AutoBE

AutoBE is an AI backend builder developed by wrtnlabs, with a core innovation in its compiler-driven code generation strategy rather than direct code output. The system contains 40+ specialized AI Agents collaborating in a five-stage waterfall process: Requirements Analysis, Database Design, Interface Design, E2E Test Generation, and Implementation.

Mechanically, AutoBE has AI construct language-agnostic ASTs first, then validates them through a three-layer compiler pipeline — Prisma Compiler, OpenAPI Compiler, and TypeScript Compiler — before producing type-safe NestJS backend code and a companion Client SDK, claiming 100% compilation success across project types.

The system supports 13+ LLMs including Claude Sonnet 4.6, GPT-5.4, GLM-5, Kimi-K2.5, and Qwen3.5 series, with an automated Benchmark evaluation pipeline covering four project types (Todo, Reddit, Shopping, ERP) measuring compilation correctness, requirement coverage, test coverage, and more. Each generation stage can be stopped independently for flexible control.

Typical Scenarios: Rapid prototyping (Todo ~53min, Reddit ~91min), no-code/low-code backend construction, automated requirement documentation, database design assistance, API specification and test automation, frontend-backend SDK bridging, LLM code generation benchmarking.

Known Limitations: High token consumption (250M+ for E-commerce scenarios), potential runtime errors (v1.0 targets 100% runtime success), no post-generation maintenance capability, AI designs may differ from user expectations requiring manual review.

Architecture Overview:

Frontend Playground (Vite, port 5173)
        ↕ WebSocket / RPC
Backend Service Layer (NestJS Server / Node.js Server)
        ↕
Agent System (40+ specialized Agents, orchestrated by waterfall stages)
        ↕
AST Intermediate Representation (language-agnostic Abstract Syntax Trees)
        ↕
Compiler Validation Layer (Prisma Compiler → OpenAPI Compiler → TypeScript Compiler)
        ↕
Generated Artifacts (NestJS backend + Prisma Schema + type-safe Client SDK)

Quick Start:

git clone https://github.com/wrtnlabs/autobe --depth=1
cd autobe
pnpm install
pnpm run playground

Visit http://localhost:5173 after startup and describe backend requirements via the chat interface. Replay is available at http://localhost:5173/replay/index.html.

Benchmark CLI:

corepack pnpm estimate                                        # All models
corepack pnpm estimate -- --model kimi-k2.5                    # Single model
corepack pnpm estimate -- --project todo                        # Single project
corepack pnpm estimate -- --model glm-5 --project shopping      # Combined

Unconfirmed Details: LLM Provider configuration format not detailed in README; distinction between NestJS Server and Node.js Server unclear; RAG + Hybrid Search implementation status unconfirmed; Java/Spring code generation has no timeline; production deployment guide pending.