Stellium · v0.0.0 Alpha

The schema lives
in the data.

Stellium is a TypeScript framework where type definitions, field definitions, and the composition edges between them are nodes in the same graph as your data. Schema migrations become a reconcile operation. LLM agents discover the data model by querying it.

One codebase runs against SQL (SQLite, Postgres, libSQL), Neo4j, in-memory JSON, or Markdown over the filesystem — same DSL, backend-appropriate emission.

Source and README: link coming soon

Pre-1.0. APIs change without notice. Do not use in production.

const Note = type('Note', {
  schemaId: '@notes',
  fields: {
    title:   field.string({ required: true }),
    content: field.string({ required: true }),
    tags:    field.string({ multiValued: true }),
  },
});

// the type definition above is itself a node
// in the same graph as your data.

Core commitment

One graph for schema and data.

TypeDefinition, FieldDefinition, and the COMPOSES edges between them are nodes — queryable with the same primitives as the records they describe. Most frameworks keep the schema in code and the data in a database. Stellium keeps both in the graph.

01
Schema migrations are diffs.

Compare your declared types against the live graph, classify each change as additive, behavioral, or destructive, and apply in dependency order inside a transaction. No hand-written migration scripts; the engine reasons over the same data the application reads.
02
The system describes itself.

An LLM agent queries TypeDefinition the same way it queries your data. The schema is discoverable by walking the graph — there is no parallel metadata channel and no documentation that can silently drift out of sync.
03
Export carries the schema.

Because the schema is data, exporting the data exports the schema with it. Sovereignty is structural: data ports out with its type definitions attached, which is what makes "leaving" actually work rather than rhetorically work.

The shape of the work

Declare a type. Generate the class. Query it.

The DSL on the left registers a type with Stellium. Running the code generator produces a typed domain class with CRUD, relationship accessors, and a query builder — the snippet on the right uses it. The query compiler emits Cypher for Neo4j, SQL for the SQL backend, and native predicate evaluation for JSON and Markdown — same input shape, backend-appropriate output.

import { type, field, relationship } from '@stellium/schema';

const TAGGED_WITH = relationship('TAGGED_WITH', {
  schemaId: '@notes',
});

const Note = type('Note', {
  schemaId: '@notes',
  fields: {
    title:   field.string({ required: true }),
    content: field.string({ required: true }),
    tags:    field.string({ multiValued: true }),
  },
  relationships: {
    topics: {
      ref: TAGGED_WITH,
      target: 'Topic',
      direction: 'outgoing',
      cardinality: 'many',
    },
  },
});

import { Note } from './generated/domain.js';

// Create
const note = await Note.create({
  title: 'Graph databases',
  content: 'Property graphs store key-value pairs...',
  tags: ['databases', 'graphs'],
});

// Type-safe query builder
const recent = await Note.query({
  title: { contains: 'Graph' },
  $orderBy: { createdAt: 'desc' },
  $limit: 20,
}).exec();

// Relationship traversal
const topics = await note.getTopics();

Intermediate representation

A backend-agnostic IR, four backend-specific emitters.

A query you write is a plain TypeScript object: shape predicates, ordering, projections, traversal. On every terminal call, the classifier partitions it into a QueryIR — a backend-agnostic plan whose only backend-aware input is the type's resolved storage mode. The active engine reads that IR and emits its native form: Cypher for Neo4j, parameterized SQL against real typed columns for SQLite, Postgres, and libSQL, or a direct predicate walk for the in-memory JSON store and the Markdown sidecar index. No lowest-common-denominator; each emission is what the backend is good at.

Per terminal call: classify into IR, then engine-specific emit. No plan cache.

await Note.query({
  title: { contains: 'Graph' },
  $orderBy: { createdAt: 'desc' },
  $limit: 20,
}).exec();

Cypher Neo4j

MATCH (n:Node)
  -[:HAS_TYPE]->
  (:TypeDefinition {name: 'Note'})
WHERE n.schemaId = '@notes'
  AND n.title CONTAINS 'Graph'
RETURN n
ORDER BY n.createdAt DESC
LIMIT 20;

SQL SQLite · Postgres · libSQL

SELECT n.*
FROM node n
WHERE n.schema_id = '@notes'
  AND n.type = 'Note'
  AND n.title LIKE '%Graph%'
ORDER BY n.created_at DESC
LIMIT 20;

Predicate walk JSON · Markdown

nodes
  .filter(n =>
    n.schemaId === '@notes'
    && n.types.has('Note')
    && n.title.includes('Graph'))
  .sort((a, b) =>
    b.createdAt - a.createdAt)
  .slice(0, 20);

Emissions above are illustrative. Real output depends on the active storage mode, indexes, and the query's classification; the SQL backend uses real typed columns and idiomatic predicates, not JSONB. The Markdown engine intentionally rejects a few IR features (raw fragments, relationship projections) and throws a typed BackendCapabilityError when they appear.

Backend profiles

Four engines. Different workloads, same surface.

With the IR in place, backend choice becomes a deployment decision rather than an architectural one. Each engine has a profile — operational footprint, query-shape strengths, persistence model — so pick whichever fits the workload. Backend selection is GraphConfig.type at initialization, and a native-query escape hatch is available for the rare case the IR doesn't cover.

SQL
SQLite · Postgres · libSQL

Suggested default

Real typed columns with B-tree indexes, not JSONB. Familiar operational surface for any team that already runs Postgres. The query compiler emits standard SQL — you can EXPLAIN it.
Neo4j
Cypher · property graph

When the graph is the query pattern

Native graph storage and traversal. Combined relationship + field queries measured at 6.6× faster than property-store mode at 100K nodes. The right choice when multi-hop reasoning is the workload.
JSON
In-memory · optional file persistence

Tests, demos, scratch graphs

Fast and ephemeral. Same query surface and same DSL as the production backends. Useful for tests that need a real graph and for early-stage projects that have not yet committed to a database.
Markdown
Filesystem · sidecar index

Portable, git-trackable, hand-editable

A graph backed by a folder of markdown files. Editable in any text editor while the framework continues to read it. Portable across machines without a migration step — the data is files.

Same DSL, same query builder, same domain classes — the engine selection is one line of config.

The strictness dial

Loose to strict by configuration.

The same codebase operates as a forgiving knowledge graph at one setting and as a strict ORM at the other. The transition is a configuration change, not a migration — the data does not move, only what the system accepts on the way in.

Forgiving

Loose mode

Missing required fields surface as warnings.
Nodes can borrow fields from types they do not carry.
Composition conflicts are tolerated.

Optimized for exploration and early-stage data. The schema is a guide, not a contract.

{
  enforceRequiredFields: false,
  allowBorrowingFields:  true,
  noShadowing:           false,
}

Enforced

Strict mode

Required fields enforced at write time.
Field borrowing rejected.
Composition conflicts caught at schema registration.

Optimized for machine reliability and production APIs. The schema is a contract.

{
  enforceRequiredFields: true,
  allowBorrowingFields:  false,
  noShadowing:           true,
}

Per-schema and per-type. Information can start loose and become strict as it matures — without changing tools.

Honest status

What's built, what's designed, what's not yet.

Stellium is a personal infrastructure project under active development. The split below tracks what exists in src/ today, what has settled architecturally but is only partially wired, and what lives only in design documents so far.

Built and operational

Schema DSL and code generation
Type composition (multi-type nodes, field borrowing, provenance)
Query compiler (classify → compile → stitch)
Reconciliation engine (diff → classify → cascade → apply)
Four pluggable backends (SQL, Neo4j, JSON, Markdown)
Dual storage modes (graph + direct) behind a uniform contract
Permission engine (RBAC + ABAC, circle-based)
Plugin system with capability gating
MCP server with progressive tool disclosure
DID-based auth, sessions, handles, rate limiting
CLI: init, generate, diff, apply, status
Data Coordinator (reactive provider/consumer wiring)

Direction settled, partial

Mutation pipeline (VECTOR Spine and Vertebrae)
Hook umbrella and Subscriber model
Three-layer auth decomposition
Configurable strictness extensions

Designed, not yet built

Presentation layer (intent-driven rendering, 30+ design docs)
Federation between Stellium instances (892-line spec)
Multi-instance isolation deployment
Real-time collaboration (CRDTs)
Reactive transformation chains (Currents)

Where to go next

Read the design. Read the code.

Everything below lives in the framework repository; the link is coming soon. Documentation is honest about which parts are settled and which are still moving — the Overview is the best single entry point if you want the comprehensive picture.

README Quick start, high-level orientation, and the maintainer's own framing of what the project is. Link coming soon
Stellium Overview The comprehensive system picture — principles, subsystem index, what's built versus designed. Link coming soon
Getting Started End-to-end tutorial that builds a personal knowledge management system through the CLI. Link coming soon
Architecture Invariants, layer boundaries, the four-layer API stack, and the rules that the codebase commits to. Link coming soon
Philosophy Compressed reasoning patterns and design biases. The why behind the architecture. Link coming soon
Source The full source. Not open to outside contributions. Link coming soon

The schema livesin the data.

One graph for schema and data.

Schema migrations are diffs.

The system describes itself.

Export carries the schema.