Bridging LLM Agents and Data Spaces Using the Model Context Protocol

This article presents an architectural mediation approach based on the Model Context Protocol to enable controlled interaction between large language model (LLM) agents and data space services. It uses the Eunomia Agent to translate data space capabilities into structured, schema-driven tools that AI agents can discover and invoke without modifying existing components. This approach was validated through a prototype implementation, proving successful end-to-end interaction across tasks such as catalog discovery, metadata retrieval, and data service invocation. The results underscore the potential of protocol-based mediation to achieve interoperable and standards-aligned integration of AI agents into data space ecosystems while upholding governance constraints. The findings offer practical guidance for organizations looking to introduce AI-driven automation into governed data-sharing environments while preserving compliance, interoperability, and clear architectural boundaries between different concerns.
2609.30341

What is the Model Context Protocol?

The Model Context Protocol (MCP) serves as a crucial interface in this architecture, providing a standardized way for LLMs to understand and interact with data spaces. It encodes various pieces of contextual information into two key elements: the model card and the context JSON. The model card acts as a metadata container that includes detailed information about the large language model itself. This information encompasses the identity of the language model and its architecture type (e.g., a specific variant of a model like LLaMA), and extends to its training specifics such as model version, the organization responsible for its development, as well as a link to an accompanying model card that offers more in-depth insights about its training processes, tokenization rules, and performance capabilities.

The context JSON, on the other hand, focuses on the interaction environment at hand, capturing information crucial to the boundaries of the ongoing conversation with the user. Key elements included here are the conversation history recording what has been previously said, the roles of entities involved in the dialogue—whether they are an 'user' or the AI 'assistant'—and system prompt information that guides how the assistant should respond based on pre-defined instructions.

Together, these two components allow the system to maintain an updated representation of the relevant model and its contextual flow, enabling the Eunomia Agent to effectively mediate interactions while preserving the necessary governance and compliance parameters of the data space.

How does the prototype compare to prior solutions?

This approach represents a significant advancement over traditional models of integrating AI into data spaces by leveraging a robust architecture with well-defined boundaries.
Prior attempts at such integration often required AI agents to have direct access to the underlying data repositories which resulted in entanglement of responsibilities, control issues, and an erosion of the security perimeter between AI services and sensitive data stores. The interoperability layer introduced in this paper explicitly leaves the existing data space infrastructure untouched, ensuring that its protection measures, governance policies, and performance optimizations remain undisturbed.

By separating concerns through a mediator like the Eunomia agent that works with the Model Context Protocol, this architectural design enables scalability and future-proofing, making it possible to introduce advanced AI automation without proving time-consuming or risky to roll back, as is sometimes the case in monolithic solutions that deeply embed AI capabilities into data space platforms. The ability to consistently invoke data space services while keeping strict policy enforcement is one of the hallmarks of this prototype demonstrating its validity and practical edge over earlier models.