CERM MCP PoC

This proof of concept explores what happens when a LangChain agent gains structured access to a full Microsoft 365 account. Rather than screen-scraping or fragile automation, the integration is built on the Model Context Protocol (MCP) — a typed tool-calling standard that lets the AI reason about which action to take and then invoke it precisely.

The MCP Server

The centrepiece is a custom Microsoft Graph API MCP server implemented in Python and served over HTTP with FastAPI. It exposes 51 specialised tools covering:

• Outlook email — read, compose, reply, search, and organise threads
• Calendar — create, update, and query events across multiple calendars
• OneDrive — list, upload, and retrieve files
• Contacts — look up and manage address book entries

Authentication is handled via device flow, supporting both personal and work/school Microsoft accounts. To keep latency low, all Graph API responses are cached with AES-256 encryption, delivering up to a 300x performance improvement on repeated queries.

The Conversational Interface

A Gradio chat UI sits in front of a LangChain agent backed by Azure OpenAI. Users type natural-language requests — “summarise my unread emails from this week” or “block off Friday afternoon” — and the agent selects and chains the appropriate MCP tools to fulfil them. Communication between the Gradio app and the MCP server uses streamable HTTP, enabling real-time, bidirectional data flow without polling.

What This Demonstrates

The project proves that MCP is a practical bridge between production enterprise systems and LLM agents — far more reliable than prompt-engineered workarounds. The whole stack runs in Docker, making it straightforward to evaluate locally or deploy to a shared environment for demos.