The Forge

A fully autonomous, 100% local software factory. Launched on a dedicated Linux PC, a chain of AIs all running on the graphics card through Ollama, with no paid API, invents an application, designs it, plans it, writes its code file by file, checks it, repairs it, then publishes it to GitHub, with a Telegram notification at each step. It can produce projects in a loop, or evolve existing ones.

In short: CrewAI imagines, Ollama plans, Aider builds, TypeScript checks, GitHub archives, Telegram keeps you posted, and a systemd service runs the factory as soon as the PC boots.

Phase 1: two CrewAI agents talk in sequence on a local model, qwen2.5-coder 14B by default. The Ideator picks a web SaaS or a mobile app and produces massive specs: name, 12 to 18 features, 15 user stories, 8 to 12 entities, 25 to 35 endpoints. The Architect derives the full tree of 80 to 120 files, each with its path, role and imports.

The planner then generates 320 development tickets across 14 thematic batches in the order of a real project: configuration, types, utilities, services, stores, hooks, base then advanced components, business components, screens, navigation, styles, tests and documentation. Each ticket is an individual Ollama call in forced JSON, saved as it goes so a failed batch never loses everything.

Phase 2: Aider runs the 320 sprints one by one in a git repository initialised for the project, with the shared memory prefixed to each ticket, the target file pre-created, thirty minutes maximum per sprint. As soon as a package.json appears, npm install runs; then TypeScript compilation is checked and, on errors, a self-repair cycle sends them back to Aider.

Phase 3: a private repository is created on GitHub, everything is committed by "The Forge Bot", and Telegram announces the project is online.

A shared-memory text file is injected into design, into every planning ticket, into every sprint and into maintenance mode: you can write directives in it by hand, and the factory records its lessons from one run to the next.

Maintenance mode hands an existing project to Open Interpreter in auto-run: analyse the code, launch the server, repair, add features, then write what it learned into memory. That is why the dedicated, isolated machine is the right approach: this mode genuinely manipulates the system.

Around it: a pre-flight diagnostic, a one-command installer, a systemd service that restarts the loop at boot, and four CrewAI tools ready but not yet wired. Experimental and unfinished: shown for the approach, not the result.