claude-code-mlx-proxy

This skill enables Claude Code to use open-source models running locally on Apple Silicon Macs via MLX framework, eliminating the need to send code to Anthropic's cloud servers. The proxy server implements Claude's Messages API, supporting streaming responses, token counting, and compatibility with thousands of models from the MLX Community on Hugging Face, including Llama 3, GLM-4.5-Air, and DeepSeek. Setup involves running a local Python server and configuring Claude Code to point to it via the ANTHROPIC_BASE_URL environment variable. Development teams gain complete code privacy, offline capability, zero API costs, and full control over model parameters and generation settings.

[{"step":"Start the MLX model server","action":"Run `python -m mlx.serve --model mlx-community/code-llama-7b-instruct-4bit --port 8000` in your terminal to start the local model server.","tip":"Use a GPU-enabled machine for best performance. Monitor GPU memory usage with `nvidia-smi` if available."},{"step":"Configure claude-code-mlx-proxy","action":"Set the proxy endpoint in your environment: `export CLAUDE_CODE_MLX_PROXY=http://localhost:8000/v1` or configure in your IDE settings.","tip":"Test the connection with `curl http://localhost:8000/v1/models` to verify the server is responsive."},{"step":"Execute coding tasks","action":"Use standard claude-code commands (e.g., 'Write a function to...') but the proxy will route requests to your local MLX model.","tip":"For large codebases, pre-load relevant files into the context window to improve response quality."},{"step":"Review and save outputs","action":"Inspect the generated code, verify functionality, and save to your project directory. Use `git diff` to track changes.","tip":"Compare outputs with previous versions using `git diff HEAD~1 -- path/to/file.py` to ensure no regressions."},{"step":"Optimize performance","action":"Adjust model parameters (temperature, max_tokens) in your proxy configuration for faster inference or higher quality outputs.","tip":"For debugging, add `--verbose` flag to the proxy server to log model interactions."}]