A WWDC26 session walkthrough on building and running agentic AI workflows entirely on a Mac using MLX, with no cloud or API keys required. The four-layer stack consists of MLX (array framework for Apple silicon), MLX-LM (model loading/quantization), MLX-LM Server (OpenAI-compatible HTTP server with tool calling), and any agent framework on top. Setup takes three steps: install MLX-LM via pip, start the server with a tool-calling model, and point an agent at localhost. Key hardware optimizations covered include M5 Neural Accelerators providing 4x faster matrix multiplication for prompt processing, continuous batching for concurrent multi-agent requests, and distributed inference across multiple Macs via Thunderbolt/Ethernet (up to 3x speedup with four nodes, supporting models like DeepSeek's 1.6T parameter variant). Live demos show an agent building a SwiftUI drawing app from scratch using OpenCode, and Xcode's Intelligence tab connecting directly to the local MLX server to identify and fix bugs — all with code staying on-device.