Vibe coding meets waterfall: AI backends that actually compile

What it does

AutoBE is a chat-based AI agent that generates complete NestJS/Prisma backends from natural language requirements. It runs a rigid waterfall pipeline—requirements analysis, database design, API specification, E2E tests, and implementation—using over 40 specialized agents. The output includes documentation, an entity relationship diagram, type-safe DTOs, and a client SDK.

The interesting bit

Instead of asking LLMs to write TypeScript directly, AutoBE makes them assemble language-neutral ASTs that are validated by built-in compilers before any code is generated. This pre-flight structural check aims to catch schema mismatches and type errors at the conceptual level, turning vibe coding into something closer to a formal build process.

Key highlights

• Benchmarks itself against 13+ LLM models across four project types, grading compilation correctness, test coverage, and security
• Top scorer glm-5 averages 86/100 (grade B); most models hover in the B-to-C range across the suite
• Generates a type-safe client SDK alongside the backend for frontend integration
• Supports stopping at any pipeline phase if you only need, say, an ERD or API spec
• Currently TypeScript-only, though the AST pipeline is designed to eventually support other languages

Caveats

• AGPL-3.0 license may complicate commercial adoption without open-sourcing derivative work
• Runs locally via a playground application; no hosted service is mentioned
• Benchmark grades top out at B for most models, with several scoring C on complex projects like Shopping

Verdict

A solid experiment if you want AI-generated backend scaffolding that values structure over chaos, particularly for NestJS/Prisma stacks. Skip it if you need a hosted platform or a language other than TypeScript today.