How to Connect Sparx EA to Microsoft Copilot: The Technical Guide

Once a business stakeholder can ask Microsoft Copilot "which applications support our Finance capability?" and get an answer drawn from your live architecture model, the conversation about EA's value changes. This guide walks the actual technical path to that point — what has to be in place, the order it goes in, and where it most often breaks.

The short answer up front: Copilot does not read a Sparx EA repository directly. A connector product sits between them. Get that connector and the model behind it right, and the rest is configuration.

The connection model: how the data actually flows

There is no native MCP server inside Sparx EA core. Connecting the repository to Copilot means deploying a product that reads the repository and republishes its contents over the Model Context Protocol (MCP) — the open standard Microsoft Copilot uses to query external systems at the moment of the prompt.

That product is EA GraphLink, shipped as part of Kernaro AI Hub since January 2026. GraphLink is a read-only MCP server you deploy on a server for enterprise-wide access. It does one thing that matters here: it uses an MDG Technology defined for your repository to translate the physical Sparx schema into a GraphQL schema, and exposes that GraphQL layer through MCP. Copilot connects to it as a federated data source.

So the path is: EA Repository → EA GraphLink (MDG-to-GraphQL translation, served over MCP) → Microsoft Copilot. Two facts follow directly from that chain. Without EA GraphLink, there is no supported route from Sparx EA to Copilot. And without a well-governed MDG Technology and clean repository content underneath it, the connection returns answers that are technically valid and architecturally meaningless.

The connector is the easy part. The MDG model and the repository content beneath it decide whether Copilot becomes a trusted source or a confident liar.

Prerequisites: what has to be true before you start

Run these checks before touching any installer. Most failed integrations trace back to one of these being assumed rather than verified.

Repository prerequisites

• The Sparx EA repository is on a server database — SQL Server, Oracle, or MySQL — not a file-based .qea/.eap used for production.
• The ArchiMate MDG profile is active and element types are applied consistently.
• Key element types carry mandatory tagged values: Application Component (lifecycle status, owner, criticality), Business Capability, Technology Service.
• The package structure is organized by architecture domain — Business, Application, Technology, Data — with no mixed-domain packages.
• Repository governance score of 14+ on the 20-item checklist (see the Repository Governance Checklist).

EA GraphLink prerequisites

• A Kernaro AI Hub / EA GraphLink license in place.
• A deployment host allocated: GraphLink runs as a service with network access to the Sparx EA repository database.
• A service account with read access to the repository database.
• An HTTPS endpoint for the GraphLink MCP server — internal with M365 tenant network access, or externally reachable with appropriate security controls.
• An MDG Technology that defines how your repository's stereotypes and tagged values map into the GraphQL schema GraphLink exposes.

Microsoft 365 prerequisites

• An M365 Copilot license at a tier that supports external MCP / federated connectors. Confirm this with your Microsoft licensing contact — it is a specific capability, not present in every Copilot configuration.
• M365 tenant administrator access for connector configuration.
• A security review of the connector configuration completed with your security team.

The six-step technical path

With prerequisites confirmed, the build runs in a fixed order. Step 3 is a deliberate gate, not a formality — do not skip it.

What Copilot can and can't do with EA data

Set expectations explicitly before rollout. Copilot retrieves and synthesizes what is in the repository — it does not invent architecture.

Copilot can answer natural-language questions about portfolio composition ("what applications do we have in Finance?"), retrieve application details (lifecycle status and owner of a named app), summarize capability-to-application mapping, filter the portfolio by tagged values ("mission-critical apps that are phasing out"), and cross-reference technology and applications ("what runs on [platform]?").

Copilot cannot, without additional modeling, explain the rationale behind architecture decisions (unless decision records are captured with enough structure), compare baseline to target architecture (unless both are modeled and separated by package scope), generate new architecture recommendations, or guarantee freshness — answers reflect the last GraphLink index, not live repository state.

Troubleshooting the common failures

Empty results for valid queries. Check that the connector scope covers the packages where the expected elements live. If you restricted package GUIDs in step 2 and the queried elements sit in an excluded package, results come back empty. Confirm the GraphLink index actually includes those elements.

Incorrect application details. The index is probably stale — re-run it and retest. If it persists, check that the tagged values in question are mapped correctly in the GraphLink MDG configuration for that element type.

Retired or test elements showing up. Add Lifecycle Status = Retired as an exclusion filter in the GraphLink MCP server configuration. Test elements should be excluded by package scope.

M365 connector validation fails. Confirm the HTTPS endpoint is reachable from the Microsoft 365 validation range, the TLS certificate is valid, and the API key has not expired.

Where this fits

This is the technical core of a broader move toward AI Augmented Architecture: making the architecture you already maintain answerable in the tools your stakeholders already open. Sparx Services delivers the connection as part of an AI Power Tools for EA engagement — GraphLink deployment, MCP server and scope configuration, the pre-deployment MDG quality assessment that step 3 depends on, and post-deployment validation. Your IT team provides the host, network access, and HTTPS endpoint; your M365 administrator configures the Copilot connector with our guidance.

FAQ

What does EA GraphLink do in this integration? It reads the Sparx EA repository and, via an MDG Technology, translates the physical Sparx schema into a GraphQL schema that it serves over MCP. That translation is what makes repository content queryable by Copilot. Without it, Copilot has no supported path to Sparx EA data.

Is MCP the same as Microsoft's Copilot API? No. MCP is an open protocol for connecting AI systems to external data sources, not a Microsoft-proprietary API. Microsoft Copilot supports MCP as a federated-connector standard, which is how GraphLink's MCP server plugs in. The same protocol is used by other AI tools such as Claude and Cursor.

What MDG quality is needed before connecting? As a practical floor: the ArchiMate MDG active and consistently applied, mandatory tagged values enforced for Application Components (lifecycle status, owner, criticality), and capability-to-application mapping populated for your primary domains. A Repository Governance Checklist score of 14+ is a reasonable proxy; below that, expect incomplete or misleading answers.

What M365 Copilot license is required? The tier that includes external MCP / federated connector support — the full M365 Copilot product, not Copilot features embedded in individual apps. Verify with your Microsoft account team, since tiers and connector availability continue to evolve.

Can I test before rolling out to stakeholders? Yes, and you should. Scope the M365 connector to a test group (the architecture team) and run the full validation set first. Do not expose the integration to business stakeholders before validation is complete.

What happens if Copilot gives a wrong answer? Wrong answers point to a stale index, an MDG quality gap, or a scope misconfiguration. The handling process is the same each time: log the answer, trace it to the underlying data issue, fix the data, and re-index. A stakeholder feedback channel for flagging bad answers is essential governance, not a nicety.