.NET: Claude Skills support

[shaofing]

Claude Skills support

[eavanvalkenburg]

Can you eloborate, which language are you using, we already support skills for Anthropic in Python: https://github.com/microsoft/agent-framework/blob/main/python/samples/getting_started/agents/anthropic/anthropic_skills.py @rogerbarreto can talk about what we support for .Net

[shaofing]

.Net

[stephentoub]

It's already supported in .NET, though it's a tad klunky at the moment and will soon be getting simpler. Today, you can do it with code like this:

ChatOptions options = new()
{
    RawRepresentationFactory = _ => new MessageCreateParams()
    {
        Betas = ["skills-2025-10-02", "code-execution-2025-08-25"],
        Container = new(new BetaContainerParams()
        {
            Skills = [new BetaSkillParams { Type = BetaSkillParamsType.Anthropic, SkillID = "pptx", Version = "latest" }]
        },
    }
};

Once anthropics/anthropic-sdk-csharp#58 merges and ships, it'll instead look like:

ChatOptions options = new()
{
    Tools = [new BetaSkillParams { Type = BetaSkillParamsType.Anthropic, SkillID = "pptx", Version = "latest" }.AsAITool()]
};

or that tool could be passed to the ChatClientAgent's ctor.

cc: @PederHP, @sd-st

[markwallace-microsoft]

@rogerbarreto it would be good to have a sample for this if we don't already

[PederHP]

One thing to note is that the tools support is only part of this. You also need to call the files endpoints in the anthropic-sdk for many use cases where you work with skills (either upload, download, or both).

I suppose general abstractions for Agents to support files would be useful, but I'm not entirely sure how this would look. Whether it is a special Agent class that wraps the file endpoints or it is extension based.

One could also argue the abstraction of making a file available to an AI model could be relevant in Microsoft.Extensions.AI (as an extension to the IChatClient builder? WithFileClient?).

AG-UI supports this kind of interaction via:

interface UserMessage {
  id: string
  role: "user"
  content: string | InputContent[]  // Can be multimodal!
}

type InputContent = TextInputContent | BinaryInputContent

interface BinaryInputContent {
  type: "binary"
  mimeType: string
  id?: string      // Reference by ID
  url?: string     // Reference by URL
  data?: string    // Inline base64 data
  filename?: string
}

I think an IChatClient could meaningfully extract files from user messages and use an IFileClient (or other abstraction) to send them to the Model API. But there's no obvious candidate for file download in AG-UI.

So, while enabling skills in the API is doable (and will be more convenient with the extension once it is merged), there are still pieces of the overall puzzle missing to have a good developer experience around skills as a feature. Ideally it should be possible to handle files uploads/downloads related to the OpenAI Code Interpreter with the same abstractions.

[stephentoub]

IChatClient supports hosted file content and inline data content as part of requests/responses; code interpreter can send down either. In the case of hosted file content, though, you're correct that there's nothing exposed in IChatClient for interacting with remote file stores. That's by design, as it's a separate concept. We might want to consider adding an IFileClient to M.E.AI, but someone would need to design exactly what that looks like and propose it over in dotnet/extensions, having factored in the differences across the various Anthropic, AWS, Gemini, OpenAI, etc. providers. My expectation is if/when such a thing exists, you could access a corresponding IFileClient from an IChatClient via its GetService method.

[rodion-m]

@stephentoub Is this feature LLM-agnostic?