Side Project Spotlight

Bill Atkins put a Mac mini in his spare room so local models can handle the coding chores he would rather skip, all without sending Stylebook's code to the cloud. Bill joins Kotaro, Steve, and Aaron to unpack his setup with OpenCode, Ollama, and Gemma, including SwiftUI migrations, crash-report utilities, RAM limits, and permission guardrails. Bill and The Trio also compare local models with their bigger cloud counterparts, explore Xcode's growing agentic toolkit, and confirm that even LLMs hate Jira.

## Chapters

00:00 Meet Bill Atkins

01:36 Why Keep AI Local

04:37 From Aider to OpenCode

09:17 Mac mini RAM and Model Choices

14:22 Permissions and Guardrails

17:17 Putting Local Models to Work

19:24 Coding Workflows and Xcode Agents

27:59 Smaller Models, Macros, and Planning

37:14 Cloud Brains and Local Tradeoffs

43:14 Bill Bot and What Comes Next

44:34 One More Thing: Stylebook

45:52 Tag

## Show Notes

- Bill has developed Stylebook with his wife since 2009, and the app is entering its seventeenth year.

- Keeping Stylebook's proprietary code away from third parties pushed Bill toward local models.

- Aider and Devstral proved the idea, but running them meant closing Xcode and Safari on Bill's 32 GB MacBook.

- Apple's WWDC26 MLX session inspired a smoother setup with OpenCode, Ollama, and Gemma on a 24 GB Mac mini.

- Bill finds Gemma particularly helpful with UIKit and SwiftUI, while smaller models still need tighter prompts and context.

- Bill locks down OpenCode so anything beyond safe directory listings and searches requires permission.

- A local model built Bill a Mac utility that turns crash reports into symbolicated stack traces.

- Aaron explains custom ACP agents in Xcode, while Steve praises its headless builds and improved SwiftUI preview tools.

- Swift macros, tests, migrations, and small utilities are ideal jobs when the work is mechanical but worthwhile.

- Aaron lands the final verdict after Steve describes Jira's API: even LLMs hate Jira.

## Links

**Bill's Local AI Setup**

Aider: https://aider.chat/ (terminal AI pair programming with local model support)

Run local agentic AI on the Mac using MLX: https://developer.apple.com/videos/play/wwdc2026/232/ (WWDC26 session on MLX, OpenCode, Xcode, and multi-Mac scaling)

OpenCode: https://opencode.ai/ (open source terminal, desktop, and IDE coding agent)

Ollama: https://ollama.com/ (runs and serves open models)

Gemma models: https://ai.google.dev/gemma/docs (Google's lightweight open models)

**Coding Agents and Workflows**

Writing code with intelligence in Xcode: https://developer.apple.com/documentation/xcode/writing-code-with-intelligence-in-xcode (Apple's agentic coding guide)

Skills for Real Engineers: https://github.com/mattpocock/skills (Matt Pocock's skills for planning, testing, and debugging)

SPS #105, How Alex Hillman Built an AI Assistant with Claude Code: https://podcast.phillycocoa.org/episodes/105-how-alex-hillman-built-an-ai-assistant-with-claude-code (a VPS assistant using Tailscale and safety guardrails)

**Notes, Tasks, and Experiments**

Obsidian: https://obsidian.md/ (local-first linked notes)

Shadertoy: https://www.shadertoy.com/ (create, browse, and share real-time shaders)

**Additional Resource**

OmniFocus: https://www.omnigroup.com/omnifocus (task management for projects and next actions)

**One More Thing**

Stylebook: https://www.stylebookapp.com/ (Jess and Bill Atkins' closet app for outfits, value, and wear)

**PhillyCocoa:** https://phillycocoa.org

Intro music: "When I Hit the Floor", © 2021 Lorne Behrman. Used with permission of the artist.