An introduction to MCP on Azure

-

Over the last few weeks I seem to be having more and more conversations about MCP servers. Be it with customers looking at agentic AI solutions, or announcements from the big tech providers. 

With this in mind, I thought I'd write this blog to help engineers and architects understand what I've learnt about how they fit into the agentic solutions that data science teams are building today.

What are MCP servers?

An MCP, or model context protocol, server is ultimately an interpolation layer that makes it easier for data science teams to integrate APIs, data sets, etc into large language model (LLM) solutions.  Within these setups, an API, data set, etc is referred to as a tool.

Historically, integrating one tool into an agentic solution has been relatively straight forward; but, that effort doesn't scale linearly. As more and more tools, the complexity rapidly increases - increasing both development and maintenance costs.

To solve this challenge, Anthropic introduced MCP. This protocol standardises the way that tools are integrated into agentic solutions and adds descriptions around the tool - meaning that the LLM can identify the right tool based on the user prompt, and use this to dynamically create the final prompt that will be sent to the LLM to answer.


How does MCP work?

With this flow in mind, we can now look at how MCP works. Behind the scenes, MCP comes in two elements:
  1. MCP client
  2. MCP server
In terms of MCP clients, most of the common LLMs available today have an MCP client built into them. They are designed to understand this protocol and how to work with it. 

Turning to look at MCP servers, these come in two flavours:
  1. COTS solutions from current software vendors such as Microsoft's recent announcements around MCP for Dataverse, Fabric, etc.
  2. Building and hosting your own MCP server (potentially accelerated by many of the open source projects available today).
Either way, we end up with the following setup:



What does this mean for Azure?

Across blog announcements, build, and Databricks Data+AI summit we have seen a number of announcements in this space over the last month. By each product we have seen announcements for the following:

  • Microsoft Fabric. Microsoft have released an open source MCP server for Fabric. Today it allows users to add RTI solutions into Fabric so that LLMs can write KQL against Eventhouses. Microsoft have also announced a number of additional Fabric objects will be added soon (check the blog for more details).
  • Azure MCP server. In public preview, Microsoft have released Azure MCP server. This allows LLMs to interact with a number of Azure components such as Cosmos DB, Azure CLI, etc. Check out the associated blog for more info.
  • Dataverse. We now get an MCP server for all data that is stored in the dataverse. Making it easy for those running D365 to integrate their data into agentic solutions.
  • Copilot studio. MCP support has been added to copilot studio, making it easier than before to add tools to the copilots being built.
  • AI foundary. Has all the tools you need to build your own MCP server for use within agentic solutions - as well as being integrated with the Azure AI agent service.
  • Azure Databricks. At summit Databricks announced MCP support for both unity catalogue and Mosaic AI, making it easier than ever to integrated data hosted in Databricks in agentic solutions.
With all these announcements, one thing is clear, MCP is here to stay and it will be the new defecto standard for integrating tools into LLM based solutions.

Comments

Post a Comment

Popular posts from this blog

Ignite 2024 - What's been announced for Microsoft Fabric

-
Image
With how popular my thoughts after Fabcon EU's keynote, I thought I'd do the same with the announcements out of Ignite. Again, we follow the themes of AI everywhere- with announcements grouped into: AI-powered Data platform Open and AI-ready data Lake AI-enabled business users Mission-critical foundation On top of these themes, we're seeing the number of items without CI/CD support being reduced. Long overdue, but at least it's now being resolved - and will be gone by the end of the year. AI-powered data platform This grouping is all about providing data teams with the tools they need to do their jobs. SQL Databases First up, we have SQL databases as a new Fabric component - basically a SaaS version of the Azure SQL DB that has been used for years. At face value, it seems like it's a small change but actually when you think about it, this looks to be a shift from Fabric being a purely analytical platform to an analytical and operational platform. Especially when you...
Read more

Power BI - Fabcon keynote, preview features, and March 2025 announcements

-
Yesterday we had an whole host of new features announced in the keynote ( see my first part in this series for more information). Whilst the main announcements focused on Fabric features, we do have a number of announcements to cover specifically around Power BI.  If this proves popular, I'll be moving to seperating out announcements for Power BI and the wider Fabric platform separately. But before we get going on new features, let's address the elephant in the room ... Power BI Price increases As of today, the list price for Power BI has gone up by a significant amount (e.g. Premium per user going from £15.00 to £18.50 per user per month - about a 19% increase). From what I'm seeing this increase is pretty much causing the pushback that partners warned Microsoft would happen. If you are licensing via O365 E5 licensing you are excluded from this price change. The cloud in the silver lining is that whilst it may not seem like Power BI has had many new features in the last ye...
Read more

Workspace topologies in Microsoft Fabric

-
Image
Since I first heard about Microsoft Fabric during private preview, I have spent many hours pondering the workspace topologies available to us. With this workspace based setup it gives us the power to spin up workspaces quickly and easily. But with that power comes great responsibility. But why is that? As time goes on I can see us struggling from workspace proliferation in exactly the same way that BI has suffered from KPI and report proliferation. That is multiple copies of the same workspace but designed for slightly different purposes. To counter that we need to carefully plan and curate our workspace creation, meaning that we need a set of topologies that we can follow to make our lives easier. But what is a workspace topology? The best way to think of it is a blueprint for how we're going to create workspaces and link them together. As I thought about this more and more, three clear topologies began to emerge. I've called these: Monolith warehouse Hub and Spoke Data Mesh M...
Read more