Question: How to configure Tool input parameters (like HttpClient)?

[willxie-eng]

I am seeking better understanding of how parameters are generated and passed into tools and if it's possible to customize input parameters (my scenario is I would like to set a custom handler for the HttpClient).

Using VS Code w/ GH Copilot as the MCP Client and debugging my C# MCP Server, I see that the client constructs a JSON object that represents the HttpClient, and the MCP Server deserializes this into an HttpClient object before passing it to the Tool.

{
  "httpClient": {
    "baseAddress": null,
    "defaultRequestHeaders": [],
    "defaultRequestVersion": "1.1",
    "defaultVersionPolicy": 1,
    "maxResponseContentBufferSize": 2147483647,
    "timeout": "00:01:00"
  }
}

My workaround is to not pass in the HttpClient at all and construct one inside the tool. It would be nice to have more control over this process or at least learn more about who generates the JSON (is it the LLM or MCP Client SDK or both?). TIA!

[stephentoub]

You've created an MCP server tool that accepts an HttpClient as a parameter? Are you trying to get that to be injected from DI? You wouldn't want to try to have that provided by the client.

[willxie-eng]

Thanks @stephentoub! Yep, making the McpServerToolType class not static and using DI works.

I'm interested in how the LLM is able to construct complex objects as inputs. I'm assuming the McpServerTool attributes auto-generates the InputSchema for methods, and that schema is used by the MCP Client & LLM to create the corresponding JSON object to pass to the server?

[stephentoub]

I'm interested in how the LLM is able to construct complex objects as inputs. I'm assuming the McpServerTool attributes auto-generates the InputSchema for methods, and that schema is used by the MCP Client & LLM to create the corresponding JSON object to pass to the server?

Yup. The arguments to the McpServerTool method can come from different places. For example, if you're using DI (or, more specifically, if an IServiceProvider finds its way to the McpServerTool.Create method under the covers), then any argument that can be satisified from that IServiceProvider are. A small set of types are hardcoded, e.g. a CancellationToken parameter is always bound to the token that's passed to the invocation, so that it flows from the caller into the tool method. All other parameters that aren't special cased end up having a schema generated for them. That schema down to the client as part the tool's description, the client passes that schema to the LLM as part of the tool description, and the LLM creates arguments according to that schema which it then sends back to the client. The client forwards that JSON from the LLM to the MCP server as part of calling the tool, and McpServerTool deserializes the JSON into the appropriate arguments.

The majority of that is handled by AIFunction from Microsoft.Extensions.AI. McpServerTool wraps AIFunctionFactory.Create. The schema generation functionality is also exposed separately as part of the AIJsonUtilities type from Microsoft.Extensions.AI.

[willxie-eng]

Thank you!