G-Code Lite is coming to iPad. See what it built

Language is structure. Expression is what its pathways allow.

G-Code Lite puts an AI coding studio on your iPad. Its agent writes, builds and checks the code. Every code change it makes waits for your review.

It started with a second‑hand iPad.

One Monday night there was nothing real to code on it, and TN Lee said a word we won't print here. Sixteen hours later, G-Code Lite was coming soon.

  1. A second‑hand iPad, bought to write code on. The AI coding app on offer only remote-controlled code running in the cloud. One unprintable word later: build the real thing.

  2. The one rule: you bring your own API key. “Your API keys and code never touch my servers, because I don't have any.”

  3. The first screen. It was ugly.

  4. A test build running on the iPad itself.

  5. The first website built inside it: an AI-built* shoe shop. Rough, but the loop worked. The model writes, you review, the iPad previews.

  6. A larger model and three design skills rebuilt it for $1.49. A budget meter went on the list.

  7. React builds ran on the iPad. npm packages installed locally, and projects kept building offline. MCP connected to Cloudflare's documentation server.

  8. It got a name: G-Code Lite.

  9. “G-Code Lite. Coming soon.”

“I built this for myself. The $99 Developer fee just saves me from re‑signing every seven days.”

“It was going to be $19.99. I dropped it to $3.9. Not as a favor, just to save myself the refund paperwork.”

TN Lee, who made G-Code Lite

* Tasks built and components tested using the Claude API.

A whole studio, in the iPad you already carry.

Describe a site. The AI builds it, then checks its own work.

Condensed from a real session on iPad Air: Cerro Nublado, an AI-built* five-page site, about 30 minutes of agent time for $5.76 of API usage.

* Tasks built and components tested using the Claude API.

It looks at its own pages.

The AI renders its pages on the iPad, reads the screenshot and the console, then fixes what it sees.

90+ looks.
123+ edits.

Project PRIDE, over seven rounds. Each look renders up to four screens of a page, at iPad landscape, portrait or phone width.

The Project PRIDE research page in the G-Code Lite preview, landscape
Landscape
The same page in portrait
Portrait

Ten pages. Seven rounds. $29.13

Project PRIDE is a ten-page research-lab site, AI-built* inside G-Code Lite on an iPad Air: 35 files and 4,507 lines, every page building with an empty console. $29.13 is the settled API cost; one request cut off by a provider error still holds $8.07 in the app's budget, and its real charge (at most $8.07) is unconfirmed.

A 30-second cut of the walkthrough recorded on the iPad after the build: the code, the packages, the chat, then the pages.

Twenty apps. Three rounds. 20/20

Twenty single-page apps, AI-built* on the iPad, each with a real npm package. After each round we fixed what G-Code Lite got wrong and ran all twenty again.

Round Pomodoro Tip Calculator Todo List Unit Converter BMI Calculator Coffee Landing Portfolio Countdown Solar Quiz Color Palette Word Counter Weather Cards Pricing Table FAQ Accordion Recipe Card Stopwatch Memory Game Quote Generator RSVP Form Habit Tracker
Round 1 9/20 $0.57 Pomodoro, canvas-confetti: passed Tip Calculator, currency.js: passed Todo List, @formkit/auto-animate: failed, hit the agent's step limit Unit Converter, convert-units: failed, hit the output limit BMI Calculator, chart.js: passed Coffee Landing, animejs: failed, lost the network Portfolio, @formkit/auto-animate: failed, hit the agent's step limit Countdown, dayjs: passed Solar Quiz, canvas-confetti: passed Color Palette, chroma-js: failed, preview too slow for the test Word Counter, words-count: failed, not run, the test could not open it Weather Cards, chart.js: passed Pricing Table, animejs: failed, hit the agent's step limit FAQ Accordion, fuse.js: failed, hit the agent's step limit Recipe Card, fraction.js: failed, hit the agent's step limit Stopwatch, @formkit/auto-animate: passed Memory Game, canvas-confetti: passed Quote Generator, animejs: failed, hit the agent's step limit RSVP Form, zod: failed, hit the output limit Habit Tracker, dayjs: passed
Round 2 17/20 $0.73 Pomodoro, canvas-confetti: passed Tip Calculator, currency.js: passed Todo List, @formkit/auto-animate: passed Unit Converter, convert-units: failed, runtime error from Node's global BMI Calculator, chart.js: passed Coffee Landing, animejs: passed Portfolio, @formkit/auto-animate: passed Countdown, dayjs: passed Solar Quiz, canvas-confetti: passed Color Palette, chroma-js: passed Word Counter, words-count: passed Weather Cards, chart.js: passed Pricing Table, animejs: failed, cut off at request 10 FAQ Accordion, fuse.js: failed, lost the network Recipe Card, fraction.js: passed Stopwatch, @formkit/auto-animate: passed Memory Game, canvas-confetti: passed Quote Generator, animejs: passed RSVP Form, zod: passed Habit Tracker, dayjs: passed
Round 3 20/20 $0.70 Pomodoro, canvas-confetti: passed, $0.0252 in 3:52 min Tip Calculator, currency.js: passed, $0.0381 in 3:57 min Todo List, @formkit/auto-animate: passed, $0.0263 in 4:20 min Unit Converter, convert-units: passed, $0.0587 in 4:55 min BMI Calculator, chart.js: passed, $0.0439 in 5:43 min Coffee Landing, animejs: passed, $0.0379 in 5:00 min Portfolio, @formkit/auto-animate: passed, $0.0397 in 4:53 min Countdown, dayjs: passed, $0.0218 in 3:43 min Solar Quiz, canvas-confetti: passed, $0.0331 in 4:33 min Color Palette, chroma-js: passed, $0.0363 in 4:56 min Word Counter, words-count: passed, $0.0287 in 4:30 min Weather Cards, chart.js: passed, $0.0347 in 4:58 min Pricing Table, animejs: passed, $0.0408 in 5:01 min FAQ Accordion, fuse.js: passed, $0.0261 in 4:43 min Recipe Card, fraction.js: passed, $0.0664 in 5:41 min Stopwatch, @formkit/auto-animate: passed, $0.0251 in 4:47 min Memory Game, canvas-confetti: passed, $0.0243 in 3:58 min Quote Generator, animejs: passed, $0.0258 in 5:10 min RSVP Form, zod: passed, $0.0445 in 6:09 min Habit Tracker, dayjs: passed, $0.0258 in 5:26 min

Round three passed all twenty in one run: 98 minutes, $0.70 in total, $0.035 an app. Between rounds, eight fixes went into G-Code Lite: a package guide for npm_install, a 64K output cap, request limits shared by the agent and its build check, an automatic build check, a faster local builder, Node's global mapped for the browser, and clearer console and error messages.

* Tasks built and components tested using the Claude API. The PRIDE lab and the figures inside it were invented by the model. Costs are API usage at Anthropic's list prices, on your own API key. Runs on G-Code Lite 0.10.0 test builds (17 to 20) on an iPad Air with M2. All three rounds are shown and only round three is scored; costs for rounds one and two leave out the amounts held for two requests cut off by lost connections. Other earlier or interrupted attempts are not counted.

Python, Node.js and C, running on the iPad itself.

A built-in terminal and three real runtimes. No server in between.

CPython 3.15.1
With pip for pure-Python packages, installed into the project.
Node.js 18.20.4
npm start, test and run, each program in a thread you can stop.
Clang 22
C and C++20 (without exceptions) compiled to WebAssembly and run on the device.
mount
Open any folder from Files, iCloud Drive or an external drive.

Four AI-built* tasks, in Python, Node.js, C/C++ and shell, written and tested on the iPad: 4 of 4 passed for $0.0139 ($0.0166 including the first C/C++ run, after which the brief was clarified).

* Tasks built and components tested using the Claude API.

You approve every code change.
You set the budget.

Your key, your provider. Keys stay in the Keychain, with no account and no G-Code server. Packages the AI installs are added to package.json right away and show in Task Output, and HANDOFF.md is saved when the budget reserve is reached; every other file it writes waits for your review.

The rest of a real toolbox.

Small things a project needs, already on the iPad.

npm on iPad
Packages from the official registry, checked with SHA-512 and built offline by esbuild 0.28.
Every page builds
Multi-page sites build all their pages, with shared scripts compiled once.
Preview offline
A live preview with its own console, even in airplane mode.
Design skills
Add taste-skill, Web Design Guidelines, UI/UX Pro Max or Redesign in one tap.
Five providers
Claude, OpenAI, DeepSeek, Qwen and OpenRouter. The agent's tools use the Claude API.
GitHub import
Public or private repositories, pinned to a commit, read-only.
MCP servers
Connect tools over Streamable HTTP. Each tool asks before it runs: once, for this task, or always.
Backups and export
Twenty verified versions, a ZIP of the project or a runnable website.

G-Code for Mac, for projects that outgrow the iPad.

A desktop agent with local models or your own keys. You decide what it may run.

  • Local GGUF Models run through a bundled llama.cpp on Metal. No Ollama or cloud key needed. Or use Claude, GPT, Gemini, DeepSeek and Qwen with your keys.
  • Manual · Plan · Accept edits · Auto · Bypass In Manual, Accept edits and Auto, commands you haven't cleared wait for your answer.
  • Terminal · Outputs A real shell and versioned outputs beside the chat.

In private development for macOS 15 on Apple silicon; developer builds are ad-hoc signed and not notarized. Ask about early access

Four businesses. Five pages each.

A gym, a shipping line, a data centre and a coffee estate, each AI-built* in G-Code Lite for $5.23 to $6.17.

* Tasks built and components tested using the Claude API. The businesses, people and figures inside these sites were invented by the model. Costs are API usage at Anthropic's list prices, on your own API key, for the final run of each site on G-Code Lite 0.10.0 test builds on an iPad Air with M2; earlier and interrupted attempts at these sites are not counted.

Coming soon to iPad.

Email us and we'll tell you when G-Code Lite is available.

Email me at launch

iPadOS 17 or later. Built for iPads with 8 GB of memory (tested on iPad Air M2). English interface. Uses your own API key.