Bridging LLM Agents and Data Spaces Using the Model Context Protocol

The Model Context Protocol enables seamless integration of LLM agents into data spaces while enforcing governance boundaries

This architecture introduces a mediation layer that bridges large language model agents with data space services using the Model Context Protocol. The Eunomia Agent converts data space functionalities into structured, schema-compliant tools—allowing AI agents to discover and execute them without altering existing infrastructure. A working prototype confirmed full workflows, including catalog searches, metadata queries, and service invocations, while preserving compliance and operational isolation. The findings highlight how protocol-driven mediation achieves standardized, policy-respecting AI integration in data-sharing environments.

The Model Context Protocol acts as the core interface, structuring interactions through two components: the model card and context JSON. The model card contains technical metadata—such as the LLM’s identity (e.g., a specific LLaMA variant), version, developer organization, and a link to its detailed model card—covering training details, tokenization, and performance benchmarks. Meanwhile, the context JSON tracks the interaction’s dynamic elements: conversation history, participant roles (user/assistant), and system prompts that guide responses. Together, these elements maintain an up-to-date model representation and contextual flow, letting the Eunomia Agent mediate interactions while upholding governance constraints.

This solution addresses key limitations of prior approaches, which often required direct AI access to data repositories, leading to security risks and blurred responsibility boundaries. Unlike those methods, this architecture preserves the data space’s existing protections, governance policies, and optimizations by introducing an interoperability layer that remains independent of the underlying infrastructure. The separation of concerns—achieved through the Eunomia Agent and Model Context Protocol—also supports scalability and future upgrades, reducing rollback risks compared to monolithic integrations. The prototype’s ability to invoke data services under strict policy enforcement demonstrates its practical advantage over earlier models.

The work was documented in an arXiv-issued submission (arXiv:2609.30341), pending DataCite DOI registration, with Andres Munoz-Arcentales as the lead author. Associated code, data, and media are available through alphaXiv, ensuring reproducibility. The findings provide actionable guidance for organizations deploying AI-driven automation in governed data-sharing ecosystems while maintaining compliance, interoperability, and clear architectural separation.