# AI writes code fast — but who keeps it maintainable? An open-source bet on metadata > Source: > Published: 2026-06-05 06:22:27+00:00 In 2026, AI writes code at incredible speed. Cursor and Claude Code agents read whole repos, the CLI became the new dev infrastructure, and "a team of agents working together" is the story of the year. But there's an uncomfortable truth: **AI writes fast, yet enterprise apps still fail to ship — maintainability gets worse, not better.** Why? For a toy app, letting an AI agent one-shot the code is great. For an enterprise system, you need: The problem with AI agents writing *code*: they emit a wall of one-shot code that immediately becomes a second source of truth alongside your project. The next agent run overwrites your hand edits; multiple agents each write their own thing. **The faster it goes, the faster the debt piles up.** AI gave us speed. Nobody gave us *scale* — the standards, boundaries, and order. And enterprise apps die on exactly that. [Oinone](https://github.com/oinone/oinone-pamirs) is an open-source, **100% metadata/model-driven** low-code framework. The bet: **data models, UI, permissions, workflows, AND the AI's output all live in one shared metadata model.** So an AI agent doesn't write throwaway code — it writes into the **same metadata** the framework and human developers already operate on: **AI for speed, the framework for scale.** That's the whole idea. ``` curl -L https://gitee.com/oinone/oinone-docker-shared/raw/master/oinone/docker-compose.yml -o docker-compose.yml docker compose -p oinone up -d # open http://127.0.0.1:88 user: admin password: admin ``` Then have the AI generate an app from a sentence, and look at what it produces — a **metadata diff**, not a pile of code. That's the difference between "AI-native" and "a low-code tool with a chatbot bolted on." Stack: Java backend + TypeScript frontend, **AGPL-3.0** (genuinely open source, the framework you run is the framework that's public). Self-hostable; in production at large enterprises. English docs are catching up; the polished 7.x live demo lands soon (the quickstart above is the real thing today). It's **not** for simple internal tools — use something lighter for those. It shines on complex, long-lived, self-hosted enterprise systems. If the "metadata as the single source of truth for AI + humans" idea resonates, a ⭐ on [GitHub](https://github.com/oinone/oinone-pamirs) (or [Gitee](https://gitee.com/oinone/oinone-pamirs)) helps more developers find it. Happy to discuss the metadata model, the AGPL choice, and how it compares to Retool/Appsmith/Budibase in the comments.