Projects Case study
pocket/grammer: an agentic AI coding IDE that fits in your pocket.
A native iOS IDE where an AI agent plans, writes, and revises code step by step — running local models on the phone's GPU, with the option to bring your own cloud key. Live on the App Store.
The product
pocket/grammer is a Developer Tools app: a workspace editor and an agentic AI coder in one native iOS app. The agent works deliberately — a conversational clarify phase, plan generation, a plan lint gate that bounces bad plans back before any code is written, then step-by-step execution with per-step review cards and per-card undo backed by a step-history ledger. The workspace side is a real editor: file tree with tabs, syntax highlighting, project-wide search, a diagnostics bar, and a JavaScript sandbox runner.
Local models, on a phone
The interesting engineering constraint: inference runs on-device. pocket/grammer does local GGUF inference via llama.cpp on the Metal GPU — probing and splitting layers, tuning thread counts, and using grammar-constrained tool calls so the model's tool use stays parseable. There's a model browser with a HuggingFace catalog and download manager, and separate plan/code role models tiered by device RAM. Your code doesn't have to leave the phone to get written.
Cloud when you want it
On-device isn't a religion, so the app also supports bring-your-own-key: connect any OpenAI-compatible provider for cloud-powered sessions, with keys stored securely on the device and redaction built in. Local first, cloud optional — the user picks the tradeoff per session.
The phone becomes the server
The part I like most: the app runs an on-device HTTP and WebSocket server with a per-install auth token, and pairs with your computer to serve a full desktop coding UI — Monaco editor, terminal panel, project search — powered entirely by the phone's models. External IDEs can even hit OpenAI- and Ollama-style endpoints, so the iPhone doubles as an inference box for the rest of your desk.
Shipped end to end
Designed, built, and shipped solo — from empty Xcode project to a product anyone can download on the App Store. Design, native Swift implementation, on-device ML plumbing, and App Store distribution, all one person.