How to register Azure AI Search as a resource?

[jaywork2020]

Pre-submission Checklist

• I have checked that this question would not be more appropriate as an issue in a specific repository
• I have searched existing discussions and documentation for answers

Question Category

• Protocol Specification
• SDK Usage
• Server Implementation
• General Implementation
• Documentation
• Other

Your Question

I have data stored in an Azure AI Search service.
How can I expose this as a resource that an MCP client can query and retrieve data from?

In other words, how do I register my Azure AI Search index so it's accessible to the client through the MCP ecosystem?

[PederHP]

There's no easy out of the box way to do this. It's not even certain you'll want to expose this as a resource, due to the way resources require the MCP Client host application to decide how to handle the resources and when to inject them into model context.

You can create a tool which wraps your Azure AI Search queries (https://learn.microsoft.com/en-us/azure/search/search-query-overview) and thus lets the model do agentic search. If you want to do classic application-layer RAG, then MCP currently doesn't have any support for it on the protocol level. There are some discussions and open proposals on how to add this, which would for example make it easier to wrap Azure AI Search semantic queries in a standardized way.

But currently you have to do a lot of the heavy-lifting yourself, and you get agentic / model-controlled search, with the advantages and disadvantages of that approach.

You could wrap some of the query types in a Resource Template, but I wouldn't recommend that - as I don't think any client applications exist which would work well with this approach.

[jaywork2020]

Thank you! I had the impression that exposing the Azure AI Search service as an MCP resource might let us handle semantic queries automatically. But I may have misunderstood the purpose or limitations of MCP resources. I’ll dig into it further and see what I can learn.

[PederHP]

I've made a proposal for a new capability in the protocol, to address what I think are shortcomings in Resource Templates, so I might be biased. You can use Resource Templates as a way to propagate arguments, but personally I would prefer a schema that looked more like tools for things like semantic queries and RAG. It gets really awkward to put an entire user message as an argument to a Resource Template (as it then goes into a "get-style" uri).

I expect to update my proposal (modelcontextprotocol/modelcontextprotocol#142) soon. I'll try to include AI Search as an example. There are other ways to accomplish this and other abstractions, but I actually think there is a bit of a void in the protocol when it comes to context enrichment / transformation that doesn't involve the model. I often see people look into Resources as exactly this kind of pseudo-tool. I am sure there are other valid abstractions to implement this - and please be aware of my bias given my active proposal. If you look further into it and agree with me that there is a gap in the protocol capabilities, it would be good to have more voices in the spec repo asking for something to provide this.