Open Ontologies

A Terraforming MCP for Knowledge Graphs
Validate, classify, and govern AI-generated ontologies. Written in Rust. Ships as a single binary.

English · 简体中文

Quick Start · Studio · Benchmarks · IES · Tools · Extending · Architecture · Docs

Open Ontologies is a Rust MCP server and desktop Studio for AI-native ontology engineering. It exposes 70+ tools that let Claude build, validate, query, diff, lint, version, reason over, align, plan, certify, and govern RDF/OWL ontologies using an in-memory Oxigraph triple store — with a full three-layer Dynamics → Causal → Planner architecture, a marketplace of 33 standard ontologies, clinical crosswalks, semantic embeddings, and a full lineage audit trail.

The Studio wraps the engine in a visual desktop environment: virtualized ontology tree with hierarchy lines, breadcrumb navigation, and connection explorer; AI chat panel with /build (IES-level deep) and /sketch (quick prototype) commands; Protégé-style property inspector; and lineage viewer.

No JVM. No Protégé.

What's New (three-layer architecture + 13 new primitives)

The full Dynamics → Causal → Planner stack plus 13 new primitives. Every piece holds the MCP-native convention: the server provides validation and scaffolding, the connected LLM (Claude over MCP) does the intelligence. No internal LLM clients, no API keys, no provider abstractions.

Three-layer architecture

13 new primitives

• onto_owl_shacl_coevolve_check + onto_owl_shacl_coevolve_incremental — SHACL validation against the OWL-RL closure, with dependency-graph routing so only shapes touching changed IRIs revalidate.
• onto_segment_retrieve — TBox-slice retrieval for ontology-grounded RAG.
• onto_extract_scaffold + onto_extract_validate — schema-guided structured extraction with typed datatype validation + conformance scoring.
• onto_cq_run + onto_verify_cq + onto_cq_verdicts_list — competency-question runner with pitfall hints + LLM-judgement loop.
• onto_classify_el — OWL-EL classification (transitive subsumption table, trivial pairs excluded).
• onto_eval_alignment — P/R/F1 over reference + computed alignment sets.
• onto_shape_combinatorics + onto_shape_induce — property-combination lattice + data-driven SHACL shape induction with support × confidence ranking.
• borderline_partition + borderline_record_verdict — generalised two-threshold review loop for any candidate set.
• onto_align_fuzzy — embedding-free fuzzy-logic adjudication with 10-rule Mamdani inference; HNSW is demoted to a candidate generator.
• onto_align_flora — end-to-end alignment pipeline pairing the signal extractor to the fuzzy adjudicator.
• onto_policy_register + onto_policy_list + onto_policy_check — authorisation gate that composes with onto_certify_action (Causal = risk; policy = authorisation).
• eval_rag + eval_rag_mmrag — Hit@k / MRR / faithfulness / token-Jaccard / ROUGE-1 scoring for retriever pipelines, with a dataset adapter.
• graph_projection_lossy_check — the auditor that pairs with onto_segment_retrieve.

Validating end-to-end

cargo run --example three_layer_pipeline

Walks Dynamics register → PDDL compile → Fast-Downward-shaped sas_plan parse → orchestrator-side IRI bind → sandbox validate → CIVeX certify → apply with OWL-RL ramification → final state inspection. Every layer through its public API, no external dependencies (Python, DoWhy, Fast Downward) required.

Zero new external Rust dependencies; everything optional gates behind Cargo features. Full test suite (160+ tests) green on default build; cargo clippy --lib --tests --examples -- -D warnings clean across both default and causal-pywhy configurations.

Quick Start (MCP / CLI)

Install

Pre-built binaries:

# macOS (Apple Silicon)
curl -LO https://github.com/fabio-rovai/open-ontologies/releases/latest/download/open-ontologies-aarch64-apple-darwin
chmod +x open-ontologies-aarch64-apple-darwin && mv open-ontologies-aarch64-apple-darwin /usr/local/bin/open-ontologies

# macOS (Intel)
curl -LO https://github.com/fabio-rovai/open-ontologies/releases/latest/download/open-ontologies-x86_64-apple-darwin
chmod +x open-ontologies-x86_64-apple-darwin && mv open-ontologies-x86_64-apple-darwin /usr/local/bin/open-ontologies

# Linux (x86_64)
curl -LO https://github.com/fabio-rovai/open-ontologies/releases/latest/download/open-ontologies-x86_64-unknown-linux-gnu
chmod +x open-ontologies-x86_64-unknown-linux-gnu && mv open-ontologies-x86_64-unknown-linux-gnu /usr/local/bin/open-ontologies

Docker:

docker pull ghcr.io/fabio-rovai/open-ontologies:latest
docker run -i ghcr.io/fabio-rovai/open-ontologies serve

serve starts an MCP server that speaks JSON-RPC over stdin/stdout — it is not an interactive CLI, so on launch it will appear to "hang" while it waits for an MCP client to connect. That is expected. To try the tools directly from a terminal instead, use the CLI subcommands (e.g. open-ontologies validate <file.ttl>); to use it with an LLM, wire it into an MCP client as shown under Connect to your MCP client.

From source (Rust 1.85+):

git clone https://github.com/fabio-rovai/open-ontologies.git
cd open-ontologies && cargo build --release
./target/release/open-ontologies init

For native Windows builds, see docs/windows.md.

Connect to your MCP client

{
  "mcpServers": {
    "open-ontologies": {
      "command": "/path/to/open-ontologies/target/release/open-ontologies",
      "args": ["serve"]
    }
  }
}

{
  "mcpServers": {
    "open-ontologies": {
      "command": "/path/to/open-ontologies/target/release/open-ontologies",
      "args": ["serve"]
    }
  }
}

{
  "mcpServers": {
    "open-ontologies": {
      "command": "/path/to/open-ontologies/target/release/open-ontologies",
      "args": ["serve"]
    }
  }
}

{
  "mcpServers": {
    "open-ontologies": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "ghcr.io/fabio-rovai/open-ontologies", "serve"]
    }
  }
}

open-ontologies serve-http --host 127.0.0.1 --port 8080

Build your first ontology

Build me a Pizza ontology following the Manchester University tutorial.
Include all 49 toppings, 24 named pizzas, spiciness value partition,
and defined classes (VegetarianPizza, MeatyPizza, SpicyPizza).
Validate it, load it, and show me the stats.

Claude generates Turtle, then runs the full pipeline automatically:

onto_validate → onto_load → onto_stats → onto_reason → onto_stats → onto_lint → onto_enforce → onto_query → onto_save → onto_version

Every build includes OWL reasoning (materializes inferred triples), design pattern enforcement, and automatic versioning.

Studio (Desktop App)

The Studio is a native desktop application that wraps the same engine in a visual environment — no browser, no server to manage. It runs entirely on your machine: the engine sidecar handles RDF/OWL operations while the UI renders the graph in real time.

Think of it as Protege meets an AI copilot. Type "build ontology about cats" and watch a 1,400-class ontology appear in the tree — classes, properties, individuals, and axioms built automatically across 13 pipeline steps. Click any node to inspect its triples, trace connections via clickable pills, and follow every change through the lineage panel.

Why virtualized tree (not 3D graph)

Prior to v0.1.12, the Studio used a D3.js horizontal tree and a 3D force-directed graph (Three.js / WebGL). Both worked for small ontologies (~100 classes) but became unusable at IES-level depth: the D3 tree couldn't handle 500+ nodes without layout thrashing, and the 3D graph froze the WebKit webview above 1,000 nodes.

The v2 deep builder changed the equation — a single /build command now produces 1,400+ classes. We replaced both views with a virtualized DOM tree: only visible rows exist in the DOM (constant memory regardless of ontology size), with hierarchy connector lines, type-filtered legend, search, breadcrumb navigation, and a connections panel. This handles the full IES Common (511 classes) and deep-built ontologies (1,400+ classes) without lag.

How it works

The Studio launches three processes that communicate locally:

1. Tauri 2 shell — native window (macOS/Linux/Windows) with a WebKit webview
2. Engine sidecar — the same Rust binary, running as an HTTP MCP server on localhost:8080
3. Agent sidecar — Node.js process running Claude via the Agent SDK, connected to the engine over MCP

When you type in the chat panel, your message goes to the Agent sidecar, which sends it to Claude. Claude decides which onto_* tools to call, the engine executes them, and the UI refreshes the graph. The entire loop — prompt to visual update — takes seconds.

Install and run

Prerequisites: Rust + Cargo · Node.js 18+

# 1. Build the engine binary (from repo root)
cargo build --release

# 2. Install JS dependencies
cd studio && npm install

# 3. Run
PATH=/opt/homebrew/bin:~/.cargo/bin:$PATH npm run tauri dev

The first launch compiles the Tauri shell (~2 min). Subsequent launches start in seconds.

Features

Keyboard shortcuts

Research Questions

The benchmarks below are not a feature tour. Each one exists to answer a specific question, and the answers include the unflattering ones.

| RQ7 | When the graph is register-scale (276k triples) rather than fund-universe-scale (1.29M), does the reference store still force the set-based workaround, and do the rule counts survive entity resolution? | insurance-register-ontology: 276,683-triple EEA insurance register, 6 layer-3 rules, rdflib 7.6.0 A/B | The workaround becomes speed, not survival. At this scale rdflib completes all six rules (133.2s total, 97s of it in one double anti-join) against 2s to 9s for the Oxigraph engine; nothing times out. Counts match the reference pipeline exactly on R1-R4 (643 / 4 / 118 / 42). R5 falls from 227 to 20 and R6 rises from 283 to 291 because the graph runs the same rule text after entity resolution, which absorbs the legitimate branch shares and attributes 8 more zombie passports via the LEI join; both differences were verified set-based. The shapes forced two more validator fixes: sh:pattern and sh:hasValue. | | RQ8 | When a published standard ships a validator but no RDF binding, does a green SHACL result mean the data is valid, or that nothing was checked? | dcat-us-binding: DCAT-US 3.0, the US federal metadata profile, 115 published examples, pySHACL vs this engine | It means nothing was checked, and the distinction is invisible without reporting reach. The corpus as published expands to 76 triples, 1 predicate, 0 DCAT, so the shapes GSA deleted select 0 focus nodes and return conforms=True. Generating a JSON-LD context from the same JSON Schema that already carries the RDF terms (231 of 270 properties) lifts the corpus to 1,510 triples and 228 DCAT statements, and the same deleted shapes, unchanged, then select 228 focus nodes and report 287 violations. This is why onto_shacl reports focus_nodes next to the verdict. The run also cost us two defects here: sh:class and sh:nodeKind are not evaluated at all (287 of pySHACL's 316 findings invisible), and sh:datatype rejects "1024"^^xsd:nonNegativeInteger against sh:datatype xsd:nonNegativeInteger. |

RQ2 and RQ4 are the ones worth reading if you are deciding whether to trust this repo. Both are negative results about work done here.

Benchmarks

How to read these numbers. Unless stated, LLM results are single-run (not averaged over seeds) and use Claude Opus 4.8. Several benchmark ontologies (Pizza, FOAF, Schema.org, OWL-Time) are widely published and may appear in an LLM's pretraining data, so a bare-LLM score is a contamination-inclusive baseline, not a clean measure of reasoning — the contribution is the delta the MCP tools add on top of that baseline, and whether that delta reproduces across models. To check exactly that, the repo ships a cross-model ablation driving the same tasks with a local Qwen3-Coder-30B as well as Claude — see benchmark/ontoaxiom/. If the tool-augmented gain holds on a second, open model, the gain is a property of the tooling, not of one vendor's model.

OntoAxiom — LLM Axiom Identification

OntoAxiom tests axiom identification across 9 ontologies and 3,042 ground truth axioms.

All conditions below are scored by a single evaluator (benchmark/ontoaxiom/score_all_conditions.py) with one shared normalizer, and both averages are reported, because the original scripts disagreed on all of that. macro = mean of per-(ontology, axiom) F1; micro = F1 over pooled TP/FP/FN.

The paper's "raw OWL hurts" result is a scoring artifact. OntoAxiom reports that an LLM given the full OWL file (F1 = 0.323) does worse than one given only class/property name lists (0.431). Those two numbers came from scripts that disagreed on three axes: the name-list scorer splits camelCase and the raw-OWL scorer only lowercases; the first reports a macro mean and the second a micro F1; and they flip pair order on different axiom types. Every one of those differences penalizes the raw-OWL condition, because that is the one where the model reads real Turtle and therefore answers in QNames (foaf:Person) and rdfs:label text ("personal mailbox" for mbox) rather than bare local names. 0.431 and 0.323 were never the same statistic.

Rescoring the same stored predictions under one evaluator flips the sign on both models and under both averages: Claude 0.451 → 0.768 macro (0.397 → 0.686 micro), Qwen 0.246 → 0.673 macro, winning 33/43 and 33/38 cells respectively. score_condition_d.py --legacy reproduces the broken 0.323 exactly, so the bug is demonstrated rather than asserted. The correction moves 0 of 5,083 name-list pairs, so it cannot flatter the baseline, and it still under-credits raw OWL: 51.8% of Claude's pairs are label text no normalizer here can match. Full analysis and reproduction: benchmark/ontoaxiom/ONTOAXIOM_SHOWDOWN.md.

Corrected, reading the ontology and SPARQL-extracting it are at parity — raw OWL wins macro (0.768 vs 0.713), extraction wins micro (0.717 vs 0.686). So the tools' edge is auditability, not F1: every MCP pair traces to a query against real triples, whereas an LLM reading a file can still hallucinate a plausible pair and no F1 score will say which.

Pizza Ontology — Manchester Tutorial

One sentence input: "Build a Pizza ontology following the Manchester tutorial specification."

/sketch vs /build — Two Build Modes

The Studio provides two build commands for different use cases. Both take the same input — "build ontology about cats" — but produce very different results:

/sketch runs 3 steps: classes + properties in one Turtle block, axioms + individuals, then save. Good for quick domain exploration or demo prototyping. Produces a complete ontology with hierarchy, properties, and individuals — but at a fraction of the depth.

/build runs a 13-step pipeline within a single persistent Claude session: foundation classes → per-branch deepening (4 passes) → gap filling → object properties (2 batches) → datatype properties → disjoints → individuals → reason → save. Each step focuses on one aspect of the ontology, staying within output token limits while building on the previous step's context. The result exceeds IES Common on every metric.

/sketch is comparable to the Pizza benchmark (95 classes, 8 properties). /build produces IES-level ontologies — deep enough for production use.

Mushroom Classification — OWL Reasoning vs Expert Labels

Dataset: UCI Mushroom Dataset — 8,124 specimens classified by mycology experts.

Ontology Marketplace — 33 Standard Ontologies

The 29 general-purpose marketplace ontologies (the four IES entries are covered separately under IES Support) fetched, owl:imports resolved, loaded, and reasoned over with both RDFS and OWL-RL profiles. Regenerate with python3 benchmark/marketplace_benchmark.py:

29/29 ontologies loaded, imports resolved, and reasoned. RDFS adds 18% more triples. OWL-RL adds 43% — transitive/symmetric/inverse properties and equivalentClass expansion discover significantly more implicit knowledge. Schema.org jumps from +4,082 (RDFS) to +14,236 (OWL-RL) inferred triples in 136ms.

Class and property counts are structural, not declaration-only: a term counts if it is typed (owl:Class, rdfs:Class, owl:ObjectProperty, owl:DatatypeProperty, rdf:Property) or used as one (rdfs:subClassOf/subPropertyOf/domain/range position). Vocabularies that never issue OWL type declarations, such as Schema.org, are therefore counted rather than reported as empty. Terms in the rdf:, rdfs: and owl: namespaces are excluded so a vocabulary is not credited with the meta-vocabulary it is written in. That exclusion is why OWL 2, RDF Schema and RDF Concepts report 0 properties: every property they define is, by definition, in an excluded namespace. FIBO's marketplace entry is the metadata module only (48 triples), which declares no terms of its own.

Compiled Claim Verification — measured vs HermiT

The claimcheck module compiles an ontology once (inferred hierarchy + disjointness, including pairs derived by sound propagation rules) into per-class token bitsets, then verifies candidate claims — the "is this set of triples consistent with the ontology?" question — with no reasoning at query time: two 64-bit ANDs per class pair, witness axiom extracted for the explanation.

Same ontology (canonical pizza.owl), same task, same machine, verdicts cross-checked against HermiT 1.4.3.456:

Correctness before speed:

• 0 disagreements with HermiT across 78,884 exhaustively-audited class pairs (13 ontologies) and 793 structurally adversarial claims.
• Sound by construction: a Rejected verdict is backed by a derivable contradiction and names the witnessing axiom. 100% contradiction recall on both fully-audited ontologies.
• Explicit incompleteness envelope: anything the compiled surface cannot decide returns Undetermined and routes to a reasoner-backed residual tier — it is never guessed.
• Closed-world vocabulary checks catch hallucinated classes/properties that open-world OWL semantics structurally cannot flag.

Offline compile: one classification pass (~120 ms for Pizza) via any complete OWL reasoner; the shipped hot path is pure Rust with no JVM dependency. Reproduction scripts: benchmark/layer3-prototype/.

Design, measurements and envelope: docs/layer3-compiled-reasoning.md. Full benchmark methodology: docs/benchmarks.md

OAEI Ontology Alignment — Anatomy Track

OAEI is the standard benchmark for ontology alignment systems. The Anatomy track aligns the mouse anatomy ontology (2,744 classes) to the human anatomy fragment of the NCI Thesaurus (3,304 classes) against 1,516 reference mappings.

The comparison below is the complete OAEI 2025 Anatomy field, reproduced from Table 9 of the official results paper (Vol-4144, om2025-oaei-paper0), including both baselines. Open Ontologies is inserted at its measured rank. Nothing is filtered.

Read this honestly. Open Ontologies ranks 9th of 13 on F1. Its precision (0.960) is third in the field, but its recall (0.730) is second-from-bottom among non-baseline systems, and the resulting F1 sits level with LogMapLt, the deliberately lightweight lexical matcher, and 0.112 below the leader. The StringEquiv baseline reaches 0.766 by normalised string equality alone, so the margin this system earns over pure string matching is +0.063 F1. That is the number to beat, and it is not yet a competitive result.

The gap is recall, and its cause is identifiable: this system carries no domain background knowledge. Anatomy is a track where UMLS and BioPortal lookups are what separate the 0.88+ band from the rest, and every system above Open Ontologies in the table either uses biomedical background knowledge, an LLM oracle, or both. Alignment here uses 7 weighted signals (label similarity, property/parent/instance/restriction/neighbourhood overlap, embedding similarity), stable 1-to-1 matching, and a label penalty when no structural evidence is available.

The defensible finding from this track is therefore not the headline F1. It is the ablation: stable 1-to-1 matching is what makes the difference, and the signal weights are nearly irrelevant once it is applied.

Removing stable matching produces 12,557 candidates against a 1,516-mapping reference. See issues #8, #9, #10; background-knowledge integration is the open work.

OAEI Ontology Alignment — Conference Track

15 of the 21 conference-track pairs, micro-averaged:

Not comparable like-for-like, and worse than it looks. The OAEI rows are Table 10 of the 2025 results paper, evaluated over all 21 pairs against the rar2 reference at each system's F1-optimal threshold. The Open Ontologies row covers 15 pairs at a fixed confidence threshold. With that caveat stated, the result is unambiguous: below every participating system and below both baselines. Best pairs are ekaw-iasted (0.588) and ekaw-sigkdd (0.533); worst is edas-sigkdd (0.211). Conference is a track where the same recall failure that costs Anatomy its rank costs more, because there is no lexical redundancy to fall back on.

IES Support

IES (Information Exchange Standard) is the UK National Digital Twin Programme's core ontology framework. It uses a 4D extensionalist (BORO) approach for modelling entities, events, states, and relationships. Open Ontologies supports the full IES stack — all three layers, SHACL shapes, and example datasets from the IES-Org GitHub repositories.

The IES Layers

The marketplace includes all three tiers of the IES framework:

onto_marketplace install ies-top     # ToLO — BORO foundations (~22 classes)
onto_marketplace install ies-core    # Core — persons, states, events (~131 classes)
onto_marketplace install ies         # Common — full ontology (511 classes, 206 properties)

Benchmark

IES is the second-largest ontology in the marketplace by class count (after Schema.org). RDFS reasoning produces the richest inference gain of any non-general ontology — State, ClassOfEntity, and Event subclasses all generating deep transitive chains.

Example Data

Load IES example datasets directly from the official repositories:

onto_pull https://raw.githubusercontent.com/IES-Org/ont-ies/main/docs/examples/sample-data/event-participation.ttl
onto_pull https://raw.githubusercontent.com/IES-Org/ont-ies/main/docs/examples/sample-data/hospital.ttl
onto_pull https://raw.githubusercontent.com/telicent-oss/ies-examples/main/additional_examples/ship_movement.ttl

SHACL Validation

onto_pull https://raw.githubusercontent.com/IES-Org/ont-ies/main/docs/specification/ies-common.shacl
onto_shacl

Data Mapping: EPC → IES

The repo includes a sample of real UK Energy Performance Certificates (benchmark/epc/epc-sample.csv) with a mapping config that transforms tabular EPC data into IES-shaped RDF:

onto_load benchmark/generated/ies-building-extension.ttl
onto_ingest benchmark/epc/epc-sample.csv --mapping benchmark/epc/epc-ies-mapping.json
onto_reason --profile rdfs

This mirrors NDTP's actual pipeline: CSV → IES RDF → validate → reason → query.

IES Building Extension — Comparison with NDTP/IRIS

The repo includes an IES Building Extension built from the UK EPC data schema and building science fundamentals, using IES 4D patterns. It was built independently — without reference to any existing implementation — then compared against the NDTP/IRIS production building ontology used in government data pipelines.

Built blind from the 105-column EPC schema, SAP methodology, and BORO 4D extensionalism — zero reference to the IRIS implementation. The two ontologies make different trade-offs: IRIS is more tightly curated with higher average hierarchy depth (2.89 vs 2.02), reflecting deliberate grouping by domain experts. Open Ontologies covers more of the EPC data schema and applies the BORO 4D pattern more systematically across the domain.

How the hierarchy emerges from building science

The ontology's depth (max 10 levels) is not hand-tuned — it follows the natural classification that building scientists use. The EPC data schema describes heating systems as flat text fields ("Condensing gas boiler with radiators"), but the underlying domain has layered structure:

graph TD
    HS[Heating System] --> CH[Central Heating]
    HS --> NC[Non-Central / Room Heating]

    CH --> WET[Wet Central Heating<br/><i>hydronic distribution</i>]
    CH --> WA[Warm Air Central Heating<br/><i>ducted air</i>]
    CH --> EC[Electric Central Heating<br/><i>storage / underfloor</i>]

    WET --> BB[Boiler-Based]
    WET --> HP[Heat Pump]
    WET --> DH[Community / District]

    BB --> CB[Combustion Boiler]
    BB --> CHP[Micro-CHP]

    CB --> GAS["Gas boiler"]
    CB --> OIL["Oil boiler"]
    CB --> LPG["LPG boiler"]
    CB --> COND["Condensing boiler"]
    CB --> COMBI["Combi boiler"]
    CB --> BACK["Back boiler"]

    HP --> ASHP["Air source"]
    HP --> GSHP["Ground source"]
    HP --> WSHP["Water source"]

    EC --> STOR["Storage heaters"]
    EC --> PNL["Panel heaters"]
    EC --> UF["Underfloor electric"]

    NC --> FIX[Fixed Room Heater]
    NC --> PORT[Portable Heater]

    FIX --> GROOM["Gas room heater"]
    FIX --> EROOM["Electric room heater"]
    FIX --> SFROOM["Solid fuel room heater"]

    style HS fill:#1a1a2e,color:#fff
    style CH fill:#16213e,color:#fff
    style NC fill:#16213e,color:#fff
    style WET fill:#0f3460,color:#fff
    style WA fill:#0f3460,color:#fff
    style EC fill:#0f3460,color:#fff
    style BB fill:#533483,color:#fff
    style HP fill:#533483,color:#fff
    style DH fill:#533483,color:#fff
    style CB fill:#e94560,color:#fff
    style CHP fill:#e94560,color:#fff

The same pattern applies to the building fabric — heat transfer physics dictates the grouping:

graph TD
    TE[Building Thermal Envelope] --> OP[Opaque Elements<br/><i>conduction-dominated</i>]
    TE --> TR[Transparent Elements<br/><i>radiation + conduction</i>]

    OP --> WALL[Walls]
    OP --> ROOF[Roofs]
    OP --> FLOOR[Floors]

    TR --> WIN[Windows]
    TR --> DOOR[Doors]

    WALL --> MAS[Masonry Walls<br/><i>thermal mass</i>]
    WALL --> FRM[Framed Walls<br/><i>stud bridges</i>]

    MAS --> CAV["Cavity wall"]
    MAS --> SOL["Solid brick"]
    MAS --> SND["Sandstone"]
    MAS --> GRN["Granite"]
    MAS --> COB["Cob"]

    FRM --> TF["Timber frame"]
    FRM --> SYS["System-built"]
    FRM --> PH["Park home"]

    ROOF --> PIT[Pitched Roof]
    ROOF --> FLT[Flat Roof]

    PIT --> COLD["Cold roof<br/><i>insulation at ceiling</i>"]
    PIT --> WARM["Warm roof<br/><i>insulation at rafter</i>"]
    PIT --> THATCH["Thatched"]

    WIN --> SGL["Single glazed"]
    WIN --> DBL["Double glazed"]
    WIN --> TPL["Triple glazed"]
    WIN --> SEC["Secondary glazing"]

    style TE fill:#1a1a2e,color:#fff
    style OP fill:#16213e,color:#fff
    style TR fill:#16213e,color:#fff
    style WALL fill:#0f3460,color:#fff
    style ROOF fill:#0f3460,color:#fff
    style FLOOR fill:#0f3460,color:#fff
    style WIN fill:#0f3460,color:#fff
    style DOOR fill:#0f3460,color:#fff
    style MAS fill:#533483,color:#fff
    style FRM fill:#533483,color:#fff
    style PIT fill:#533483,color:#fff
    style FLT fill:#533483,color:#fff

Each level in the tree is a real building science distinction — central vs room heating, hydronic vs warm air, combustion vs electric, masonry vs framed, cavity vs solid. An independent building scientist, given the same EPC data values, produces these same intermediate groupings (verified by clean-room reproduction). RDFS reasoning traverses these chains transitively, which is why a 10-level hierarchy generates 662 inferred triples from 3,229 raw.

EPC Column Coverage Benchmark

Both ontologies tested against 36 key EPC data columns — can each ontology receive and represent the data from that column?

Queries derived from published DESNZ/ONS EPC statistical reports — not from either ontology's class structure. Full benchmark: benchmark/epc/

Use onto_align to map it to other domain ontologies:

onto_load benchmark/generated/ies-building-extension.ttl
onto_align <other-ontology.ttl>

Hierarchy Enforcement — Automated Inference Improvement

The hierarchy enforce pack detects flat spots in any ontology and suggests intermediate grouping classes. This is the same process used to deepen the building extension — now codified as a repeatable tool:

onto_load my-ontology.ttl
onto_enforce --pack hierarchy
# → flags classes with >5 direct children
# → reports max depth, avg depth, hierarchy density

Tested on IES Common (511 classes), the tool found 24 flat spots. A clean-room agent — with no prior context — proposed 38 intermediate grouping classes based solely on the domain meaning of the flagged children:

graph LR
    subgraph Before["IES Common — before"]
        EP1[EventParticipant] --> P1["Prosecutor"]
        EP1 --> P2["Observer"]
        EP1 --> P3["Driver"]
        EP1 --> P4["Supplier"]
        EP1 --> P5["WeaponLocation"]
        EP1 --> P6["...52 direct children"]
    end

    subgraph After["IES Common — after hierarchy enforce"]
        EP2[EventParticipant] --> R[RoleInEvent]
        EP2 --> L[LocationInEvent]
        EP2 --> A[AssetInEvent]
        R --> LR2[LegalRole]
        R --> IR[InvestigativeRole]
        R --> CR[CommercialRole]
        LR2 --> Q1["Prosecutor"]
        LR2 --> Q2["Signatory"]
        IR --> Q3["Observer"]
        IR --> Q4["Investigator"]
        CR --> Q5["Supplier"]
        CR --> Q6["Negotiator"]
        L --> Q7["WeaponLocation"]
        L --> Q8["TargetLocation"]
        A --> Q9["VehicleUsed"]
    end

    style EP1 fill:#e94560,color:#fff
    style EP2 fill:#1a1a2e,color:#fff
    style R fill:#16213e,color:#fff
    style L fill:#16213e,color:#fff
    style A fill:#16213e,color:#fff
    style LR2 fill:#0f3460,color:#fff
    style IR fill:#0f3460,color:#fff
    style CR fill:#0f3460,color:#fff

The same tool, applied to any ontology, produces the same kind of improvement. The intermediate classes emerge from domain knowledge — not from reference to any other implementation.

Further Reading

Tools

70+ tools organized by function — available as MCP tools (prefixed onto_) and CLI subcommands:

Extending Open Ontologies

Four extension surfaces, in increasing order of coupling — full map in ECOSYSTEM.md:

Everything holds the MCP-native convention: extensions provide validation and scaffolding; the connected LLM does the intelligence.

Architecture

Engine

flowchart TD
    subgraph Clients["Clients"]
        Claude["Claude / LLM\nMCP stdio"]
        CLI["CLI\nonto_* subcommands"]
        Studio["Studio\nHTTP REST"]
    end

    subgraph Server["Open Ontologies Server"]
        direction TB

        subgraph Transport["Transport Layer"]
            MCP_HTTP["MCP Streamable HTTP\n/mcp"]
            REST["REST API\n/api/query · /api/update\n/api/save · /api/load · /api/lineage"]
        end

        subgraph ToolGroups["70+ Tools"]
            direction LR
            Core["Core\nvalidate · load · save · clear\nstats · query · diff · lint\nconvert · status"]
            DataPipe["Data Pipeline\nmap · ingest · shacl\nreason · extend · import-schema"]
            Lifecycle["Lifecycle\nplan · apply · lock · drift\nenforce · monitor · lineage"]
            Advanced["Alignment + Clinical\nalign · crosswalk · enrich\nenrich · embed · search · similarity\ndl_explain · dl_check"]
            Version["Versioning\nversion · history · rollback"]
        end

        subgraph Core2["Core Engine"]
            GraphStore["Oxigraph Triple Store\nRDF/OWL in-memory\nSPARQL 1.1"]
            SQLite["SQLite\nlineage events\nversion snapshots\nlint/enforce feedback\nembedding vectors"]
            Reasoner["OWL2-DL Reasoner\nSHOIQ tableaux\nRDFS · OWL-RL"]
            Embedder["Embedding Engine\ntract-onnx (ONNX)\ntext + Poincaré structural"]
        end
    end

    subgraph External["External Sources"]
        PG["PostgreSQL\nschema import"]
        SPARQL["Remote SPARQL\nendpoints"]
        OWL["OWL URLs\nowl:imports chains"]
        Parquet["Parquet / Arrow\nclinical crosswalks\nICD-10 · SNOMED · MeSH"]
        Files["Files\nCSV · JSON · XML\nYAML · XLSX · Parquet"]
    end

    Claude -->|"MCP stdio"| MCP_HTTP
    CLI -->|"subcommands"| MCP_HTTP
    Studio -->|"sessionless"| REST

    MCP_HTTP --> ToolGroups
    REST --> ToolGroups

    ToolGroups --> GraphStore
    ToolGroups --> SQLite
    ToolGroups --> Reasoner
    ToolGroups --> Embedder

    Reasoner --> GraphStore
    Embedder --> SQLite

    DataPipe --> Files
    Advanced --> Parquet
    Core --> OWL
    Core --> SPARQL
    DataPipe --> PG

Studio

flowchart TD
    subgraph UI["React UI (Vite + Tailwind CSS)"]
        Graph["Virtualized Tree\nDOM + virtual scroll"]
        Chat["AI Chat Panel\nZustand store"]
        Inspector["Property Inspector\nInline SPARQL edit"]
        Lineage["Lineage Panel\nAudit trail"]
        Save["Named Save\n⌘S → ~/.open-ontologies/"]
    end

    subgraph Tauri["Tauri 2 Shell (Rust)"]
        IPC["Tauri IPC\ninvoke / event"]
        ChatState["ChatState\nstdin/stdout pipe"]
    end

    subgraph Engine["Engine Sidecar (Rust / Axum)"]
        MCP["/mcp — MCP Streamable HTTP\nonto_* tools"]
        REST2["/api/query · /api/update\n/api/save · /api/load-turtle\n/api/stats · /api/lineage"]
        Store["Arc<GraphStore>\nOxigraph"]
        DB["SQLite"]
    end

    subgraph Agent["Agent Sidecar (Node.js)"]
        SDK["Claude Opus 4.8\nAgent SDK"]
        Proto["stdin/stdout JSON protocol"]
    end

    Graph -->|"SPARQL SELECT/UPDATE · REST"| REST2
    Inspector -->|"SPARQL UPDATE · REST"| REST2
    Lineage -->|"GET /api/lineage"| REST2
    Save -->|"POST /api/save"| REST2
    Chat -->|"invoke send_chat_message"| IPC
    IPC --> ChatState
    ChatState -->|"stdin { type: chat }"| Proto
    Proto --> SDK
    SDK -->|"MCP tools/call"| MCP
    SDK -->|"stdout { type: text/tool_call/done }"| Proto
    Proto -->|"Tauri emit agent-message"| Chat
    MCP --> Store
    REST2 --> Store
    Store --> DB

Design decisions

Stack

Documentation

Citation

Open Ontologies is described in a preprint. The alignment engine implements the stable-matching method introduced there; the Causal layer builds on the intervention-verification framework of CIVeX.

• Open Ontologies: Tool-Augmented Ontology Engineering with Stable Matching Alignment. Fabio Rovai, 2026. arXiv:2605.09184
• CIVeX: Causal Intervention Verification for Language Agents. Fabio Rovai, 2026. arXiv:2605.09168

@article{rovai2026openontologies,
  title   = {Open Ontologies: Tool-Augmented Ontology Engineering with Stable Matching Alignment},
  author  = {Rovai, Fabio},
  journal = {arXiv preprint arXiv:2605.09184},
  year    = {2026},
  doi     = {10.48550/arXiv.2605.09184},
  url     = {https://arxiv.org/abs/2605.09184}
}

See CITATION.cff for machine-readable metadata. It powers GitHub's "Cite this repository" button.

Maintainer

Maintained by The Tesseract Academy (Kampakis and Co Ltd), a UK research and data-science practice. Applied case studies built with this toolkit are published in the Tesseract Foundational Research programme.

License

MIT

Sponsor

If this work is useful to you, you can support its continued development through GitHub Sponsors.