[FEATURE REQ] Custom tools and format (cfg) support

[bharathm03]

Describe the feature or improvement you are requesting

This is a feature request to add support for the recently introduced "Custom Tools" feature in the OpenAI API. This feature enables defining tools that can accept a plain string as input, rather than a structured JSON object. It also includes the ability to constrain the model's output for these tools using a context-free grammar (CFG) with either Lark or regex syntax.

Additional context

This feature is already available in both openai-node and openai-python. It would be helpful my app.

OpenAI documentation link:
https://platform.openai.com/docs/guides/function-calling#custom-tools

[jsquire]

Hi @bharathm03. Thanks for reaching out and for your inquiry. Custom tools are not yet represented in the REST API spec used for client generation. Once the spec is updated to include this area and a new version published, we'll include it in the client. In the meantime, you can access it by using the overload on the client that allows you to drop down to protocol details. Translating the example from the guide that you linked would look something like:

var client = new OpenAIResponseClient("gpt-5", "<< YOUR API KEY >>");

// Form the OpenAI payload directly so that we can access the missing tool.   
var rawResponse = await client.CreateResponseAsync(
    BinaryContent.Create(BinaryData.FromString("""
{
  "model": "gpt-5",
  "input": "Use the code_exec tool to print hello world to the console.",
  "tools" : [
  {
    "type": "custom",
    "name": "code_exec",
    "description": "Executes arbitrary Python code.",
  }]          
}
""")));

// The call returns a raw JSON response.  Translate to the associated model so that it is easier to work with.
var response = ModelReaderWriter.Read<OpenAIResponse>(rawResponse.GetRawResponse().Content);
Console.WriteLine($"The model answered:'{response.GetOutputText()}'");

[bharathm03]

Thanks. Is there a way to use it with Microsoft.Extension.AI?

[stephentoub]

Thanks. Is there a way to use it with Microsoft.Extension.AI?

If you know that your IChatClient is backed by Responses, you can use any ResponseTool as an AITool via the AsAITool() extension method, so once the OpenAI library exposes a ResponseTool for CFG tools, it'll work. In the meantime, irrespective of M.E.AI, you can use a helper like:

static ResponseTool CreateResponseToolFromJson(JsonElement toolJson)
{
    ResponseCreationOptions options = new();
    options = ((IJsonModel<ResponseCreationOptions>)options).Create(BinaryData.FromString($$"""{ "tools" : [{{toolJson}}] }"""), ModelReaderWriterOptions.Json)!;
    return options.Tools[0];
}

to deserialize a ResponseTool from JSON representing one, e.g.

ResponseTool customTool = CreateResponseToolFromJson(JsonElement.Parse("""
    {
        "type": "custom",
        "name": "timestamp_grammar",
        "description": "Saves a timestamp in date + time in 24-hr format.",
        "format": {
            "type": "grammar",
            "syntax": "regex",
            "definition": "^\\d{4}-(0[1-9]|1[0-2])-(0[1-9]|[12]\\d|3[01]) (?:[01]\\d|2[0-3]):[0-5]\\d$"
        }
    }
    """));

Note, though, that with custom/CFG tools, just specifying the tool is only half the battle. As with function calls, the model is going to send back a response item that represents the call request, and until that's properly exposed via the OpenAI APIs, you're going to be stuck manually handling such responses.

Please feel free to ask follow-up questions / open issues over in the dotnet/extensions repo.

[bharathm03]

thanks. I'll try use this approach.