This page is everything I've shipped — phone apps, local-AI tools, agents, and robots. All of it runs on hardware you own, and all of it is free to fork on GitHub.
📰 Read the build log — all articles and announcements →
Built start to finish on this stack, shipped to both stores. Private by default, pay once or nothing.
601-class object detection, OCR, offline translation and LiDAR depth — 100% on-device. Free, no ads, no cloud calls.
Describe a song in any of 19 languages and own a full-length original in minutes. No subscription, every song 100% yours.
Complete K–4 learning adventure — 2,000+ activities, a narrator that reads everything aloud, 100% offline. No ads, no accounts, nothing tracked.
Run Claude Code 100% on-device. MLX-native Anthropic-API server with a six-model local lineup — Muse-Glimmer 30B (multimodal), Gemma 4 31B, Qwen 3.5 122B at 65 tok/s, DeepSeek V4 Flash with 1M context. Zero cloud, zero API fees. The same stack is AirGap: the full private-AI build for firms that can't put client work in the cloud.
The model side of the stack. 15 MLX-quantized open-weight models — Llama 3.3 70B, Gemma 4 31B and 12B, Hermes 4 14B, Qwen3-VL 32B vision, Qwen3.6 27B, Nemotron-3 Nano Omni, Muse-Glimmer 30B — plus CoreML detectors. The exact files that ship with claude-code-local and the RealTime AI Camera iOS app. Live download counts, no cloud inference, run on your own M-series Mac.
Porting frontier model architectures to run natively on Apple Silicon — the layer everything else stands on.
NVIDIA's tri-modal Nemotron Omni — text, vision and audio — running entirely on Apple Silicon. Pure MLX runtime, parity-verified against NVIDIA's reference.
Architecture support for Muse-Glimmer multimodal (vision + language) models in mlx-vlm on Apple Silicon.
Muse-Glimmer text-model support submitted upstream to Apple's mlx-lm. PR #1710, open — contributing at the framework level, not just consuming it.
CogVideoX-5B video generation running on Apple Silicon — the setup that made local text-to-video real on an M4 Pro.
One Mac: brain, ears, eyes and hands — plus the phone in your pocket. Ambient computing that never phones home.
The full private AI workstation build: one Mac, wifi off, nothing leaves the building. A capability I keep sharp — the whole stack is open source.
Backup brain for Claude Code. Hit your plan limit or an outage — flip one command and keep working on local models.
Hands-free voice loop — Apple speech recognition in, cloned-voice TTS out. Both directions fully on-device.
The Windows side of the voice loop — dictate into any app, have any text read aloud in natural neural voices.
Local MLX + Chrome DevTools Protocol. Drives a real browser autonomously — iframes, Shadow DOM, ProseMirror.
Control Claude from your iPhone via iMessage. Text commands in, screenshots and videos out.
Open a browser tab → you're driving your Mac mini. Live screen, real PTY shell, on-device voice.
macOS screen + facecam recorder with a local HTTP API. Pairs with Claude Code for record-and-send pipelines.
FastAPI automation server for iOS development — remote Xcode builds, install-to-device, app launch, live monitoring.
Agents doing real work on real machines — and the leaderboard that keeps them honest.
The local agent leaderboard — filesystem-judged agent tasks on consumer hardware, with reference-validated judges.
Wake up to AI-scored job matches every morning. Reads your full credentials, scores roles across 50+ companies. Mac and Windows.
395 dormant X follows removed in 40 minutes using your own logged-in browser — no scraper, no API key, no stored credentials.
Small, dependency-free tools that solve real problems in a local-AI shop.
Find the license landmines in a local AI stack — including the ones inside the model weights. No dependencies, runs offline.
A credits file that names the people, not just the packages — resolves every dependency back to the humans who wrote it.
The x402 payment protocol for PHP — let AI agents pay your site in stablecoins. Zero dependencies, works in WordPress/WooCommerce.
Building toward a modular-robot-parts future — local AI vision in things you can actually touch.
1kW autonomous tank platform — LIDAR, 601-class object detection, iMessage natural-language control, Jetson brain.
Roboflow on your phone — label photos and train your own on-device vision model for a robot, privately, from the camera roll.
Open-source single-seat amphibious electric VTOL (FAA Part 103 ultralight). Lands on water and takes off again. Full CAD.
Side quests that turned into real results — research, thesis defenses, and one little kid's tutor.
Cross-cohort test of whether AGTR1+ dopamine neurons are selectively lost in Parkinson's — 579,931 cells, four independent datasets, same direction in all four.
Free 1st-grade tutor — adaptive, voice-guided, click-only safe. Play it now in a browser, no login, no app.
An HSU Master's Thesis project at Cal Poly Humboldt Marine Sciences. Four single-page features — molecule-to-whale scale of the sea, marine and freshwater restoration, and a 13-chapter scientific-diving deep dive. Hand-coded SVG, offline-ready, each page reads itself aloud. Built end-to-end in Claude Code.
Films, worlds and toys — all generated or hand-built on the same local stack.
Turn a story into a finished film on one Mac — stills, motion, voices, music, grade and sound. The pipeline behind The Lucid Engine, Antarctica, and The Bear Sister.
Immersive WebXR virtual cannabis marketplace built with A-Frame. Live at marijuanaunion.com/marketplace/.
All of this is free and open, and you can run everything on your own Mac without paying me a dime. If it saved you an afternoon, a few dollars a month keeps me building in the open.
Matt Macosko, working solo out of Arcata, CA. Nice Dreamz LLC. Decades of physical-product iteration before AI accelerated everything into software. Started with a chicken problem. Still figuring it out.
Three ways to engage, depending on who you are: