mlx-code 0.0.4

Coding Agent for Mac

mlx-code

A lightweight, powerful Coding Agent for Mac. Built on top of Apple's MLX framework, mlx-code provides fast, local inference with built-in prompt caching and robust tool-calling capabilities.

It features a multi-provider local server, a terminal-based chat REPL, and a dedicated TUI for inspecting logs.

Features

• Local MLX Inference: Powered by mlx-lm for optimized performance on Apple Silicon. Includes intelligent prompt caching.
• Multi-Provider Compatibility: Seamlessly translates and handles requests formatted for Claude, Gemini, Codex, DeepSeek and standard OpenAI APIs.
• Built-in REPL & Tools: Comes with pie, a fully-featured chat REPL with tool execution and reasoning token support.
• TUI Log Viewer: Includes a Curses-based Terminal UI for filtering, inspecting, and tracking JSON logs in real-time.
• Server Mode: Easily spin up a local server compatible with standard LLM tooling.

Quick Start

Install via pip and launch the agent immediately:

pip install mlx-code
mc

Command Line Interfaces

The package installs three primary command-line tools:

• mc (Main Agent/Server): Runs the core agent and local API server (defaults to 127.0.0.1:8000).
• me (REPL): Launches the interactive pie chat REPL.
• md (Diagnostics/Logs): Opens the TUI viewer to navigate and filter JSON logs generated by the agent.

Options

You can customize the model, server, and behavior using command-line flags.

# Use Gemini CLI as the harness
mc --harness gemini 

# Use a custom local LLM backend
mc --model mlx-community/Qwen3.5-4B-OptiQ-4bit

# Use DeepSeek V4 Flash API
me --deepseek

# Run the server only
mc --nocc

# General shell piping and chaining works too
echo "explain symgraph.py" | mc | cat - PLAN.md | mc

(For a full list of mc server arguments, run mc --help)

Credits

• main.py: Built on MLX and MLX LM by Apple.
• pie.py: Adapted from pi by Mario Zechner (MIT License).

Licence

Apache License 2.0 — see LICENSE for details.