Does the integration work with Health Cloud or Financial Services Cloud?

Yes. The connection is at the platform level, so it works in any Salesforce cloud that has Agentforce enabled. The industry clouds add their own data objects; the EA connection adds architecture intelligence to whichever context you are working in. The configuration steps are the same.

Insight · Integration

How to Connect Sparx EA to Salesforce Agentforce via MCP: Implementation Guide

Q: Is one EA connection available to all Agentforce agents, or is it per-agent?

The MCP connection is registered once at the org level. Individual agents then reference it in their topic configuration, so a customer-service agent, an internal support agent and an architecture-review agent can all query the same repository with different scopes and instructions.

Q: What Agentforce release supports external MCP servers?

Agentforce gained native MCP client support through 2025, reaching beta in early 2026, and hosted MCP connections are generally available for Enterprise Edition orgs and above. Confirm your org is on a current release and that the MCP feature is enabled before registering EA GraphLink.

Q: How is authentication handled between Agentforce and a private EA GraphLink server?

For a private deployment, a secure outbound path is needed so Agentforce, which runs in Salesforce's cloud, can reach the server. Options include Salesforce Private Connect, a hardened reverse proxy with IP allowlisting, or hosting EA GraphLink in a cloud environment with an authenticated public endpoint.

By Sparx Services · February 24, 2026 · 7 min read

If your organization runs on Salesforce, the case for putting architecture intelligence inside Agentforce writes itself. The systems that support a given account, the platforms reaching end-of-life, the dependencies behind a customer-facing service — that knowledge already lives in your Sparx EA repository. The work is making it answerable from the place your customer-facing and operations teams already spend their day.

The connection runs over the Model Context Protocol (MCP) — the same standard you would use to wire Sparx EA into Microsoft Copilot. What changes is the registration surface and the user surface: you configure it in Agentforce rather than Copilot Studio, and the answers land in Salesforce CRM and Service Cloud rather than Teams and Outlook. One MCP server can feed both ecosystems at once.

What this covers

Prerequisites — what has to be in place on the Sparx EA side and the Salesforce side before you start.
The six-step build — registering the MCP server, configuring an Agentforce topic, testing queries, hardening for production, building actions, and optionally joining EA data through MuleSoft.
Where quality comes from — why repository governance, not the connector, sets the ceiling on answer quality.
Common questions — concurrency with Copilot, per-agent reuse, release requirements, private-network auth, and access scoping.

There is no native MCP server inside Sparx EA.

The MCP endpoint that Agentforce connects to is provided by EA GraphLink, part of Kernaro AI Hub. EA GraphLink is a read-only MCP server deployed on a server for enterprise-wide access. It relies on an MDG Technology defined for your repository, which translates the physical Sparx schema into the GraphQL schema that MCP exposes. Get that translation right and Agentforce sees clean, queryable architecture; skip it and the answers degrade.

Before you start: prerequisites

Three things determine whether the integration is a smooth afternoon or a week of firewall tickets. Confirm all of them first.

EA GraphLink is deployed and reachable

The EA GraphLink MCP endpoint is active over HTTPS and you can confirm it responds.
Authentication is configured — an API key or OAuth 2.0 credential is available for the Agentforce integration.
The MDG Technology backing the repository is in place, so the element types and query patterns your Agentforce use cases need are actually exposed through the GraphQL schema.
You have tested the endpoint independently (curl or Postman) before touching anything in Salesforce.

Agentforce is licensed and enabled

Agentforce is available in your Salesforce org, with the MCP connection feature turned on.
You hold System Administrator or Agentforce Admin permissions in that org.
Your org is on a current release — Agentforce reached MCP client beta in early 2026, and hosted MCP connections are generally available for Enterprise Edition and above.

The network path is open

Agentforce, running in Salesforce's cloud, must be able to reach the EA GraphLink endpoint.
If EA GraphLink sits inside your corporate network, you will need Salesforce Private Connect or a hardened outbound gateway to bridge cloud to private network.
If EA GraphLink is hosted in a cloud environment with an authenticated public endpoint, connectivity is far simpler.

Agree the scope before you configure anything. Decide which questions Agentforce should answer from the repository. Salesforce-specific use cases tend to look like: "Which systems support the customer portal this account uses?", "What does our Customer 360 platform actually consist of?", or "Which applications affecting customer service are reaching end-of-life?" These questions shape the topic you configure in step 2.

The six-step build

The sequence below takes you from a registered data source to working agentic actions. Steps 1 through 4 stand up the read path; step 5 is where Agentforce starts doing things with the data; step 6 is an optional integration workstream for organizations that need EA data joined with CRM data.

Register EA GraphLink as an external MCP source

In Salesforce Setup, add EA GraphLink as an external data source of type MCP (Model Context Protocol). Give it a clear label (Enterprise Architecture Repository), point the server URL at your EA GraphLink endpoint, and supply the API key generated for the Agentforce integration. Save, then run Test Connection — a healthy result returns the server's capability declaration. If it fails, check the URL, the key for stray whitespace, your firewall's outbound allowlist, and the EA GraphLink logs for the connection attempt.

Configure an Agentforce topic for EA queries

A topic defines what an agent can talk about and do. Create one called Enterprise Architecture Intelligence. Its description tells the agent it can answer questions about applications, technology platforms, business capabilities, dependencies and architecture decisions — and that it must query the repository for specifics rather than relying on general knowledge. Its instructions reinforce that: always query the EA source, and if the repository returns nothing, say so and point the user to the architecture team. Grant the topic the action to query the EA source plus draft a follow-up summary.

Test with representative queries

In Agentforce preview mode, run the questions your users will actually ask. "Which applications support our Customer Onboarding capability?" should return the realizing application components. "Are any customer-facing applications reaching end-of-life in the next 12 months?" should return names and dates. "What does our CRM platform depend on?" should return connected systems. "What are our cloud infrastructure standards?" should return the relevant accepted decisions. If the names don't match what your architects modeled, you are usually looking at a governance gap — inconsistent stereotypes or naming — not a connector fault.

Harden the topic for production

Refine based on what testing surfaced. If queries return too much (all 400 applications when a filtered answer was wanted), add scoping instructions — for example, filter customer-facing questions to a named business domain. If the repository holds content that shouldn't reach a customer-service context, add exclusion rules (do not surface the Security Architecture package). Add clarification handling for ambiguous queries, and set the response register to its audience: plain language for customer-facing agents, precise terminology for internal technical ones.

Build actions from EA data

This is where Agentforce moves past answering questions. Wire EA data into Salesforce workflows: when the agent spots an end-of-life system in a service context, have it create a Salesforce task for the EA queue, auto-populated from repository data. When a case involves a system the repository flags as high risk, set an architecture-risk indicator on the case and add the detail as a comment. When complaints cluster around a capability, query the Business Owner tagged value on that capability element and notify the owner through Chatter.

Optionally, join EA data through MuleSoft

MuleSoft is the Salesforce-stack counterpart to Microsoft Fabric: it lets EA data be combined with other enterprise data for richer intelligence. Deploy a MuleSoft API that reads EA GraphLink's GraphQL interface on a schedule into a managed store, join it with CRM data (accounts, products, cases), and expose the enriched set through an Experience API for dashboards — or back to Agentforce as a secondary context source. This answers questions the direct connection can't, like "which applications support products sold to our top 20 accounts?" It is a heavier workstream, warranted only when EA and CRM data genuinely need to meet.

Where answer quality actually comes from

It is tempting to treat this as a connectivity project. It isn't. The connection is the easy part. The quality of what Agentforce returns is bounded by the governance of the repository behind it. If your application portfolio uses inconsistent stereotypes, half-populated tagged values, or naming that drifts between teams, those gaps surface directly in the agent's answers — now in front of customer-facing staff. The MDG Technology that EA GraphLink depends on is also your quality gate. Time spent tightening it pays back every time an agent runs a query.

This is the same discipline behind every integration we deliver, and it is the heart of AI Augmented Architecture: the repository is the context layer, and its quality sets the ceiling on everything built on top of it. The Salesforce path and the Microsoft path differ only at the registration and user surface — the engineering that makes either one trustworthy is the same.

Frequently asked questions

Can Agentforce and Microsoft Copilot both connect to the same EA GraphLink server?

Yes. EA GraphLink supports concurrent connections from multiple MCP clients, so Agentforce and Copilot can both query the same Sparx EA repository from a single deployment. For organizations running both Salesforce and Microsoft 365, that means EA intelligence in both ecosystems off one server, with no conflict between the connections.

Is one EA connection available to all agents, or is it per-agent?

The MCP source is registered once at the org level. Individual agents access it by referencing it in their topic configuration, so a customer-service agent, an internal support agent and an architecture-review agent can all query the same repository with different scopes and instructions.

What Agentforce release supports external MCP servers?

Agentforce built up native MCP client support through 2025, reaching beta in early 2026, and hosted MCP connections are now generally available for Enterprise Edition orgs and above. Confirm your org is on a current release and that the MCP feature is enabled before you register EA GraphLink.

Does it work with Health Cloud or Financial Services Cloud?

Yes. The connection is at the platform level, so it works in any Salesforce cloud with Agentforce enabled. The industry clouds add their own data objects and processes; the EA connection adds architecture intelligence to whichever context you're in. The configuration steps don't change.

How is authentication handled with a private EA GraphLink server?

For a private deployment, a secure outbound path lets Agentforce — which runs in Salesforce's cloud — reach the server. Options include Salesforce Private Connect, a hardened reverse proxy with strict IP allowlisting of Salesforce's egress ranges, or hosting EA GraphLink in a cloud environment with an authenticated public endpoint. We design the network path as part of the integration work.

Can we limit which EA data Agentforce can read?

Yes. EA GraphLink access controls restrict which packages, element types and tagged-value dimensions are exposed. Configure a dedicated key with read access only to the packages relevant to the Agentforce use cases — say, the application portfolio and business capabilities, but not security architecture — so sensitive content never reaches the Salesforce context.

For Salesforce-first organizations, this is how EA intelligence stops living in a tool only architects open and starts informing CRM, service operations and Agentforce workflows. Configure the Solution is where we stand the integration layer up and target the queries your teams will actually run.

Bring your architecture into Salesforce.

Talk to a practitioner about connecting your Sparx EA repository to Agentforce — from EA GraphLink deployment through agent topics and the first agentic actions.

Book a call →

Keep reading