⏳ This skill is pending AI review.

Scores will appear once the review pipeline completes.

v1.1.0

ontomcp-driver

mareasw@mareasw⭐ 23 stars· last commit 1d ago· 0 open issues

Use OntoSkills MCP tools to discover and apply structured skill knowledge.

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

—/10

// RATINGS

⭐GitHub Stars
⭐⭐ 23 on GitHubGitHub ↗

Growing

🟢ProSkills Score
—
📍

Not yet listed on ClawHub or SkillsMP

// README


What is OntoSkills?

OntoSkills transforms natural language skill definitions into validated OWL 2 ontologies — queryable knowledge graphs that enable deterministic reasoning for AI agents.

The problem: LLMs read skills probabilistically. Same query, different results. Long skill files burn tokens and confuse smaller models.

The solution: Compile skills to ontologies. Query with SPARQL. Get exact answers, every time.

flowchart LR
    CORE["OntoCore<br/>━━━━━━━━━━<br/>SKILL.md → .ttl<br/>LLM + SHACL"] -->|"compiles"| CENTER["OntoSkills<br/>━━━━━━━━━━<br/>OWL 2 Ontologies<br/>.ttl artifacts"]
    CENTER -->|"loads"| MCP["OntoMCP<br/>━━━━━━━━━━<br/>Rust SPARQL<br/>in-memory graph"]
    MCP <-->|"queries"| AGENT["AI Agent<br/>━━━━━━━━━━<br/>Deterministic<br/>reasoning"]

    style CORE fill:#e91e63,stroke:#2a2a3e,color:#f0f0f5
    style CENTER fill:#abf9cc,stroke:#2a2a3e,color:#0d0d14
    style MCP fill:#92eff4,stroke:#2a2a3e,color:#0d0d14
    style AGENT fill:#6dc9ee,stroke:#2a2a3e,color:#0d0d14

Why OntoSkills?

ProblemSolution
LLMs interpret text differently each timeSPARQL returns exact answers
50+ skill files = context overflowQuery only what's needed
No verifiable structure for relationshipsOWL 2 formal semantics
Small models can't read complex skillsDemocratized intelligence via graph queries

For 100 skills: ~500KB text scan → ~1KB query

→ Read the full philosophy


Quick Start

# Install the MCP runtime and bootstrap your client
npx ontoskills install mcp --claude

# Or install the Python compiler separately
pip install ontocore
ontocore compile

→ Full installation guide


Components

ComponentLanguageDescription
OntoCorePythonNeuro-symbolic compiler: SKILL.md → OWL 2 ontology
OntoMCPRustMCP server with sub-ms SPARQL queries
OntoStoreGitHubVersioned skill registry
CLINode.jsOne-command installer (npx ontoskills)

Document knowledge base

OntoCore also compiles a folder of documents into a knowledge base, not just skills. Every statement is an atomic claim with a verbatim quote and a locator (document, page, section, URL), and answers are deterministic — no model is called at query time.

cd examples/demo-corpus && ./fetch.sh     # 14 public documents, hashes checked

ontocore ingest . -o kb --estimate        # cost estimate, no API calls
ontocore ingest . -o kb --resume          # compile (paid, one model per chunk)
ontocore ingest . -o kb --resume --semantic  # + local vectors, still no network
ontocore ask "How quickly must a personal-data breach be notified?" --kb kb
ontocore view --kb kb --questions-file questions.txt
  • the compiler never writes a quote: the model returns span ids, the quote and locator are copied from the segmented text; a list item also carries the span that opens its list as context evidence, so an item with no subject of its own is still readable;
  • a deterministic verifier (verbatim, locator, page, section, numbers) gates every claim and ships weak / no verdicts visibly instead of hiding them; ask answers only claims it passed, view shows every claim with its badge;
  • a number that names where a rule lives (Clause 8.2, Directive (EU) 2019/790) is a note, not a failure; a quantity that no quote carries is refused when the claim is built and recorded in rejected.jsonl;
  • ask ranks with BM25 and groups by document, so a cross-document question shows which document says what; when the KB carries vectors.jsonl it fuses BM25 with a local MiniLM ranking (reciprocal rank fusion), offline and with no paid call, and falls back to BM25 alone if the optional runtime is absent;
  • when two documents give the same topic different numbers, the answer reports a tension with both citations rather than picking a winner;
  • output: claims.jsonl, graph.ttl, evidence.csv, text/, and a self-contained read-only view.html with copy-citation.

Measured on that corpus with deepseek-v4.1-flash, one full run: 5,680 claims from 14 documents and 556 pages, 99.8% ok at the verifier (10 weak, 0 no), $1.27; all three demo questions cite claims from at least two documents, one of them with a cross-document tension. Retrieval is measured, not assumed: against a 20-question gold set declared before the first run, BM25 hits 13/20 and the hybrid 16/20 — both miss the pre-registered gate (18/20 and 0.60 recall), and RESULTS.md reports the numbers with the diagnosis. The weak breakdown and both measurements are in benchmark/docbench/RESULTS.md.

See docs/phase2-plan.md for the pipeline and examples/demo-corpus/ for the corpus manifest, licences and sources. Measured extraction results are in benchmark/docbench/RESULTS.md.


Documentation


License

MIT License — see LICENSE for details.

© 2026 MareaSW

// HOW IT'S BUILT

// FILES

2 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • site/public/agent-skills/ontomcp-driver/SKILL.md1.3 KB02751af15eb8upstream ↗
    ---
    name: ontomcp-driver
    description: Use OntoSkills MCP tools to discover and apply structured skill knowledge.
    ---
    
    ## Tools
    
    - **prefetch_knowledge** — One call: search + fetch context + compact. **Use this first.**
    - **search** — Find skills by keyword, alias, or structured filters.
    - **get_skill_context** — Full skill context with knowledge nodes (procedures, constraints, anti-patterns).
    - **evaluate_execution_plan** — Validate a multi-skill plan against state chains.
    - **query_epistemic_rules** — Cross-skill knowledge rules filtered by kind/severity.
    
    Responses are compact by default. Use `format: "raw"` for full JSON.
    
    ## Workflow
    
    1. Call `prefetch_knowledge` with a query describing the task goal.
    2. Read the returned knowledge nodes — they contain procedures, constraints, and anti-patterns.
    3. For multi-skill plans: call `evaluate_execution_plan` to validate state chains.
    4. Write code following the knowledge nodes.
    
    ## Critical rules
    
    - **Always prefetch first** — don't guess skill names.
    - **CRITICAL/HIGH severity nodes are hard constraints** — never skip them.
    - **Check requiresState** — skills fail without their preconditions.
    
    ## State semantics
    
    Skills form a state machine: `requiresState` (preconditions) / `yieldsState` (outcomes).
    Skill A's yieldsState must match Skill B's requiresState for valid chaining.
    
  • README.mdupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

Plain-text copy, no login: /skill-bundles/mareasw-ontoskills-site-public-agent-skills-ontomcp-driver.md

// REPO STATS

23 stars
0 open issues
Last commit: 1d ago