ontary — the ontology SDK

Author an ontology once — typed objects, links, actions, functions, and a security policy — and get a governed runtime: scope- and sensitivity-aware reads, audited business-verb actions, and an MCP server for AI agents. The engine is domain-agnostic; the ticket domain below is just an example.

Python 3.12+ · pydantic-only core · mypy --strict · offline make verify.

Release metadata: version 0.19.0, store schema v14. Releases are annotated tags (v0.19.0).

Documentation: https://ryoochi0112.github.io/ontary/ (English / 日本語).

Install

From PyPI:

pip install "ontary[mcp]"

mcp serves the ontology to AI agents; postgres adds PostgresStore. The core depends only on pydantic. Pin an exact version: ontary is pre-1.0 and any minor may break you (see Compatibility).

uv add "ontary[mcp]==0.19.0"

or in pyproject.toml:

[project]
dependencies = ["ontary[mcp]==0.19.0"]

Without an index, install the tagged git ref:

uv add "ontary @ git+https://github.com/ryoochi0112/ontary@v0.19.0"

Quickstart

Two object types, one business-verb action, a SQLite store, and an MCP server for one queue-scoped agent. Runs as-is (tests/test_docs.py executes it).

from ontary import (
    ActionContext, ActionError, ActionParams, Consumer, DirectProperty,
    ObjectStore, Ontology, OntologyObject, SelfScope, Source,
    build_mcp_server, prop, target,
)

ontology = Ontology(name="tickets", scope_levels=["queue"])

@ontology.object(layer="L0", scope=[SelfScope(level="queue")])
class Queue(OntologyObject):
    id: str = prop(primary_key=True)
    name: str

@ontology.object(
    layer="L0", owned={"escalated": False},
    scope=[DirectProperty(level="queue", property_name="queue_id")],
)
class Ticket(OntologyObject):
    id: str = prop(primary_key=True)
    subject: str
    queue_id: str = prop(scope_level="queue")
    escalated: bool | None = prop(default=False)

class EscalateTicket(ActionParams):
    ticket_id: str = target(Ticket)

@ontology.action(
    EscalateTicket, target=Ticket, roles=["Agent"],
    display_name="Escalate ticket", description="Mark a ticket urgent.",
    api_name="EscalateTicket",
)
def escalate(ctx: ActionContext, params: EscalateTicket) -> dict[str, str]:
    ticket = ctx.get(Ticket, params.ticket_id)
    if ticket is None:
        raise ActionError("ticket does not exist", code="PRECONDITION_FAILED")
    ticket.escalated = True
    ctx.save(ticket)
    return {"ticket_id": params.ticket_id}

ontology.validate()
store = ObjectStore(ontology.registry)          # ObjectStore(registry, "tickets.db") persists
source = Source(source_system="demo")
store.insert("Queue", {"id": "queue-a", "name": "Billing"}, source)
ticket_id = store.insert(
    "Ticket", {"subject": "Invoice mismatch", "queue_id": "queue-a"}, source
)
agent = Consumer(actor_id="agent-1", role="Agent", scope_level="queue",
                 scope_id="queue-a", kind="human")
client = ontology.bind(store).for_consumer(agent)
client.execute(EscalateTicket(ticket_id=ticket_id))
assert client.get(Ticket, ticket_id).escalated is True

server = build_mcp_server(ontology, store, agent)   # server.run() serves stdio

The full Org → Queue → Ticket → Comment ontology with functions and a multi-consumer MCP server is examples/tickets/.

Command line

The package installs ontary and ontary-mcp.

• ontary validate pkg.module:attr validates an ontology and reports diagnostics. Use --json for JSON output.
• ontary serve pkg.module:attr --dev [--store PATH] [--port N] serves an ontology over localhost MCP.
• ontary version prints the version.

ontary serve your_app.ontology:ontology --dev --store ./dev.sqlite --port 8000

ontary-mcp is a placeholder that exits with instructions. Full flags and exit codes: CLI reference.

Where to go

License

MIT — see LICENSE. ontary continues the ontos SDK, which Atrae, Inc. authored and released to the author for open-source development.