American Code is an independent AI lab — the kind the open models in your Ollama library come from, aimed at the layer around them: distributed AI, on‑device intelligence, hardware, and developer tools. A lot is already shipped. A lot more is coming.
Hardware, apps, distributed AI, developer tools — a slice of what we’ve shipped. It’s a running list, not a finished one.
A global inference fabric made of consumer Macs.
An Apple MLX–based mesh network that turns idle Apple Silicon into a distributed AI inference grid. No cluster to rent, no queue to wait in — the machines find each other and get to work. Open source, open protocol, AGPL.
mlxmesh.net ↗A physical control surface for your AI agents.
A custom RP2040 hardware macropad paired with a Tauri/Rust desktop app that orchestrates coding agents across Claude, Cursor, Windsurf, and VS Code. Auto‑commits, codebase analysis, security checks, multi‑agent monitoring — and a real red emergency‑stop key.
ai-vibedeck.com ↗An IDE that runs your models, not the cloud’s.
An open‑source Ollama model IDE built on the VS Code platform: multi‑model orchestration, device‑specific model recommendations, and Ralph‑based code / test / repair loops that keep grinding until it’s green. Local by default. AGPL.
github.com/american-code ↗An NVIDIA DGX cluster, in your pocket.
A native iOS app for monitoring and commanding an on‑prem NVIDIA DGX GPU cluster — live utilization, job control, and anomaly checks routed through an on‑prem LLM. The whole rack, from your phone.
Native iOS · on-premCore ML models for anyone, from raw data.
A native macOS app that let anyone build Core ML models straight from their own data — vision, semantics, and numeric prediction — no ML degree, no pipeline, no data science team. Train on your machine, ship on your machine.
macOS · c. 2018Full-spectrum RF scanner, in your pocket.
A native iPadOS app that turns a tablet and a commodity RTL‑SDR dongle into a spectrum analyzer and automatic emitter classifier — 24 MHz to 1.766 GHz. Full DSP chain in Swift/Accelerate, Metal waterfall display, band‑plan‑aware classification, and BLE coverage out to 2.4 GHz.
Active developmentNeural TTS for research papers. On your device. No cloud.
An iPad app that reads academic PDFs aloud using Kokoro 82M — a StyleTTS2 model reimplemented in MLX Swift — at inference time with zero network calls. Complete synthesis pipeline in Swift: G2P frontend, PLBERT prosody encoder, iSTFT vocoder. Concept map extracted alongside.
Active developmentOpenAI-compatible inference server. Pure Swift. No cloud.
A production‑grade Swift 6 inference server built on Hummingbird 2 and MLX Swift — exposes the OpenAI API over local models on Apple Silicon. Prefix‑trie KV cache, deadline‑bounded batch assembler, LRU model pool with lazy loading from Hugging Face. No Python, no Docker, no cloud.
Active developmentThe collective communications layer for Apple Silicon clusters.
An NCCL‑equivalent written in Swift — auto‑tuned
topology (ring vs. tree vs. hierarchical), in‑flight wire
compression on slow links, and a C/Swift API that lets MLX,
llama.cpp, and raw‑Metal workloads adopt it by renaming
nccl → mccl.
Run @triton.jit kernels unmodified on a Mac GPU.
A Triton compiler backend targeting Apple Metal — Swift core that emits MSL, compiles metallib, and drives the Metal runtime. The matmul tutorial at 75–76% of Apple’s own MPS; FlashAttention‑2 beating the MPS composite at s ≤ 1024. Triton targets only CUDA/ROCm; this removes that lock‑in.
Active developmentNsight Compute for Apple Silicon.
Roofline analysis against measured (not spec‑sheet) chip peaks,
per‑dispatch counter capture via MTLCounterSampleBuffer,
static occupancy from MTLComputePipelineState, and a kernel
differ. Instruments doesn’t understand ML kernel shapes; metalscope
does.
Your EC2 rack, on your wrist.
Agentless SSH‑based monitoring for EC2 hosts — host status, systemd services, listening ports, Docker containers, and journald logs, surfaced on iPhone and glanceably on Apple Watch. Add a PEM key, point it at a box. Nothing installed on the server.
Active developmentNative diagramming that exports the real thing.
One SwiftUI codebase, one renderer — so the on‑screen
canvas, PNG export (1×–6×), and PDF export (true
vector, selectable text) all call the same drawing path and can never
drift. Native .structor package format reopens and edits
losslessly.
Tinkercad‑style 3D modeling. Free. No account.
Build from parametric shapes, carve with holes via CSG booleans, draw‑and‑extrude sketches, and import existing STLs to remix. Exports watertight STL files ready for any slicer. No subscription, no cloud, everything on‑device.
Active developmentYour personal normal, learned from your own data.
A native iPadOS app that learns what “normal” means for every metric your iPhone and Apple Watch Ultra record, surfaces the signals that move together, and explains today’s picture — goals sized to your age, height, weight, and sex. Full HealthKit integration. Nothing leaves the device.
Active developmentVantage (Swift remote desktop) — and more not yet named. Always something in the forge.
Follow along ↗The same kind of lab the open models in your Ollama library come from — pointed at the layer around the model. Distributed inference on consumer silicon. A hardware deck for driving AI agents. An IDE that runs your own models. A DGX cluster you command from your phone. Different problems, same instinct: pull it apart, understand it, and wire it up in a way no one bothered to before.
Every project starts as a question and ends as a paper. Fifteen technical white papers and four academic research papers — on federated inference, on‑device AI, model interpretability, mechanistic interpretability, and the tooling around all of it.
Read the papers →We like hard problems and people who’d rather build the future than wait for it. The channel is always open.