Exposing A Custom Class to Dynamics 365 MCP Server

 

Breaking the Forms: Direct Code Invocation for Agents in Finance & Operations Apps

Most AI agents interacting with enterprise ERPs—like Microsoft Dynamics 365 Finance & Operations—operate like super-fast human users. They navigate forms, click buttons, and fill in fields. This "form interaction" pattern is powerful because it instantly makes millions of existing UI-based actions available to your AI.

But what if the business logic you need isn't on a form? Or what if navigating a complex wizard is too slow and fragile for a critical backend process?

This is where the find_actions and invoke_action tools shine. They allow developers to bypass the UI layer entirely and expose direct code execution to the Model Context Protocol (MCP) server.

The "Escape Hatch": Direct Code Access

While standard MCP tools rely on form metadata, find_actions and invoke_action are special. They act as a bridge directly to your application's X++ classes.

• find_actions: Allows the agent to query the system and discover which custom code classes are available for it to use.
• invoke_action: Allows the agent to trigger a specific class and pass parameters directly to it.

This capability is essential for scenarios where complex calculation engines, heavy data processing, or headless business logic need to be triggered by an AI agent without a user interface.

How it Works: The Architecture

To make a piece of code accessible to an agent, you don't just "open up" the codebase. You follow a structured framework to ensure security and discoverability.

The Workflow:

1. Developer Action: A developer writes a class in the Finance & Operations app that implements the ICustomAPI interface.
2. Security: The class is associated with a Menu Item. The security roles assigned to that Menu Item determine who (and which agents) can call it.
3. Registration: The class is registered as an AI Tool in the Synchronize Dataverse Custom APIs form.
4. Discovery: The MCP Server uses find_actions to see this registered class.
5. Execution: The Agent uses invoke_action to run the code.

Developer Guide: Exposing Your Logic

If you are a developer looking to build these tools, here is the high-level checklist:

1. Implement ICustomAPI

Your X++ class must implement this specific interface. This standardizes the input/output structure so the AI agent knows exactly how to talk to your code.

2. Security is Key

Just because it's code doesn't mean it bypasses security. You must link your class to a Menu Action Item.

• If the agent's user account doesn't have access to that Menu Item, find_actions will return nothing, and invoke_action will fail.
• This leverages the standard Role-Based Security (RBS) of Finance & Operations.

3. Sync to Dataverse

Once your code is ready and secured, it needs to be "advertised." You do this via the Synchronize Dataverse Custom APIs (CustomApiTable) form.

• Classes that are correctly configured will appear in this list.
• Syncing them makes them visible to the outside world (the MCP Server).

Summary

Feature	Standard MCP Tools	invoke_action Tools
Interaction Type	UI / Form-based	Direct Code Execution
Best For	Data entry, standard workflows	Complex logic, calculations, headless processes
Setup	Zero-code (uses existing forms)	Pro-code (requires X++ development)
Security	Form permissions	Menu Item permissions

By using invoke_action, you move beyond "simulating a user" to "integrating with the core," giving your agents the ability to perform complex, high-speed operations that go far beyond what a human could do on a screen.