nicedreamzapp/claude-code-local

What it solves

Claude Code Local allows users to run the Claude Code terminal interface entirely on-device using local AI models on Apple Silicon. This eliminates reliance on cloud APIs, removes usage limits, ensures 100% privacy by keeping code on the local machine, and removes the network latency associated with cloud calls.

How it works

The project provides an MLX-native server that speaks the Anthropic API natively. Unlike traditional setups that use a translation proxy (e.g., Claude Code $\rightarrow$ Proxy $\rightarrow$ Ollama $\rightarrow$ Model), this server removes the proxy layer, allowing Claude Code to communicate directly with the local model. This architecture reduces task completion time significantly (from 133 seconds to 17.6 seconds in benchmarks).

Who it’s for

Developers using Apple Silicon Macs (from 16GB MacBook Airs to high-end M-series Max chips) who want to use Claude Code without API keys, subscriptions, or internet connectivity, particularly those handling confidential codebases.

Highlights

• Native Anthropic API Support: Eliminates translation proxies for 7.5x faster performance.
• Diverse Model Roster: Supports a wide range of models including Llama 3.3 70B, Gemma 4 31B, Qwen 3.8 27B, and DeepSeek V4 Flash.
• High-Context Options: DeepSeek V4 Flash provides up to 1 million tokens of context with a persistent disk KV cache.
• Abliterated Models: Provides custom MLX builds of models with suppressed refusal directions for fewer benign-but-edgy refusals.
• Hardware Optimized: Specifically tuned for Apple Silicon using MLX and specialized engines like ds4 for DeepSeek.