the goal
make it cheap and safe to let an ai agent write most of the code in a real product — backend, web, mobile — without the parts silently drifting apart.
the problem
agents are fast but local. they happily add a field to a nest controller and forget the nuxt call site, or rename a flutter model and leave the server schema behind. in a normal monorepo nothing stops them — types live separately per runtime, and the first real feedback is a 500 in production.
how it solves it
one contract, three runtimes.
- openapi + asyncapi as the single source of truth for requests, events and domain models.
- codegen into typed clients and server stubs for nest (typescript), nuxt (typescript) and flutter (dart) — so a change in the spec breaks compile everywhere it matters, before the agent can ship it.
- prisma 7 for the data layer, better auth for sessions, turbo + pnpm for the plumbing.
- repo layout and readme written for agents: small, labelled surfaces, clear "do/don't" boundaries, scripts agents can actually run.
the boilerplate itself is tiny on purpose. the idea isn't "use this framework" — it's "clone this, delete what you don't need, keep the contract-first habit".
source → github.com/evgentus-cy/claude-driven-nest-nuxt-flutter-monorepo
● oss · tooling · mit