OpenAI Provider

Connect to the OpenAI API (api.openai.com) for chat completions, embeddings, image generation, and speech.

Quick Start

builder.Services
    .AddCoreAIServices()
    .AddCoreAIOrchestration()
    .AddCoreAIOpenAI();

Services Registered

Service	Implementation	Lifetime
IAIClientProvider	OpenAIClientProvider	Scoped
IAICompletionClient	OpenAICompletionClient	Scoped
Connection source	—	Scoped

Configuration

Connection Setup

Provide an API key through your connection source:

builder.Services.AddCoreAIConnectionSource("OpenAI", options =>
{
    options.Connections.Add(new AIProviderConnectionEntry
    {
        Name = "my-openai",
        ClientName = "OpenAI",
        // Set API key and optional endpoint
    });
});

Constants

Constant	Value
OpenAIConstants.ClientName	"OpenAI"

Capabilities

Capability	Supported
Chat completions	✅
Streaming	✅
Embeddings	✅
Image generation	✅
Speech-to-text	✅
Text-to-speech	✅

Configuration Example

A full appsettings.json configuration for OpenAI:

{
  "CrestApps": {
    "AI": {
        "Connections": [
          {
            "Name": "openai-primary",
            "ClientName": "OpenAI",
            "ApiKey": "sk-..."
          }
        ]
      }
    }
}

Or register connections programmatically:

builder.Services.AddCoreAIConnectionSource("OpenAI", options =>
{
    options.Connections.Add(new AIProviderConnectionEntry
    {
        Name = "my-openai",
        ClientName = "OpenAI",
        // API key is read from configuration or set directly
    });
});

OpenAI-compatible endpoints

The OpenAI client is often useful beyond api.openai.com. It can also connect to providers that expose OpenAI-compatible chat and embedding endpoints.

Common examples include:

Provider	Notes
OpenAI	Native target for this client
Azure OpenAI	OpenAI-compatible request shape with Azure-specific endpoint and deployment conventions
Anthropic	Provides OpenAI-compatible endpoints in addition to its native Claude API
Google Gemini	Offers an OpenAI-compatibility layer for supported Gemini models
Groq	OpenAI-compatible chat/completions API
Mistral	OpenAI-compatible API surface for supported models
xAI	OpenAI-compatible API for Grok models
Together AI	OpenAI-compatible endpoints for hosted open-weight models
Fireworks AI	OpenAI-compatible endpoints for hosted open-weight models
OpenRouter	OpenAI-compatible multi-provider routing layer
DeepSeek	OpenAI-compatible chat and embedding-style endpoints
Perplexity	OpenAI-compatible chat/completions endpoint
Ollama / vLLM / LocalAI	Common self-hosted OpenAI-compatible endpoints

Some providers expose both OpenAI-compatible and native APIs. Anthropic is one example: you can use its OpenAI-compatible endpoint through this client, or use the dedicated Claude Orchestrator when you want the native Claude-oriented integration path.

tip

Never commit API keys to source control. Use environment variables, user secrets, or a vault provider for the values you map into your CrestApps:AI:Connections configuration.

Available Models

Model	Type	Context Window	Best For
gpt-4.1	Chat	1M tokens	Complex reasoning, coding, instruction following
gpt-4.1-mini	Chat	1M tokens	Balanced performance and cost
gpt-4.1-nano	Chat	1M tokens	Fast, cost-effective for simple tasks
o4-mini	Reasoning	200K tokens	STEM, math, coding with chain-of-thought
gpt-4o	Chat	128K tokens	Multimodal (text + vision), general purpose
gpt-4o-mini	Chat	128K tokens	Budget-friendly multimodal
text-embedding-3-small	Embedding	8K tokens	Cost-effective embeddings
text-embedding-3-large	Embedding	8K tokens	Higher-quality embeddings
dall-e-3	Image	—	Image generation
whisper-1	Speech-to-text	—	Audio transcription
tts-1 / tts-1-hd	Text-to-speech	—	Voice synthesis

info

Model availability and capabilities change frequently. Check the OpenAI models documentation for the latest information.

Streaming

The OpenAI provider fully supports streaming responses. When streaming is enabled, tokens are sent to the client as they are generated rather than waiting for the complete response:

// Streaming is handled automatically by the orchestrator when the
// chat interaction is configured for streaming (the default for real-time chat).
// No additional configuration is needed.

Streaming is the default behavior for the chat interactions module. The OpenAICompletionClient uses the IChatClient.GetStreamingResponseAsync() method from the Microsoft.Extensions.AI abstraction.

Function Calling

OpenAI models support function calling (tool use), which is the foundation for the Custom AI Tools system. When tools are registered and assigned to a profile, the OpenAI provider automatically:

1. Serializes tool definitions as JSON Schema in the request
2. Parses tool call responses from the model
3. Invokes the matching AITool via the orchestrator
4. Sends tool results back to the model for the final response

All GPT-4 and newer models support parallel function calling (multiple tools invoked in a single turn).