Initial release of a fully local, zero-API-key multi-agent system for Apple Silicon. Two LLM agents (Coder and Sheriff) collaborate via a shared mlx_lm.server instance to write, test, and debug Python scripts autonomously. Key features: Single-server, two-persona architecture optimized for 16 GB Apple Silicon Macs Self-correcting feedback loop with bounded autonomy (max attempts, stagnation detection, token budget) Orchestrator-driven code execution with markdown extraction fallback Three showcase scenarios: calendar generation, CSV analytics pipeline, gradient descent from scratch CLI interface via Typer with auto server lifecycle management Tech stack: MLX, mlx-lm, Agno, Typer, Rich, Pydantic Model: Qwen2.5-Coder-7B-Instruct-4bit (auto-downloaded, ~4 GB)