Open Ontologies is a Rust MCP server and desktop application for building and checking RDF and OWL ontologies. It exposes more than 70 onto_* tools that let a connected language model create, validate, query, diff, lint, version, reason over, align and govern ontologies against an in-memory Oxigraph triple store. The server holds no model client of its own: validation and scaffolding live in the binary, and the intelligence comes from whichever model is attached over MCP. It ships as a single binary with no Java runtime.

The engine layers a Dynamics, Causal and Planner stack on top of the store. Dynamics registers action schemas and applies them in atomic ticks with static causal laws, default-value laws, OWL-RL ramification and seeded non-deterministic outcomes. The Causal layer certifies an action, optionally using PyWhy backdoor identification behind a Cargo feature. The Planner compiles PDDL, calls a Fast Downward subprocess and validates the result by sandbox simulation, keeping the solver on the client side. Other primitives cover SHACL and OWL co-evolution checks, TBox-slice retrieval for ontology-grounded retrieval-augmented generation (RAG), schema-guided extraction, competency-question runs, shape induction and policy checks.

A Studio desktop application wraps the same engine: a Tauri 2 shell hosts the interface, the Rust binary runs as an HTTP MCP sidecar on localhost, and a Node.js sidecar runs Claude through the Claude Agent SDK connected to the engine over MCP. It has a virtualised ontology tree that stays responsive past 1,400 classes, a Protégé-style inline triple inspector, a lineage panel backed by SQLite, and a chat panel where /build runs a 13-step pipeline and /sketch runs three steps.

Access is available over stdio, over HTTP with serve-http, and through an Obsidian plugin that runs the engine as a managed sidecar with validate-on-save for Turtle files and a vault-to-RDF mapper. The HTTP surface includes mutating routes for SPARQL UPDATE and for reading and writing files on the server; when no bearer token is configured, that layer is not installed and every route except /health is reachable by anything that can open the port.

On the OntoAxiom axiom-identification benchmark, rescoring stored predictions under a single evaluator puts MCP extraction and raw OWL reading at parity:

Approach	Input	macro F1	micro F1
Bare Claude Opus	Name lists	0.451	0.397
Bare Qwen3-Coder-30B	Name lists	0.223	0.176
Claude Opus, raw OWL	Full Turtle	0.768	0.686
Qwen3-Coder-30B, raw OWL	Full Turtle	0.673	0.667
MCP extraction	Full OWL	0.713	0.717

Features

MCP tool surface: more than 70 onto_* tools for building, validating, querying, reasoning, aligning, versioning and governing ontologies
Triple store: in-memory Oxigraph with SPARQL SELECT and UPDATE, plus a native OWL2-DL tableaux reasoner and OWL-EL classification
Validation: SHACL against the OWL-RL closure, incremental revalidation routed by a dependency graph, and a closed-world vocabulary gate for terms SHACL accepts as merely unknown
Dynamics layer: action schemas applied in concurrent atomic ticks with invariants, default-value laws and reproducible non-deterministic outcomes
Causal layer: action certification with a structural proxy by default and opt-in do-calculus through PyWhy
Planner layer: PDDL compilation, classical planning through Fast Downward, and sandbox-simulated plan validation
Alignment: fuzzy-logic adjudication with 10-rule Mamdani inference, an end-to-end pipeline, and precision, recall and F1 evaluation
Retrieval support: TBox-slice retrieval, a lossy graph-projection auditor, and Hit@k, MRR, faithfulness and ROUGE-1 scoring
Studio: Tauri desktop app with a virtualised tree, inline triple editor, lineage panel and an agent chat with /build and /sketch

Discussing Open Ontologies· 0

0 threads · General