Agentic Data Space Railway X

Description

The workshop shows how AI assistants such as Copilot, Claude, IDE-based agents or chatbot clients interact with a Tractus-X-based data space. The session uses Railway-X as the demonstration context and focuses on how an AI client can access data space capabilities without replacing the existing Industry Core Hub and REST-based integration layer.

It follows a practical structure: it starts with the fundamentals of the Model Context Protocol, the Industry Core Hub and the Railway-X scenario. It then moves into a Copilot demo and self-exploration, where participants connect an AI client to Railway-X. The final part discusses current limitations, productive use cases and architecture options for future agentic data spaces.

Responsible, Contact Person

• Johannes Schramm Manager Tech Innovation, Capgemini, johannes.schramm@capgemini.com
• Lena Hoffmeier Senior Consultant Data Strategy, Capgemini, lena.hoffmeier@capgemini.com

Prerequisites/Requirements

1. A running IC-Hub instance (v0.7.0+) with the MCP KIT addon enabled.

2. An API key or Keycloak credentials for IC-Hub authentication.

3. An MCP-compatible AI client (Claude Desktop, Claude Code, or Cursor)

Technical Architecture

The architecture combines three primary layers:

Layer	Technical Role
AI Client	User-facing interface such as Copilot, Claude, IDE agents or custom chatbots
MCP Server	Model-facing interface exposing tools, resources and prompts
Industry Core Hub	Central domain component containing business logic, configuration and authentication

The overall interaction pattern is:

User
  ↓
AI Client (Copilot / Claude / Agent)
  ↓
MCP Server
  ↓
Industry Core Hub Backend
  ↓
EDC & Data Space Services

The Industry Core Hub remains the central system component. MCP acts as a thin abstraction layer that allows AI agents to discover and invoke capabilities already provided by the hub.

REST APIs versus MCP Tools

One key message of the workshop is that MCP does not replace REST APIs. Instead, MCP introduces an AI-oriented interaction model.

Aspect	REST API	MCP Tools
Consumer	Applications, dashboards, partner systems	LLM-based AI agents
Discovery	Developers read documentation beforehand	Models discover tools at runtime
Contract	OpenAPI / Swagger	Tool descriptions + JSON schema
Operation Style	CRUD-oriented endpoints	Intent-oriented actions
Example	GET /parts/{id}/twin	get_digital_twin(part_id)

The REST API continues to serve as the system of record, while MCP provides an intent-based layer that makes the same business capabilities accessible to LLMs.

MCP Tool Coverage and Data Space Operations

The following table maps the available MCP tools to the underlying data space or local IC-Hub operation. It clarifies whether a tool triggers a data space interaction, reads or writes local IC-Hub data, or orchestrates multiple steps across DTR, EDC and HTTP transfer.

MCP Tool	DSP Operation	Notes
list_known_partners	Local database read	No data space call — reads the IC-Hub partner registry.
list_partner_twins	DTR catalog request	EDC negotiation followed by a partner DTR query for twin shells.
get_twin_details	DTR shell lookup	EDC negotiation followed by a specific AAS shell descriptor fetch.
list_twin_submodels	DTR submodel descriptors	EDC negotiation followed by listing submodel descriptors, without payload transfer.
fetch_submodel	Submodel HTTP transfer	EDC contract negotiation, then data transfer over HTTP.
fetch_partner_dpp	DPP submodel transfer	Auto-discovers the DPP submodel descriptor, then performs EDC + HTTP transfer.
fetch_dpp	Local submodel server read	No EDC — reads DPPs hosted by this IC-Hub.
fetch_partner_catalog	EDC catalog request	Connector Discovery by BPNL, then a DSP catalog request per EDC endpoint. Returns datasets and raw ODRL policies.
fetch_partner_asset	DSP negotiation + transfer	Connector Discovery, full contract negotiation for an asset_id, EDR token retrieval, then data-plane fetch.
list_my_catalog_parts	Local database read	No data space call — reads the IC-Hub catalog registry.
get_session_summary	In-memory session state	No calls — returns entities accumulated in this session.
create_catalog_part	Local database write	Creates a catalog part entry with status draft. No data space call.
update_catalog_part	Local database write	Updates catalog part metadata. No data space call.
create_serialized_part	Local database write	Creates a serialized part instance. Auto-creates missing catalog parts and partner mappings.
share_catalog_part	DTR + EDC orchestration	8-step flow: partner → agreement → twin → DTR registration → EDC asset → policy → contract → PartTypeInformation aspect.
register_business_partner	Local database write	Registers a partner entry. No data space call.
create_catalog_part_twin	DTR shell registration	Registers a PartType AAS shell descriptor in the DTR.
create_serialized_part_twin	DTR shell + aspect creation	Registers a PartInstance shell and auto-creates SerialPart V3 aspect.
attach_twin_aspect	Submodel upload + DTR registration	Uploads submodel payload and registers the descriptor in the DTR shell.
share_dpp	DTR twin sharing	Shares a DPP twin with a business partner via twin exchange.

Challenges and Current Limitations

The workshop explicitly highlights several limitations of the current architecture.

Challenge	Technical Description	Required Improvement
Hallucinations	LLMs may generate incorrect results	Validation and guardrails
Submodel correctness	Generated submodels are not validated	Automated schema and semantic validation
Policy handling	Policies are automatically accepted	Policy evaluation and approval workflows
Retry behavior	Models do not retry failed operations	Workflow orchestration and recovery mechanisms
Context consumption	MCP tools continuously consume context	Context optimization strategies
Missing business semantics	Agents understand structure but not meaning	Ontology-based semantic layer

Particularly important is the distinction between technical interoperability and business understanding. While MCP enables technical access to data space services, agents still require additional contextual understanding to interpret data correctly.