Skip to content

What are OntoSkills?

OntoSkills are a neuro-symbolic skill platform for deterministic agents. They turn SKILL.md sources into validated OWL 2 ontologies, serve compiled skills through a local MCP runtime, and distribute published packages through OntoStore.


Why OntoSkills?

The determinism problem

LLMs read skills probabilistically. Same query, different results. Long skill files are expensive for large models and confusing for small ones.

  • Non-deterministic reading — LLMs interpret text differently each time
  • Token waste — Large models burn tokens parsing long documentation
  • Small model limits — Complex skills are unreadable by edge models
  • No verifiable structure — Relationships between skills are implicit

The ontological solution

OntoSkills transforms skills into formal ontologies with Description Logics (OWL 2):

  • Deterministic queries — SPARQL returns exact answers, every time
  • Entailment reasoning — Infer dependencies, conflicts, capabilities
  • Democratized intelligence — Small models query what large models read
  • Formal semantics — No ambiguity in skill relationships

Performance comparison

OperationReading FilesOntology Query
Find skill by intentO(n) text scanO(1) indexed lookup
Check dependenciesParse each fileFollow dependsOn edges
Detect conflictsCompare all pairsSingle SPARQL query

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


How it works

OntoCore Architecture

The compilation pipeline

  1. Extract — Claude reads SKILL.md and extracts structured knowledge
  2. Validate — Security pipeline checks for malicious content
  3. Serialize — Pydantic models → RDF triples
  4. Verify — SHACL gatekeeper ensures logical validity
  5. Write — Compiled .ttl files to ontoskills/

The runtime

  • OntoMCP loads compiled .ttl files from ontoskills/
  • Agents query via SPARQL through the MCP protocol
  • OntoStore is built in by default
  • Third-party stores can be added explicitly with store add-source

Key capabilities

CapabilityDescription
LLM ExtractionClaude extracts structured knowledge from SKILL.md files
Knowledge ArchitectureFollows the “A is a B that C” definition pattern (genus + differentia)
Knowledge Nodes10-dimensional epistemic taxonomy (Heuristic, AntiPattern, PreFlightCheck, etc.)
OWL 2 SerializationOutputs valid OWL 2 ontologies in RDF/Turtle format
SHACL ValidationConstitutional gatekeeper ensures logical validity before write
State MachinesSkills can define preconditions, postconditions, and failure handlers
Security PipelineDefense-in-depth: regex patterns + LLM review for malicious content
Static LintingDetects dead states, circular deps, duplicate intents
Drift DetectionSemantic diff between ontology versions

What gets compiled

Every skill is extracted with:

  • Identity: nature, genus, differentia (Knowledge Architecture)
  • Intents: What user intentions this skill resolves
  • Requirements: Dependencies (EnvVar, Tool, Hardware, API, Knowledge)
  • Knowledge Nodes: Epistemic knowledge (8-12 nodes per skill)
  • Execution Payload: Optional code to execute
  • State Transitions: requiresState, yieldsState, handlesFailure
  • Provenance: generatedBy attestation (LLM model used)

Components

ComponentLanguageDescription
ontoskillsCLIUser-facing installer and manager
OntoCorePythonSkill compiler for SKILL.md sources
OntoMCPRustMCP server with 5 semantic tools (incl. search_intents)
OntoStoreGitHub repoOfficial compiled skill store
skills/MarkdownHuman-authored source skills
ontoskills/TurtleCompiled ontology artifacts

Use cases

Use CaseHow OntoSkills Help
Enterprise AI AgentsDeterministic skill selection via SPARQL queries
Edge DeploymentSmaller models query large skill ecosystems
Multi-Agent SystemsShared ontology as coordination layer
Compliance & AuditEvery skill carries attestation and content hash
Skill MarketplacesOntoStore and third-party stores enable plug-and-play distribution

Next steps