doorman11991/marrowscript

MarrowScript is a small programming language and compiler aimed at backend developers who want to describe an entire system in one file and get a full Node.js backend out the other end. You write a .marrow file using a few high-level building blocks: entities for data, capabilities for operations, events, flows for multi-step sagas, channels for websockets, and policies for things like rate limits. The compiler then produces a project with an Express API, PostgreSQL migrations, JWT auth, a TypeScript SDK, OpenAPI spec, React hooks, an admin panel, Dockerfile, and GitHub Actions CI. The same input always produces the same output. To use it, you install the compiler from npm with npm install -g marrowscript-compiler and you need Node.js 18 or later. A typical flow is marrowc init my-app --domain saas_platform to scaffold a project, marrowc compile to generate the backend, then docker compose up -d and npm run dev to start it locally on port 3000 with an admin panel at /admin. Postgres and Redis start automatically from the generated docker-compose file. Entities have fields with types like string, uint, uuid, json, list, and set, plus constraints, state machines, and relations. Sensitive fields can be tagged so they are redacted from logs, and renamed fields produce safe ALTER TABLE migrations. Capabilities declare preconditions, effects on entity fields, events to emit, and how they should run, including transactional or eventual mode, timeouts, idempotency, and retry behavior. Events go through a transactional outbox with delivery modes like at_least_once, at_most_once, and exactly_once. Flows define sagas where each step has a compensating action that runs on failure. The README also covers using the language with large language models. There is a suggested prompt for getting a model to write .marrow files, and two reflect commands that read an existing TypeScript codebase and produce a .marrow stub, either statically or with help from an LLM. MarrowScript can also call LLMs from inside the generated backend: you declare a model with an OpenAI-compatible endpoint and a prompt block with retry, cache, and validation rules, and the compiler emits the adapters, retry logic, and trace tables.