Everything in the Cyber ecosystem lives in one graph.
Concepts, skills, roles, courses, labs, library topics and the people learning them. Cyber Library, Cyber Courses, CyberCTF and Cyber Bench are all clients of that one graph. "What do I need to learn before this?" is a walk along REQUIRES edges. A graph database is the natural home for that, so the backend runs on Neo4j, and Neo4j only.
The database was never the problem. The TypeScript around it was.
What it looked like without an ORM
Three things kept hurting.
Every row was a guess. The driver hands you records, not types. Each value is untyped until you cast it, and integers come back as the driver's own Integer class. I ended up writing a wrapper whose whole job is to make every column unknown instead of any, plus a toInt helper to read numbers. Nothing tells you that u.emial is a typo until production does.
No migrations. The GraphQL layer doesn't manage constraints, so the backend creates them itself: a hand-written list of about 25 CREATE CONSTRAINT strings, run on cold start. Renaming one means writing the DROP by hand, too.
No Prisma. On Postgres I'd reach for Prisma without thinking. But Prisma doesn't support Neo4j: the request was closed as not planned in December 2024. The other Neo4j libraries each solve part of the problem. None gave me a schema, a typed client and migrations in one place.
I had already written a Neo4j adapter for Better Auth. Every new app meant writing the same glue again.
So I built the tool I wanted
Cyphra is a typed ORM for Neo4j. You describe the graph once:
node User {
id String @id @default(cuid())
email String @unique
posts Post[] @relationship(type: "AUTHORED", direction: OUT)
}
cyphra generate writes a client that knows every property and relationship:
const active = await db.user.findMany({
where: { posts: { some: { title: { contains: "graph" } } } },
select: { email: true },
});
active is typed from the select. Misspell email and it's a compile error, not a 3 a.m. bug.
What you get
| Feature | What it replaces |
|---|---|
One .cyphra schema | Constraints scattered in code |
| Generated typed client | unknown rows and manual casts |
cyphra migrate dev | Hand-written CREATE CONSTRAINT lists |
| Nested writes in one transaction | Multi-step Cypher by hand |
db.$cypher | Nothing. Raw Cypher stays one call away |
cyphra dev | Setting up Neo4j locally: it starts Neo4j 5 in Docker and writes your .env |
cyphra pull | Rewriting a schema for an existing database |
The API will feel familiar if you've used Prisma: findMany, include, select, upsert, $transaction. That's on purpose. The goal is to make Neo4j as easy to pick as Postgres.
Not a wrapper around Cypher
Graphs are the point, so relationships are first-class: relationship properties are typed too, and connect won't create duplicate edges. When the ORM isn't enough, you write Cypher. Cyphra doesn't hide the database, it removes the boilerplate around it.
Hosting
The library and CLI are free and MIT. If you don't want to run Neo4j yourself, there's Cyphra Cloud in early access, and the Kubernetes operator behind it is open source too. Self-hosting stays first-class.
Try it
- Website and docs: cyphra.dev
- Code: github.com/Cyphra-dev/cyphra (MIT)
It's young, so tell me what breaks. Issues and PRs welcome.
Ship the fix. Then share the patch.
