AWS open-sources an ontology layer so agents stop guessing

What it does

It’s a semantic context layer for AWS: you connect data sources, it discovers schemas and ingests documents, induces formal ontologies from them, and builds a unified knowledge graph. Agents then query that graph — via SPARQL federation and an MCP server — instead of freestyle retrieval. The pitch is that context gets validated against business logic before it reaches the agent, so decisions are consistent and explainable rather than vibes-based.

The interesting bit

The Scan → Model → Serve pipeline is old-school data engineering wearing new clothes: ontology induction with HermiT/ELK reasoners, a Virtual Knowledge Graph via Ontop, and Smithy-generated API contracts as the source of truth. The genuinely modern piece is the MCP server — agents traverse the graph as tools, with namespace isolation and role-based access control deciding what they’re allowed to see.

Key highlights

• Three-stage workflow: scan sources, model ontologies, serve context to agents over MCP
• SPARQL federation through a Virtual Knowledge Graph (Ontop) over your existing data
• Access control with namespace-scoped roles plus platform-wide platform-admin/platform-viewer
• Smithy models generate OpenAPI specs, Python interfaces, and a TypeScript client
• Includes Athena federation connectors (reference plus Databricks) and a React/Cloudscape frontend

Caveats

• It’s a read-only mirror — no pull requests accepted, issues and feedback only
• The prerequisite surface is wide: Python 3.12, Node, Java 17–24 (21 recommended, 25 unsupported), Docker, Maven, Gradle, uv, pnpm. The README spends real time on JDK versioning traps, which tells you something
• The README recommends tagged releases over main, which is honest but also a hint about stability

Verdict

Worth a look if you’re building agent systems on AWS and want context that’s governed and auditable rather than stuffed into a prompt. If you’re not already comfortable with ontologies, SPARQL, and a multi-language monorepo, the learning curve will do the filtering for you.