to the project's README. The list of supported models and Claude Code versions can change, so check the current documentation before you start, rather than relying on outdated guides.

Finally, be prepared for a different user experience. Local inference typically introduces noticeable latency compared to cloud APIs. For compliance-driven scenarios, this is a fair trade; however, if your primary motivation is just cost savings, you might find the performance disappointing.

claude-code-local, in its current form, feels like a promising early solution pointing in the right direction. Its most significant contribution is proving that tools like Claude Code can indeed operate independently of the cloud — a crucial capability for data-sensitive environments. As Apple Silicon Macs with larger memory become more common and local models continue to evolve, projects like this will become increasingly difficult to overlook.