Nocobase bets AI agents should extend apps, not generate them

Nocobase tells agents to extend a production app instead of generating one from scratch, per its GitHub README.

That's the gap that kills generated apps after the demo, when auth, permissions, and data migrations actually have to hold up. Nocobase has picked up roughly 23,400 stars on GitHub arguing its scaffold already clears that bar.

Its README names v0 and Lovable-style tools as the full-generation approach. Those tools write an entire app from a prompt, and they're good at producing something that looks finished in a demo. Where they struggle is the parts users don't see until week three. Who can access what. How data moves when the schema changes. What happens when a table needs a new required field on records that already exist.

I've hit this wall building solo products. Generating a CRUD app from a prompt is the easy 80%. The rest is permission boundaries, migrations that don't corrupt existing rows, and auth that survives a schema change. Generated code falls apart there because it was never built against a real data model that has to persist.

Scaffold versus generate

The pitch is that an agent working against Nocobase's scaffold inherits those systems already solved, instead of reinventing them badly under a deadline. What the README doesn't say is how much of an agent's output still has to route through Nocobase's existing data model. That's the real limit on how far this approach stretches. Scaffold-versus-generate is the axis I'd use to judge any agentic app builder now, and it's not one most comparisons draw.