ontogate

Declare your domain once — get both the docs and the enforcement for AI agents.

Status: pre-release. This package currently ships a version stub while v0 is under active development. The API described below is the design target, not a published surface.

What ontogate will be

The domain rules you've been hand-writing into markdown — product statuses, order invariants, "never issue a coupon for a sold-out item" — declared once in TypeScript (zod-native), compiled into two rails:

• Prompt rail — generated domain docs for AGENTS.md, so agents are guided to behave well.
• Runtime rail — a typed MCP server with staged write-actions, human-in-the-loop approval, and a tamper-evident audit log, so agents are physically stopped when they don't.

const issueCoupon = defineAction({
  name: 'issueCoupon',
  target: Product,
  input: z.object({ productId: z.string(), amount: z.number() }),
  policy: {
    approval: 'required',
    audit: true,
    where: { field: 'status', op: 'neq', value: 'soldout' },
  },
  execute: async ({ productId, amount }) => {
    /* your existing backend call */
  },
});

v0 scope (in development)

• ontogate init — deterministic scanner (Prisma / OpenAPI) that extracts your ontology instead of asking you to type it. No LLM calls, no API keys — ever.
• ontogate mcp — typed MCP server over your declared objects, links, and actions.
• ontogate approvals — CLI approval queue for staged actions.
• ontogate audit verify — hash-chain verification of the audit log.
• ontogate studio — a live, read-only map of your domain graph.

License

MIT