WWDC26: Run local agentic AI on the Mac using MLX | Apple

Overview

Angelos, an engineer on Apple’s MLX team, demonstrates how to run agentic AI workflows entirely on a Mac. Unlike a basic chat model, an agent repeatedly asks a model what to do, uses tools such as shell commands or file access, evaluates the results, and continues until the task is complete.

Local Agent Stack

The presentation describes four layers:

1. MLX — Apple’s open-source array framework for Apple silicon, handling computation, Metal acceleration, and memory management.
2. MLX-LM — Tools for loading, running, quantizing, and fine-tuning language models, including models from Hugging Face.
3. MLX-LM Server — A local HTTP server with an OpenAI-compatible API, structured tool calling, and support for reasoning models. Its standard interface lets compatible agents use a local model much like a cloud API.
4. Agent — A tool or framework that supports the OpenAI chat-completions protocol, such as OpenCode, Xcode, Pi agent, or a custom script.

The speaker also notes that tools including Ollama, LM Studio, and vLLM build on MLX and MLX-LM.

Setup Guide

The tutorial’s setup has three steps:

1. Install MLX-LM with pip.
2. Start mlx_lm.server with a model that supports tool calling. The speaker recommends starting with a small model to test the setup.
3. Configure an agent to use the local server’s URL and the model name it expects.

An OpenCode example shows how to define a local provider. The server runs on localhost, and the agent can use it without needing to know whether the model is local or cloud-hosted.

Performance and Scaling

The talk highlights three challenges for local agents:

• Prompt processing: Agents repeatedly process growing tool outputs and context. The speaker says M5 Neural Accelerators make matrix multiplication four times faster than on M4, and that MLX’s specialized multiplication and attention kernels can translate this into nearly the same prompt-processing speedup. MLX selects the appropriate kernels automatically.
• Concurrent agents: MLX-LM Server uses continuous batching to handle simultaneous requests. Incoming requests can join a batch already being processed, helping subagents avoid waiting in a queue.
• Large models: MLX distributed inference can spread a model across Macs connected by Thunderbolt or Ethernet. The speaker gives DeepSeek’s 1.6-trillion-parameter model as an example, saying its weights require more than 800 GB of memory. Distributed inference can enable larger models and parallelize prompt processing. Setup uses mlx.launch and a hostfile; macOS 26.2 adds Thunderbolt RDMA support. The speaker reports speedups of up to three times with four nodes.

Demonstrations

• Repository review: OpenCode uses a local model to fetch recent MLX pull requests with the GitHub CLI, read diffs, summarize changes, and flag items needing attention. The model runs locally, while the Git commands access the network.
• SwiftUI app creation: An agent builds a drawing app from a blank Xcode project, writes and builds the code, addresses build errors, and then adds rounded end caps at the user’s request.
• Xcode bug fixing: Xcode connects to the local MLX server through its Intelligence settings and a locally hosted provider. The model inspects a deliberately introduced bug, edits the code, and helps verify the fix by building and running the app.

Key Takeaways

The presentation emphasizes local execution for privacy, offline availability, and avoiding cloud API costs. It concludes that MLX-LM, its server, and the demonstrated workflow are open source and available.

No product reviews are presented; the video is a technical overview, setup guide, and set of demonstrations.

Main speaker/source: Angelos, an engineer on Apple’s MLX team; published by Apple Developer.