MCP Server Code Accessing URL Context Set by MCP Client?

[JeffFerguson]

Pre-submission Checklist

• I have verified that this discussion would not be more appropriate as an issue in a specific repository
• I have searched existing discussions to avoid duplicates

Discussion Topic

This may be in the documentation somewhere, but I thought that I would ask.

I would like MCP clients to be able to set a URL as context to the queries to my server (I think that these may also be known as "resources"). How do I access that contextual URL from my [McpServer] code? Is there anything that I need to do server-side to enable that URL-based contextual functionality? Right now, I have something like this:

[McpServerTool, Description("description goes here")]
public static async Task<string> IsValid(string url)
{
    return "stuff";
}

Ideally, I would like the clients to be able to set the URL as query context, rather than in the query itself.

[PederHP]

How do you want the client (host) to do this in practice. Resources aren't quite what you describe. They are a capability where the client can list/retrieve/subscribe to resources on the server.

Tools are normally passed to the LLM. If you don't want url in your example to be part of the tool schema, as given to the model (because it is set in other ways), then you shouldn't take it as an argument but your tool class should access through some other means (ie DI).

Your example confuses me a little, because if url was set from elsewhere the method is parameterless and you give the LLM an IsValid() tool, which will be called based on the description and the additional instructions provided (if the client utilizes that field in the capability exchange). That seems like an odd use case without knowing more.

Can you explain more about the context of what you're trying to do? When you say clients should be able to set the URL, do you mean the LLM, the application, or the user?

[JeffFerguson]

Thank you for responses! Let me offer a more concrete example.

From within Visual Studio Code, I can open up GitHub Copilot, which serves as an MCP client, and I can chat with my MCP Server in that user interface. Part of that user interface is an "Add Context" button, through which I can select a file, folder, or other entity, to offer context to the prompts I will be writing. In the example below, I have added .gitignore to the context.

When I now start writing prompts, the prompts will (presumably) act against .gitignore because I have set that context. (Please correct me if my understanding here is flawed.)

Given these conditions, here are my questions:

• What code would I write in my MCP Server to understand that .gitignore was the file selected for the context?
• Would I be able to "attach" a URL to some Internet resource to the context instead of a local file?

[PederHP]

You cannot access the context of the client from the MCP server in this manner, I am afraid. When you pick a file for context in VS Code it is added to the model context. The tools an MCP server provides are also added to the context, as are resources chosen by the user.

Also you're not chatting with your MCP server. The chat is with an LLM that the client (VS Code) has a different connection to. MCP provides additional context for that interaction. (there's something called sampling which lets the server use this LLM but it's in a limited fashion and not really relevant to this topic). The "add context" is for the LLM, not for your server.

You could create a tool with a list of "selected files" or "context files" and hope the model puts them in (if it has the file names) when it's called, but some would argue this is a security issue as users generally don't want their context leaking into remote MCP servers (as it may contains confidential information). And you'd have to prompt the model to remember to call your MCP method with all selected files. And if you want other users that use your MCP server to do this, it won't really work.

Another way to explain this is that the MCP server provides services to the client host (VS Code in this case), but the client doesn't share that much back. Pretty much nothing really. So you can provide tools for VS Code that the LLM in GitHub Copilot can use, but you cna't hook into what's going on there. You can do that with VS Code extensions I believe, but they're unrelated to MCP and they have some limitations as well, but might be able to do what you're looking for.

So you can't create an MCP server which lets you add urls as context in VS Code and then execute code against them. You could create a tool which takes a url (as in your example) and instruct the model to call it (in chat) and then do the is_valid and further instruct the model to then call another tool to fetch the page.

But, and I mean no disrespect, to make use that kind of convoluted tool calling and prompt interaction you should familiarize yourself better with the basics of MCP, and even then it's probably a fair amount of engineering effort.

There's more advanced stuff you could do with elicitation to trigger adding urls in the front end and checking them server side, but again it's complicated and needs a good understanding of how the protocol works.

There's certainly no easy "out of the box" way to do this.

[JeffFerguson]

Thank you very much for your insight!