How to Set Up the New Dynamics 365 ERP MCP Server

Key Capabilities

The dynamic MCP server can interact with business data in D365 F&O through natural language. For example, a prompt like:

"Please create a customer with ID NL-002, Name is Next365 Netherlands, Group 20, Currency Euro and having the next address: Johan Huizingalaan 163A Amsterdam"

The MCP server, combined with the Claude Sonnet 4.5 model, can even infer missing information — such as recognising the country based on the city name.

Prerequisites

A Tier 2 or PPAC-deployed environment running Dynamics 365 F&O version 10.0.46

Setup Steps

1. Enable the Feature

Navigate to Feature management and enable (Preview) Dynamics 365 ERP Model Context Protocol server. If the feature is not visible, click Check for updates at the top-right of the list page.

2. Configure Allowed MCP Clients

Navigate to System administration > Setup > Allowed MCP clients. Here you can enable pre-configured clients (Copilot Studio, VS Code/GitHub Copilot) or add new clients by specifying the Client ID of an App registration in Entra ID.

3. Create a Copilot Studio Agent

In Copilot Studio, create a new agent and provide a name, description, and instructions.

4. Enable Claude Sonnet 4.5 Model

Microsoft recommends using the Claude Sonnet 4.5 language model from Anthropic for best results. External models are not enabled by default — a Global Administrator must allow them:

Open Microsoft 365 Admin Center (admin.microsoft.com)
Browse to Copilot > Settings > Data access tab
Click AI providers for other large language models and enable the Anthropic provider

Note: Enabling the provider can take several hours. After enablement, change the agent model from the default GPT-4.1 to Claude Sonnet 4.5.

5. Add the MCP Server as a Tool

Add the MCP server as a tool in your agent. Ensure you select the correct Dynamics 365 ERP MCP (Preview) server (both the previous and new preview servers may be available). Create or select a connection, then click Add and configure.

6. Test the Agent

Test the agent by asking questions in natural language. Responses may be slow in preview as data is sent externally and the MCP agent navigates the application by finding menu items, opening forms, and filtering data. The more specific the question, the better the results.

If you get consistently wrong results, adjust the agent's instructions to help it understand specific scenarios (e.g. distinguishing between transfer journals and movement journals).