Matt Macosko

Something was eating my chickens. That is the whole origin story.

Protecting them needed a camera that could see at night. That needed a vision model small enough to live on a robot. That needed AI running on my own hardware instead of somebody else's cloud. Fifty-three repos later I am somehow doing neuroscience, and every one of them still traces back to the coop.

Local AI on Apple Silicon · Humboldt County, CA · open to work

matt@ineedhemp.com · looking for a role building on-device / local AI.

• claude-code-local: Claude Code running 100% on-device on Apple Silicon, with its own MLX server and tool-call parser. 3,300+ stars, 625 forks, and fixes merged from contributors on hardware I don't own.
• Upstream in Apple's MLX: Muse Glimmer (Meta) text model support merged into ml-explore/mlx-lm (+302 lines). Thanks to @zcbenz for the review and merge.
• nemotron-omni-mlx: first open Apple Silicon runtime for NVIDIA Nemotron Omni's vision and audio towers, 23/23 parity tests against NVIDIA's PyTorch reference.
• Agent-12: a filesystem-judged local agent leaderboard, independently reproduced by @galashko.
• Shipped apps: 4 apps live on both the App Store and Google Play (Hive Strike live on Play, in App Store review), 1,386 App Store downloads and 133 Google Play installs.

Every app, one row, both stores. Private by default, pay once (or nothing), no subscriptions.

App	What it is	iPhone	Android
Song Forge	Describe a song, get a finished original. Any genre, 19 languages, rendered on my Macs at home, never stored. 10 free songs, no account, no email.
RealTime AI Cam	Point your camera, it names 601 objects. Fully offline.
RealTime Space	Fly a real-physics starship through the solar system and 8,920 real stars.
Brainforest	2,000+ K-4 lessons, reads everything aloud, fully offline, one-time $1.99. No ads, no tracking. Built for my daughter.

| Hive Strike | One bee against sixteen worlds, sixteen bosses and forty-eight kinds of insect. Every sprite, background, animated scene and music bed generated locally on one Mac, no cloud, no stock art. | | |

Android source ships in each repo's android/ folder.

Trinidad Head — my own terminal, built from scratch for Windows and Mac. Glass windows with a neon glow, a different color for every open window, and your own questions easy to spot when you scroll back through an AI chat. Project page →

nemotron-omni-mlx — NVIDIA's Nemotron Omni sees, hears, and reasons, but its vision and audio towers had no open runtime for Apple Silicon. So I wrote one. Pure MLX, MIT, 23/23 parity tests against NVIDIA's PyTorch reference. 67.7 tok/s with an image, 147 tok/s with audio, wifi off. Credit to NVIDIA for the open weights and yayr for the 4-bit conversion.

Most of these repos look standalone. They're not. They're one system spread across three machines that talk to each other 24/7:

• 💻 M5 Max MacBook Pro · 128 GB — heavy compute: model inference, video, image, training
• 🖥️ M4 Mac mini · 64 GB — always-on workhorse: autonomous agents, FiaOS web shell, free Claude for the whole mesh
• ☁️ Hostinger VPS — public gateway: HQ Dashboard, the Ohm chat agent, sites, cron

The idea: Apple Silicon is fast enough now that the everyday AI I want doesn't need a cloud subscription. Build local-first, stay useful when the internet goes down. The same stack is airgap-ready, which matters to the law / medical / compliance-sensitive firms I work with through Nice Dreamz LLC.

                          ┌─────────────────────────────┐
                          │   iPhone (iMessage bridge)  │
                          │   ask · send · image · vid  │
                          └────────────┬────────────────┘
                                       │
       ┌───────────────────────────────┼───────────────────────────────┐
       │                               │                               │
┌──────▼──────────┐         ┌──────────▼──────────────┐     ┌──────────▼────────┐
│  M5 Max         │         │   Mac mini (always-on)  │     │  VPS (Hostinger)  │
│  workstation    │◀──SSH──▶│   the workhorse         │◀───▶│   gateway         │
└─────────────────┘         └─────────────────────────┘     └───────────────────┘

Thirteen always-on agents run across those nodes: a GitHub watcher that drafts issue replies with local AI and texts me the draft, a 9am iMessage business briefing, email triage, a Reddit brand watcher, and the trick that powers it all — mini_claude, which lets every agent use my Claude Max plan for $0 per call.

AirGap is my name for the full private-AI build: a complete AI workstation on one Mac, wifi off, nothing leaving the building. It started as a packaged offering; these days I treat it as a capability I keep sharp rather than something I sell. Everything it needs is open source on this page, so any competent IT person can stand one up without me. That's the point.

If your firm handles data that can't touch the cloud and you want one built and your team trained on it, email info@nicedreamzwholesale.com. What it looks like → · Why it matters now →

Sorted by GitHub stars.

Run Claude Code 100% on-device on Apple Silicon. Current lineup led by Qwen 3.8 27B and Gemma 4 31B, four modes (code, browser, voice, phone), Anthropic-compatible MLX server, setup.sh picks the right model for your RAM. Your code never leaves your Mac.

MLX Apple Silicon Claude Code Python

Talk to your Mac, hear it reply in your own cloned voice. 100% private, works on a plane, and the Ohm web chat brings the same loop to any browser at zero API cost.

Pocket TTS Apple Silicon Flask

Local MLX models driving Chrome via DevTools Protocol. Handles what other agents can't: cross-origin iframes, Shadow DOM, ProseMirror. 56-second demo →

MLX Chrome DevTools Protocol Local AI

One logged-in browser, many agents, nobody fights. Leases each agent an ordinary background tab in your real browser, with your real logins, and every agent sees only its own. ~300 lines, one dependency. Verified on Linux by a stranger who then found a bug in my own docs.

Chrome DevTools Protocol Multi-agent Local AI

The biggest free open-vocabulary detector that fits on an iPhone: 601 classes, OCR, LiDAR depth, all on-device. The seed of the robot's vision stack.

YOLOv8 CoreML Swift LiDAR

Control Claude Code from your iPhone via iMessage. Commands by text; screenshots, recordings, and YES / NO / YES_TO_ALL approval buttons come back.

iMessage AppleScript macOS Automation

Started with protecting my chickens from predators. Now a 1 kW autonomous tank robot you text from your phone: 601-class vision on a Jetson, dual PTZ cameras, 3D LIDAR, and it pulls a cart of firewood. Actually. The honest goal: a Lego-clip-in modular robotics kit running the same local-AI brain as the desk. VisionBuilder trains its eyes from photos on my phone.

Jetson Orin Nano YOLOv8 Teensy RPLidar Robotics

2,000+ K-4 lessons with a friendly narrator that reads everything aloud. Themes she picks (unicorns, dinosaurs, space, cats, Bluey), difficulty that climbs as she masters skills and gently drops when she struggles, click-only kid-safe interface. Fully offline, one-time $1.99, no ads, no tracking, no internet needed. Source is open: brainforest-app, iOS and Android in one repo.

iOS Fully Offline Adaptive Learning Kid-Safe

Four real jobs in four real browser tabs at once, each answer checked by script, no AI judge (Sep 16, 2026, M5 Max, browser-agent + browser-broker):

Every model we run is also scored on real agent tasks at the open Agent-12 leaderboard.

Divine Tribe (vaporizer hardware, American-owned since 2013) plus three more stores — 140+ products, 5,000+ units shipped, all automated by the mesh above. Divine Tribe pays the bills; everything else is downstream.

I run Divine Tribe — vaporizer hardware, since 2013. Cinch ships those orders. Everything else is downstream of paying the bills.

Nice Dreamz LLC is the umbrella on top. AirGap AI is the showcase: what a full private-AI build looks like for the kind of firm that can't put its data into a cloud chatbot. It's a capability I maintain, not a product I'm pushing. The three-node mesh is the demo: it's been my daily driver for a year and a half. I'm also building the Cannabis Device Safety Institute in parallel, a standards body for cannabis-consumption hardware.

The long arc is the robot. Same eyes, different bodies.

Still figuring it out. No investors. No team. One MacBook, one Mac mini, one VPS, and a lot of late nights.