Presents the Eunomia Agent, an MCP-based architectural mediator that enables controlled, policy-compliant interaction between LLMs and sovereign Data Spaces.
The Eunomia Agent is a data space agent that provides catalog management, dataset metadata, and data service registration, remaining unchanged and accessible via public REST APIs.
To enable controlled interaction, an external Architectural Mediation Layer operates between MCP clients and the Eunomia Agent, exposing data space capabilities as structured MCP tools (e.g., list_datasets, get_dataset_metadata, query_data_service) without modifying the core data space components.
This approach ensures policy compliance and governance by centralizing identity verification, contract negotiation, and policy enforcement within the Eunomia Agent, while the mediation layer handles translation and validation.
The material presents the Eunomia Agent, an architecture for mediating between large language model agents and sovereign Data Spaces using the Model Context Protocol (MCP). It addresses a practical gap in agentic AI systems: LLMs can reason, plan, and invoke tools, but they should not directly access governed data ecosystems without structured control, policy evaluation, and trust boundaries. The paper positions MCP as a protocol layer that allows LLM agents to discover, request, and consume data-space capabilities in a standardized way, while the mediator enforces compliance, authorization, and usage constraints on the data-space side.
Its key contribution is an architectural mediation pattern that separates the LLM runtime from the Data Space infrastructure. Rather than treating an LLM agent as an untrusted direct consumer of data services, the Eunomia Agent acts as a controlled interface that translates high-level model intent into governed operations—such as querying, retrieving, or invoking data-space resources—subject to policy checks, consent rules, and access controls. This design emphasizes interoperability, auditability, and least-privilege access, making it easier to integrate LLM agents into environments where data sovereignty, regulatory compliance, and organizational trust are first-class requirements.
The work matters because it offers a concrete path toward deploying LLM agents in data-governed, enterprise, or cross-organizational settings without compromising data-space principles. As MCP becomes a more common substrate for model-to-tool and model-to-context integration, the paper’s mediation approach is relevant for building systems that are both agentic and compliant: capable enough to use external data intelligently, but constrained enough to respect policy, ownership, and sovereignty boundaries.