Apple Intel
★
Planner
done · 0:14
on-device
TransistorKit is a Mac app that turns a one-line idea into a real, installable app on iOS, Android, TV, the web, and your server — built in parallel by a fleet of orchestrated coding agents running a blend of AI models. You talk to one of them. The rest get to work.
TransistorKit blends four coding runtimes into one fleet — Claude Code, OpenAI Codex, Apple Intelligence on-device, and any local model you've downloaded via MLX. Then it routes intelligently: cheap roles on cheap brains, hard roles on the heavy ones, and an automatic fallback when something runs out or misbehaves.
Plug in Claude, Codex, Apple Intelligence, and any code-tuned model you've pulled from Hugging Face. Then point each role at whichever brain you trust with it — Planner on the free on-device model, Builders on Claude, Reviewer on Codex. Wire it once; the harness routes per role.
Hit the Claude weekly cap mid-fleet? The Orchestrator hands the next builder to GPT-5.5 — same brief, same project, no reset, no new conversation. You don't re-paste the spec. You don't re-explain the design. The thread stays alive and the work keeps moving.
When an agent goes off the rails — wrong tool call, looping on a question, dumping code into chat instead of writing files — the Orchestrator catches it, kills the agent, and respawns the role on a different runtime. You can step in if you want, but you don't have to.
TransistorKit isn't "a wrapper around a chatbot." It's a fleet orchestrator, a live-preview rig, an asset pipeline, and a build tracker — in one Mac app, driving real native code.
You talk to The Orchestrator. It spawns Planners to write the spec, Builders to ship the code, and Reviewers to check the result — one per platform, all running in parallel. Every agent on the board, with status, runtime, last action, and a full transcript a click away.
Four runtimes plug into the same harness — Claude, Codex, Apple Intelligence on-device, and MLX for local Hugging Face models. Blend mode lets you route each role independently. Keep the Planner on Apple Intelligence (free, no quota). Send the Builder to Claude (heavy). Stay offline when you want to.
Builders run side by side, each in their own working directory, all reading the same SPEC.md. iOS gets SwiftUI. Android gets Compose. Web gets the framework you picked. Build numbers bump automatically on every dispatch — CURRENT_PROJECT_VERSION, MARKETING_VERSION, versionCode, package.json — so a deploy is always monotonic.
TransistorBridge SDKs — Swift, Kotlin, TypeScript — are auto-injected at build time. Your app calls report(screen:) as the user navigates. TransistorKit knows which screen is on the device, so "in the Home screen, make the title bigger" actually finds the right source file instead of guessing. Debug server is embedded; no separate process to run.
When a builder needs an image, it drops a line into ASSETS_NEEDED.md. TransistorKit parses it into a gallery with one card per missing asset. Bring an OpenAI or xAI key and generate them in-app, sized per slot. The App Icon Wizard takes one 1024×1024 and emits every platform's required size. Drag-and-drop if you'd rather supply your own.
Every file lands in a real project folder you can open in Xcode or Android Studio. State persists across launches — projects, chat history, fleet sessions. A local log on your machine captures every action, ready to grep if something goes sideways. Export the whole workspace as a timestamped bundle whenever you want.
When a builder finishes, TransistorKit auto-injects a tiny SDK into every platform — Swift, Kotlin, TypeScript. The running app reports its current screen back over an embedded HTTP server. That's why "in the Home screen, make the title bigger" finds the source file instead of guessing.
The same brief lands on every platform that's running. The change shows up on your device the moment the builder finishes its turn.
import TransistorBridge
struct HomeView: View {
var body: some View {
VStack {
Text("Today's reading")
.font(.largeTitle)
}
.reportScreen("Home") // ← the bridge
}
}HomeView.swift.
Per-role routing means the cheap parts run on the cheap brain. Planner on Apple Intelligence, Builder on Claude, Reviewer back on-device. Or all four. Or none — TransistorKit doesn't care.
Heavy lifting — opus / sonnet, full file + shell tools.
Alternative cloud builder. Same file + shell scope.
Free, no network, no quota. Real tool layer.
Bring your own model. Qwen-Coder, DeepSeek-Coder-V2-Lite.
Every action lands in a local log file you can search with grep. Every file lands in a real folder. Every spawn is timestamped. Open Inspect on any agent for the full transcript — kind tag, timestamp, body. Copy it, grep it, paste it into a bug report.
Across every session, ever. Split by provider in the band.
Counted from real files on disk, not bragged from token output.
By hand at 50 LOC/working day — McConnell's solo-dev focused rate. Convert to calendar time and that's about 945 working days, or 4.5 years of weekday work.
{ "tokens": { "claude": 1894211, "codex": 523884 }, "linesOfCode": 47231, "firstSeenAt": "2026-04-12T16:08:11Z" }
When an agent notices the harness misbehaving — a missing action, a confusing prompt, a tool that's not wired right — it files a short note to the maintainer. The note shows up in our queue. We fix it. You get the fix in the next version. Sometimes the maintainer pings you back mid-session with a clarification, which the orchestrator folds into the conversation without breaking your thread.
The feature is off by default and lives behind a single switch — flip it on during the welcome wizard, or any time in Settings → Telemetry. When it's off, the entire channel is stripped from the orchestrator's vocabulary; nothing leaves the box.
TransistorKit is a single-window Mac app. Drop in your credentials, pick a brain, describe what you want to build. Everything else is real files in real folders.
Bring your own keys: Anthropic, OpenAI, xAI. Apple Intelligence works out of the box on supported hardware.