A Local Open-Weight Model Builds Its First Web App from a GitHub Issue

By Joel Dare - Written September 3, 2026

Today I had a breakthrough.

I watched as a small web app materialized, automatically, from a free large language model (LLM) running on my laptop.

It started with some stray thoughts, which I wrote into a new GitHub issue, describing a test app I had in mind.

I labeled the issue with “gemma”.

My router (a shell script) noticed the label and made a call to Pi, the minimal coding agent.

Pi read my static instructions then connected to an LLM running locally.

But, I wasn’t excited yet.

This was the planning pass. I had designed it into the system. It’s default mode.

It’s easy for a local model to write prose so it didn’t surprise me when it returned with a well-written plan based on the stray thoughts I had aimlessly typed into the issue description.

The complete plan it had written looked good to me.

I labeled the issue again. This time I selected “gemma” and “project”.

Again, my router noticed the labels and made a call to Pi.

About five minutes later the model wrote a comment on the issue telling me a prototype was ready.

Honestly, I still didn’t think it was going to work. I doubted it would run, if it existed at all.

I opened the link and found a working web app exactly as the plan had described.

It was a tic-tac-toe game. The simplest example I could think of without being completely trivial.

The game wasn’t perfect. There were only two rows of three squares for X’s and O’s. I needed a 3x3 grid. But I quickly realized this was a typo in my original request, not an error the model had made. I had asked for a 3x2 grid.

I added a comment to the GitHub issue outlining several changes, including fixing the grid.

I labeled the issue as “gemma” and “project” and watched as it picked it up again, made my changes, and then commented that the new prototype was ready.

Tic-tac-toe board the local model generated

The Dark Factory Series

Terms - Privacy