How to expose Resources correctly?

[IvanMurzak]

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

Working on Unity-MCP server for Unity Engine. A bit confused with understanding the Resource subject. An average Unity project has thousands of assets, gameObjects and C# components (attachable to a gameObject).

It seems the protocol is expecting to expose FULL list of the resources because they are all in the current context… it doesn’t seems right.

Should I use Tools instead as a Get function to request more specified data request and allowing filtering? I do not see how to use resources/list in a helpful way.

[tianfudhe]

I have the same question.

• Under what condition should we design a client that can expose resources?
• and what to expose?
• Besides, when the developers are implementing a client of their own, when and how to leverage the resource of servers?

[tianfudhe]

also need more explaination for Context, notification, roots, sampling.

most open-source clients only support tool, making it difficult to study/understand how these concepts fit what scenarios.

[IvanMurzak]

My question is in the context of MCP server, not a MCP client. Anyway seems like it is messy a bit in the protocol if the same issue happens on the Client as well. Not many clients support it. There is feature support matrix

[hesreallyhim]

I have the same question.

• Under what condition should we design a client that can expose resources?
• and what to expose?
• Besides, when the developers are implementing a client of their own, when and how to leverage the resource of servers?

Server exposes resources, not client. If the resources are dynamic, and there's lots of them, then you use resource templates.

[hesreallyhim]

Besides, when the developers are implementing a client of their own, when and how to leverage the resource of servers?

A resource is literally any data the server wants to expose to the client, usually something like a file, a database entry, etc., totally application-specific:

Resources allow servers to share data that provides context to language models, such as files, database schemas, or application-specific information.

How you leverage them is basically asking "what should this server do"? It's a "how long is a piece of string" question.

[IvanMurzak]

Asked the same question in discord chat. And got a good answer in the discord channel. Let me forward it:

Message from Peder:

Resources are usually an application facing primitive. Most client hosts do not forward them directly to the model, but instead allow the user to interact with them in various ways or use them for specif purposes (like tool icons).

I would not recommend exposing assets as resources. Instead consider what functionality you want to expose to the model. Is it a Unity Editor integration? Is it for connecting to a running Unity Engine based executable?

If Editor then start "small", and provide tools for navigating project and scene hierarchicies and for creating assets, game objects and manipulating components on game objects.

Message from Matt Lenhard:

Support for resources is still pretty hit or miss for what its worth. You can see a feature support matrix here if you are curious. That being said, like Peder mentioned I wouldn't expose all of the assets as resources.

[rlb567]

Interesting. I've been looking into exposing a lot of data via MCP to allow the client/model to manipulate it for reporting and surfacing unknown facts. I was about to create a couple of million resources as a test, but was struggling to see how that would help a client.

Like you, I considered creating tools to return subsets, but if I create a tool then I'm predetermining some of what the user might actually want to do. My data is retail sales data and am investigating how we might use an AI to answer a simple query like "which products sell better when it is raining". I can expose all that as resources, even the weather at time of sale, but it is a lot.
Or I can create a resource template /product/{id}/{weather} but that's exponential in terms of possible combinations.

I think I'm starting to see that for my specific use case, I suspect MCP will need to add something specific to allow us to expose "tables" of data, and also to describe "output schema" in a standardised helpful way. But all ears for alternatives I haven't considered.

(Still love MCP, we can expose all the easy stuff already)

[hesreallyhim]

this is the purpose of resource templates. if new resources or even types of resources are being created, then server can provide the capability "listChanged" for resources (meaning the list of resources has changed)

Sorry if I'm not following you...

Or I can create a resource template /product/{id}/{weather} but that's exponential in terms of possible combinations.

what's exponential? the schema is just that line above - the values for the placeholders are "limitless" but that's just like any API...

[drusellers]

Ok, so the way that I'm reading this is that we SHOULD NOT expose thousands of resources, this makes sense given the current protocol but I guess I'm still confused as to what a resource is then. :D That's ok it's still early days.

I feel a pull to expose data about my project state (in this case a RabbitMQ cluster) but maybe I return that data via a tool - and that tool can have what ever "filters" are needed. Like give me all "durable queues" etc.

[evalstate]

In that situation, a Resource Template is probably more appropriate.

For a CRM MCP Server for example you could have a URI scheme such as:
my-crm-server://customers/id/23222.

But you wouldn't list all of your Customer ID's in the List, so you can use a template instead:
my-crm-server://customers/id/{customer_id}.

This means that tool calls can work with Customer Ids, or embed resources using that URI scheme and get a consistent response.

Thing to figure out is which parts of your application should view/manipulate resources. In some cases, you want the LLM to conduct inference on them directly, in others you may want the LLM to instruct the MCP Server to do something with a Resource.

For the Unity example, this pairing with roots might also be appropriate - where file:// resources can be made accessible to Host and Server.

[hesreallyhim]

Ok, so the way that I'm reading this is that we SHOULD NOT expose thousands of resources

Why not? Every single file on your machine could be a resource. That's why you use a Resource Template. So I think I agree with @evalstate