Claude Code without the cloud bill or the NDA panic

What it does

This repo turns an Apple Silicon Mac into an on-device inference stack that exposes local models through an Anthropic-compatible API. That lets Anthropic’s own claude-code CLI run against quantized local brains—Qwen, Gemma, Llama, and DeepSeek V4 Flash—instead of cloud servers. The project positions itself as the “brain” of a four-part ambient-computing stack, with sibling repos handling voice and remote browser access.

The interesting bit

The author builds and hosts custom “abliterated” MLX quantizations, then wraps them in launchers and a local server so Claude Code needs zero reconfiguration to go offline. For the 1-million-context DeepSeek V4 Flash, it swaps Apple’s MLX for Antirez’s bare-metal ds4 engine—pure C and Metal kernels—and persists the KV cache to disk so the massive system prompt prefills exactly once. It is essentially a compatibility layer pretending to be a datacenter, which is a lot of engineering to avoid a network request.

Key highlights

• Qwen 3.5 122B, Gemma 4 31B, Llama 3.3 70B, and DeepSeek V4 Flash (1M context) served via MLX or Antirez’s ds4 engine.
• Fully offline and airgap-ready; the README demos NDA review with Wi-Fi physically disabled and lsof running live.
• Custom “abliterated” MLX quantizations hosted on HuggingFace, not generic mirrors.
• RAM floor is 32 GB for Gemma, 96 GB for Qwen, and 128 GB for DeepSeek.
• Voice and browser-remote access are handled by separate repos in the four-part “ambient” stack.

Caveats

• The hardware bar is high: the README lists minimum RAM at 32 GB for the smallest model and 128 GB for DeepSeek V4 Flash.
• The documentation is interleaved with marketing for a pre-configured Mac mini (“AirGap Box”) and a $19 agent pack, which blurs the line between open-source project and product storefront.
• Voice and browser features rely on companion repositories that are mentioned but not detailed here.

Verdict

If you handle sensitive code or documents on a maxed-out Apple Silicon Mac and want Claude Code without egress, this is a direct path. If you are RAM-constrained, running Intel, or simply prefer other front-ends, it is not for you.