What Is the Sparx EA MCP Server? Technical Guide for AI Integration

The Model Context Protocol is an open standard that lets an AI assistant discover an external data source, understand what it contains, and query it at the moment of the question. Apply that to an EA repository and an assistant such as Claude or GitHub Copilot can answer questions about your architecture directly from the model, which applications are end-of-life, what a capability depends on, which diagrams reference a given component, instead of working from a stale export. The catch is that the repository has to be exposed through an MCP server first, and Sparx does not provide one out of the box.

Two products, two MCP servers

As of October 2026, two products give Sparx EA an MCP interface. They are not the same thing, and the difference matters more than the marketing suggests.

Put plainly: Kernaro AI Hub’s MCP server is the shared, server-side, read-only window onto the repository; AI Power Tools is the local, hands-on workbench that can change the model. Some practices doing AI-Augmented Architecture will want both, because they solve different problems. The AI tools for Sparx EA comparison sets them beside the other options, and a deeper walkthrough of the IDE routes lives in connecting Sparx EA to Cursor and Claude Code.

How the two paths fit together

Both servers read from the same underlying Sparx EA repository. What changes is where the server runs, who connects to it, and whether the assistant can write back.

What an MCP server exposes from the repository

Whichever product you use, the value depends on serving the model as structured, typed data rather than raw rows. A well-built EA MCP server makes the following available to a connected assistant:

Typed elements. Each element is returned as its modeled type, an ArchiMate Application Component, a Business Process, a Technology Node, not as a generic UML object with a stereotype string the assistant has to parse. That lets you ask for “all Application Components in Finance with lifecycle status End-of-Life” and get a clean, filtered list.

Relationships. Realization, Serving, Assignment, Composition and the rest are queryable with their source and target, so an assistant can resolve “which components realize the Customer Onboarding capability?” from the actual relationship graph.

Tagged values. Custom attributes defined by your MDG extensions come through as structured dimensions. A governed Lifecycle Status enumeration (Active, Migrate, Eliminate, End-of-Life, Decommissioned) becomes a vocabulary the assistant can filter and count on precisely.

Package structure and documentation. The repository hierarchy is navigable, so queries can be scoped to a package, and element notes, including the Context, Decision and Consequences of an Architecture Decision Record, are readable for questions that need narrative context as well as attributes.

Diagram metadata. Diagram names and the elements that appear on them are exposed. The assistant reads what a diagram contains; with AI Power Tools it can also drive the EA UI to validate a diagram visually.

Why MDG governance decides the answer quality

This is the part most teams underestimate. The quality of the AI’s answers is bounded by the governance of the model. An MCP server serves the repository as it actually is. Where elements are correctly stereotyped and tagged values follow a governed enumeration, the assistant receives precise, typed data and returns precise answers. Where the model is a pile of generic UML elements with ad-hoc notes, someone typed “EOL” instead of “End-of-Life,” or recorded lifecycle status as free text, the assistant inherits that mess.

Ask a well-governed repository “how many applications are in End-of-Life status?” and you get a number and a list. Ask the same of a repository where that fact lives in scattered free-text notes and the answer is unreliable, because there is nothing structured to query. This is exactly why we treat MDG Technology governance as a prerequisite, not an afterthought; the same argument runs through MDG Technology as your AI quality gate. Kernaro AI Hub’s reliance on an MDG-defined GraphQL layer for its MCP server makes this explicit: no governance, no usable schema.

Which AI tools can connect

Because MCP is an open standard, an EA MCP server is vendor-neutral in principle. In practice, which client works with which server is a product decision, so here is the current picture:

• Claude: Claude Desktop chat and Claude Code in the Claude desktop app, through the AI Power Tools extension (the server) and plugin (the skills).
• GitHub Copilot: Agent mode in VS Code, through AI Power Tools’ VS Code extension: the same MCP server, with the same skills loaded by VS Code.
• Other MCP clients, such as Cursor or an in-house application. These can connect to a server-side MCP endpoint such as Kernaro AI Hub’s. Confirm client support against Sparx Systems’ current documentation before you plan around it.
• Microsoft 365 Copilot, no live connection to AI Power Tools. Microsoft 365 Copilot answers questions, read-only, from a published Power BI semantic model built from your EA data. If you have Kernaro AI Hub, the Kernaro Connector for Power BI, an extension to Hub, is the preferred route to Power BI, and Copilot can also reach Hub’s read-only API through Copilot Studio or a Microsoft 365 connector.

Why this matters for the job

An MCP server serves two jobs. The shared, read-only one gets the repository’s answers to the people who ask, so the portfolio question does not need an architect in the middle. The local, read/write one lets AI do the modeling work: building the diagrams for every application and keeping the mappings current, with the architect directing it. By hand, both are slow, and the results go stale.

What the server cannot supply is the intent. The architect decides which questions the business should be able to ask, what the assistant may read and change, and what counts as complete. An assistant is only as good as the person prompting it, and the MDG governance above is the standard the answers are held to. For reporting in Power BI, Microsoft 365 Copilot, Fabric or a warehouse, the connection is a reporting database and not an MCP server, and the reporting page covers it.

We can teach your architects to work this way on your own repository, or build and update the model for you and hand it over. Sparx Services, a NovoCircle practice independent of Sparx Systems Pty Ltd, sells training, mentoring and consulting, not the tools. Kernaro AI Hub is licensed by Sparx Systems, separately from your Sparx EA modeling licenses. AI Power Tools for Sparx EA is free, with nothing to buy or activate. The mechanics are in the NovoCircle documentation (opens in a new tab), the free tooling is described on AI Power Tools for Sparx EA, and the options side by side are on AI tools for Sparx EA.

Frequently asked questions