Comunica MCP SPARQL Poster

Comunica MCP SPARQL

Improving the Accuracy of AI Agents
using SPARQL-based Access to Decentralized Knowledge Graphs

Ruben Taelman, Jitse De Smet

Full paper

Why AI Agents + Knowledge Graphs?

LLMs hallucinate and generate plausible but incorrect answers based solely on training data.
Knowledge Graphs provide accurate, structured, live data (DBpedia, Wikidata, Solid pods, …).
Model Context Protocol (MCP) lets AI agents call external tools at runtime without retraining.
Goal: expose any Knowledge Graph via a SPARQL-based MCP interface using Comunica.

AI Agent Querying Knowledge Graphs

How It Works

Requires only a single config change:

{
  "mcpServers": {
    "comunica": {
      "command": "npx",
      "args": ["-y", "@comunica/mcp-sparql",
               "--mode", "stdio"]
    }
  }
}

Works with any MCP client, such as Claude Desktop, GitHub Copilot, VS Code, …

MCP Tools Exposed to AI Agents

query-sparql

Execute SPARQL queries over one or more remote sources: SPARQL endpoints, Linked Data docs, Solid pods, TPF interfaces, HDT files. Supports updates, federation, HTTP proxies, and authentication.

query-sparql-rdf

Execute SPARQL queries over a serialized RDF dataset as a string. Accepts any RDF media type (Turtle, JSON-LD, N-Triples, …).

Also: mcp-sparql-solid, mcp-sparql-link-traversal, mcp-sparql-file, mcp-sparql-hdt

Preliminary Findings

LLMs effectively query DBpedia and Wikidata (internalized during pretraining).
LLMs struggle with less-known KGs (UniProt, DBLP) unless explicitly guided.
LLMs can hallucinate SPARQL endpoint URLs and fail at federated queries.
"What movies do Brad Pitt and DiCaprio both star in?"
→ Claude alone: 1 answer. With Comunica MCP: 2 accurate answers via DBpedia.
Future: KG discovery, schema tools, guided federation.

Conclusion

Connects AI agents to one or more KGs via SPARQL.
Towards trustworthy, explainable agentic AI.
Open-source (MIT), on npm: @comunica/mcp-sparql

npmjs.com/@comunica/mcp-sparql