Worldant is an actively in-progress runtime for AI-owned software worlds.

It is being rebuilt around a directory-first model: an agent edits files, the daemon validates and builds a .dist snapshot, and the running world exposes typed commands, mutations, workflows, and reactive streams.

This documentation describes the target architecture. APIs are not stable yet.

Why it exists

AI should be able to expand its ability by creating real software capability: schema, logic, durable jobs, reactive views, and typed interfaces. The result must still be normal software. It must have a database truth, authorization, crash recovery, deploy safety, and a way for humans and generated UI to use the same surface.

Worldant is that runtime boundary.

What Worldant owns

• World directory discovery and validation.
• SQL migration validation and schema reflection.
• PgPaw/Postgres as database truth and realtime data plane.
• Commands for quick read/compute logic.
• Mutations for committed writes.
• Durable workflows and replay-safe steps.
• Reactive SQL subscriptions and cursor-backed event streams.
• A live npm remote-function protocol.
• MCP and HTTP surfaces for outside agents.

Worldant does not own generated UI. Ragent owns UI generation and consumes Worldant through the same typed package/protocol as any other client.

Core rule

All durable truth enters PgPaw/Postgres.
All visible clients subscribe back from PgPaw/Postgres.

Commands may compute. Mutations may commit. Workflows may wait and resume. Reactive streams may notify UI, agents, and other consumers. None of those paths becomes a hidden state system outside the database.

Shape

world/
  migrations/
  generated/
  commands/
  mutations/
  workflows/
    steps/
  reactive/
  tests/

worldant build
  -> .dist/run/<run_snapshot_id>
  -> PgPaw metadata
  -> generated npm contract

worldant run
  -> daemon serves latest valid snapshots
  -> workflow runs pin to their snapshot

Status

Worldant is actively in progress. The current proof target is:

• A generated chess world with schema, mutation, command, workflow, and reactive board subscription.
• A live npm package that exposes the world as typed functions.
• A Ragent UI that imports that package and interacts through committed database truth.
• A long-running workflow that survives restart and resumes from PgPaw state.

See the repository thesis.md for the current design details.