Technical note · Published on 25 September 2026
The Forge, a software factory that doesn't work yet
A chain of local AIs that invents, plans, writes and publishes an application without me. What it does, where it fails, and why I show it anyway.
I just wanted to test something. Can a chain of AIs running locally, on my graphics card through Ollama and with no paid API at all, invent an application, plan it, write it and publish it without me touching the keyboard?
The setup fits in one line. CrewAI imagines, Ollama plans, Aider builds, TypeScript checks, GitHub archives, Telegram keeps me posted, and a systemd service runs the factory as soon as the PC boots.
Two agents start by talking. The Ideator picks a web SaaS or a mobile app and produces massive specs, 12 to 18 features, 15 user stories, 8 to 12 entities, 25 to 35 endpoints. The Architect derives a tree of 80 to 120 files from it. Then a planner splits everything into 320 tickets, in 14 batches in the order of a real project, each saved as it goes so a failed batch never loses everything.
Aider then runs the 320 sprints one by one, thirty minutes maximum each, with a shared memory prefixed to every ticket. As soon as a package.json appears, npm install runs, the TypeScript compilation is checked, and the errors go back to Aider in a self-repair cycle. At the end, a private repository is created and "The Forge Bot" commits.
A maintenance mode hands an existing project to Open Interpreter, which analyses, launches, repairs, and writes what it learned into the memory. That's why a dedicated, isolated machine is the right approach, this mode genuinely manipulates the system.
The answer to my question is "not yet". The project is experimental and unfinished, and I show it for the approach, not the result.