AutoBE
Turns natural-language requirements into staged, buildable TypeScript backends.
What does this agent do, and when should you use it?
AutoBE is a locally run AI backend builder for moving from a prototype toward a production backend. Through its playground chat interface, users guide a waterfall flow covering requirements analysis, database design, API design, E2E tests, and implementation. A Facade Controller coordinates more than 40 specialized functional agents, which construct and validate language-neutral ASTs before code generation. Its intended outputs include a TypeScript, Prisma, and NestJS backend; database and API documentation; test code; and a type-safe client SDK. The repository presents a 100% buildability goal, while also documenting that runtime behavior, design interpretation, and ongoing maintenance still require human work.
After pnpm run playground starts the local chat application, a user submits backend requirements in stages. The Facade Controller routes work through Analyze, Database, Interface, Test, and Realize; AutoBeDatabase validates the database phase, AutoBeOpenApi produces OpenAPI-related output for the interface phase, AutoBeTest analyzes tests, and Realize uses a Hybrid Compiler. Agents create language-neutral ASTs expressed through predefined schemas, validate them against type rules, and then generate TypeScript and Prisma code. The resulting backend includes API controllers, DTO structures, implementation logic, E2E test functions, and a type-safe SDK for frontend calls and testing.
- A non-programmer founder wants a first todo-app backend prototype, beginning with a requirements report, schema, and API specification.
- A backend team needs reviewable ERDs, controllers, DTOs, and an initial implementation for a new NestJS and Prisma service.
- A test engineer wants generated E2E test functions that exercise endpoints through a type-safe client SDK.
- An engineering lead wants to inspect benchmark results across models for Todo, Reddit, shopping, and ERP backend-generation tasks.
- A development team wants an initial generated backend to extend with coding assistants while retaining specifications and tests as maintenance inputs.
What are this agent's strengths and limitations?
- Produces a staged set of backend artifacts—requirements, database, API, tests, and implementation—rather than only an application code snippet.
- Uses AST validation and database, OpenAPI, and test compilers before code generation as its documented buildability mechanism.
- Generates a type-safe client SDK that is used for both frontend integration and E2E testing.
- Includes an evaluation pipeline across more than 13 models and four project types, with JSON report output paths.
- The repository explicitly says buildability does not guarantee runtime success; database, endpoint, and business-logic failures can still occur.
- Complex projects can consume substantial AI tokens: the documented range is roughly 30M to 250M+, with a simple todo app around 4M.
- Generated database and API designs may differ from the user's intent and should be reviewed before implementation or production use.
- Ongoing maintenance is outside the current scope; teams must handle fixes, features, performance work, and security updates after generation.
- The supplied material lacks provider-specific credential and configuration instructions, so connecting a real model may require additional setup.
How do you install or deploy this agent?
Prerequisites: Git and pnpm. Run:git clone https://github.com/wrtnlabs/autobe --depth=1cd autobepnpm installpnpm run playground
The playground is served at http://localhost:5173. The supplied material says AutoBE can use multiple LLM providers and local models, but does not document provider credentials, environment variables, or setup steps.
How do you use this agent?
Open http://localhost:5173 and describe the backend you want in the chat interface. You can request, in order, requirements analysis, a database schema, an API interface specification, E2E test functions, and API implementation, or stop after any stage. Test and benchmark chat replays are available at http://localhost:5173/replay/index.html. Run benchmarks with corepack pnpm estimate, optionally adding --model and --project.