THE APPS

The apps built
by AutoDev LG

Let's be honest: they still look pretty basic.
But no human wrote a single line of their code.

Here's how it works. First, a human says in one line what app to build, like “Make me an Invaders game.” From that request, Claude Code writes a spec and hands it to AutoDev LG. From there, the design, the tests and the implementation are all done by local LLMs running on the GPUs in my home PC. No cloud AI is used at that stage.

(In the future, writing the spec will be handed to AutoDev LG itself, too.)

Both are small Windows desktop apps that run on nothing but Python's standard library (Tkinter). They don't even touch the internet. Plain? Yes. But they work. (Well, one of them does.)

taskboarda task board

✅ 3 PASSes in a row

Plain on the outside. Strangely paranoid on the inside.

The taskboard window: a list of task IDs, titles and statuses, with Add, Change Status and Delete buttons — The real thing, built by loop68 (v142) on 9/29. The tasks shown are samples added for the screenshot.

📄 See the spec given to the local LLMs →

What it does

Add and delete tasks
Switch status: todo → doing → done
List tasks by ID, title and status (sorted by ID)
Nothing is lost when you close it (saved to a JSON file). Japanese titles work too
Weird input, like an empty title, gets an error pop-up

Where it's surprisingly serious

Saves in a way that can't corrupt the data file even if the PC dies mid-save (write a temp file, flush it to disk, then swap it in)
If a save fails, it rolls everything back to how it was before the operation
A deleted task's number is never reused
Empty, broken or garbled data files are detected and refused

Hold on: three buttons, mostly gray. Yet it guards its data like a bank. The spec said so, and the local LLMs followed it to the letter.

~360

lines of app code (4 files)

471

lines of tests (longer than the app)

54 min

for the latest automatic build

17 / 89

PASSes so far

89 attempts so far, 17 PASSes. The first PASS came on 9/14, and on 9/26 it finally made 3 in a row. Now the work is keeping “PASS” the normal outcome while cutting the time it takes.

INVADERSan Invaders game

🚧 In progress (0 PASSes yet)

Red squares invade. You are a green square. ...It's mostly squares.

The INVADERS prototype: on a black background, red square invaders in 4 rows of 8, yellow bombs falling, and a green square player at the bottom — The 9/27 prototype (inv16), actually running. Yellow squares are bombs; the white dot is your shot.

📄 See the spec given to the local LLMs →

How to play (from the spec)

← → move Space fire P pause R restart

A 20 × 15 grid with invaders in 4 rows of 8
The invaders march sideways, drop one row at the edge and turn around (you know the move)
One shot on screen at a time. A hit is worth 10 points
Game over when your lives hit 0 or the invaders reach the bottom. Wipe them all out to win
The high score is saved to a file, and the random seed that picks which invader drops a bomb is fixed so tests can reproduce it

Where it's struggling right now

15 attempts so far, 0 PASSes
The prototype in the picture failed 14 of its 45 tests, with bugs like “the score doesn't go up when you hit an invader” and “wiping out every invader ends in game over”
Lately it often stumbles even earlier, at the “write the tests” stage. Getting past that is the next wall

Hold on: destroy every invader and... “Game Over.” A bold new take on game design. ...Yes, it will be fixed.

0 / 15

PASSes so far

○○○

goal: 3 PASSes in a row

With far more state and movement than taskboard, this is the assignment that really tests the framework. On the day it finally PASSes, expect loud celebrations in the dev log.

Someday, just “make me an Invaders game,”
and out comes a game you can play.

That's the end goal of AutoDev LG. We're only partway there. Come watch these plain little apps slowly turn into proper ones.

See the dev log →How it works →