MCP matters because it changes the relationship between AI and enterprise systems. Instead of asking a model to hallucinate an integration, you give it a small set of explicit tools. Each tool has a name. Each has a schema. Each has a boundary.

The agent should not get a skeleton key. It should get a toolbox.

Why MCP fits JDE

JD Edwards is already full of carefully constructed boundaries: applications, roles, tables, business functions, AIS services and Orchestrations. MCP does not replace those boundaries. Used well, it exposes them to AI in a way the agent can understand.

For JDE, the most natural first MCP server is an AIS Orchestration server: list what the configured user can call, inspect the schema, execute the requested Orchestration and return the structured result.

The three-tool pattern

  • Discover. What Orchestrations are available?
  • Inspect. What inputs does this Orchestration require?
  • Execute. Run it with a constructed payload and return the response.

That is deliberately narrow. Narrow is good. Narrow means auditable. Narrow means understandable. Narrow means the agent has power without ambiguity.

Where file workflows fit

Many AI workflows need files as well as JDE actions: source spreadsheets, generated CSVs, logs, PDFs, extracts or response documents. A separate file I/O MCP server can give the agent access to a configured folder without giving it an open machine.

Small servers, strong boundaries

AI MCP Servers are intentionally small. A JDE AIS server should be a JDE AIS server. A file server should be a file server. The agent can combine tools, but governance can assess them one by one.

How this supports BrainStorm

BrainStorm uses the same architectural idea: conversation on the front end, approved tools behind it, documents and citations beside it, JDE identity underneath it. MCP is not just plumbing. It is the discipline that keeps the AI useful without making it unbounded.

When the MCP client is remote

Local STDIO MCP is one problem; making those tools reachable from Microsoft Copilot or another enterprise AI client is another. Remote access introduces an identity boundary that deserves its own design. See Secure MCP for Copilot and Enterprise AI, MCP Gateway Comparison: What Happens to User Identity? and MCP User Identity vs Shared Service Accounts.

Steward is designed to put that enterprise boundary in front of compatible tool-oriented STDIO MCP servers while keeping access to configured protected resources bound to the actual requesting user identity, without substituting another backend identity. The same principle can extend to file shares, servers and databases using integrated OS authentication.

What to avoid

Avoid a single “do anything” server. Avoid raw database write tools. Avoid ambiguous tool names. Avoid service accounts where user identity matters. Avoid letting the agent discover power that no human has consciously approved. MCP is powerful because it can be precise; precision is the point.

Need an AI tool surface for JDE?

Expose the Orchestrations you want agents to use, and no more. Small MCP servers make that practical, testable and governable.

Discuss your workflow View related products