MCP 服务器

mcp-d365-connector

Enables natural language querying of Dynamics 365 Finance & Operations by using Azure OpenAI to generate OData queries, fetch data, and refine responses.

README

MCP Server for Dynamics 365 Finance & Operations

This project demonstrates how to build a custom Model Context Protocol (MCP) server that integrates Azure OpenAI with Dynamics 365 Finance & Operations (D365 F&O). The server leverages natural language input to generate valid OData queries, fetch data from D365, and process results using a Large Language Model (LLM).

🔧 Architecture Overview

![Architecture Diagram] alt text

Components:

Client (User): Sends natural language queries.
MCP Server (FastAPI): Parses the input, uses Azure OpenAI to generate an OData query.
D365 F&O: Returns matching business data using OData.
Azure OpenAI: Used twice — once to generate the OData query, and again to post-process the response.
Output: Cleaned and focused data returned to the calling system.

📦 Technology Stack

Python 3.10+
FastAPI
Azure OpenAI (openai Python SDK)
Requests (for HTTP calls)
OAuth2 for secure access to D365 F&O

🛠 SDKs Used

OpenAI SDK – For communicating with Azure OpenAI deployments.
Requests – For REST API calls to D365 and token endpoints.
Pydantic + FastAPI – For building the RESTful MCP interface.

🌐 Public MCP Servers (Non-Microsoft)

While this implementation is custom-built for D365 F&O, several public MCP servers exist in the ecosystem — especially for open-source or OSS-focused LLM projects:

LangChain MCP Examples
Haystack LLM Agent Servers
DSPy Agent Context Servers
AutoGen Frameworks with context protocol layers

These are usually used for document retrieval, QA systems, and autonomous agents.

📘 Example Scenario

"Show me top 10 customers from California with credit limit over $10,000"

How It Works:

User sends a natural language request.
MCP Server uses LLM to generate this OData query:

CustomersV3?$filter=State eq 'CA' and CreditLimit gt 10000&$top=10

MCP Server calls the D365 F&O endpoint with this query.
Response is then passed back to the LLM for post-processing to extract just the Customer Name and ID.
Final cleaned output is returned to the caller.

💡 LLM Prompting Strategy

Two distinct prompts are used:

Prompt #1 (Generate OData Query) – Guides the LLM to only return a valid OData URL fragment.
Prompt #2 (Refine Response) – Extracts only CustomerAccount and CustomerName from the raw OData response.

📂 Getting Started

Clone the repo.
Copy .env.sample to .env and fill in your keys.
Run the server:

uvicorn main:app --reload

🔐 Security

Make sure .env is added to .gitignore. Do not commit your credentials or secrets.

Testing

Use CURL or tools like PostMan to send API call: Invoke-RestMethod -Uri http://127.0.0.1:8000/api/mcp -Method POST -Body '{"name":"Test","context":"Get Customers and first only"}' ` -ContentType "application/json"