AI SDK for .NET

Status: Phases 1-8 Complete ✅ | Production Ready | .NET 10.0+

A comprehensive .NET SDK for integrating with AI providers, offering a unified interface for language models, embeddings, structured output, and streaming. Single package, 33 providers included.

Current Status

Test Coverage: 182 tests passing (122 Core + 60 ASP.NET Core)

Features

• 📦 Single Package - Everything in one NuGet package, no dependency hell
• 🚀 Idiomatic .NET - Async/await, IAsyncEnumerable, records, dependency injection
• 🔌 33 Providers Included - OpenAI, Anthropic, Azure, Google, Groq, Mistral, Cohere, GoogleVertex, AmazonBedrock, Replicate, Perplexity, DeepSeek, OpenAI-Compatible, Vercel, xAI, HuggingFace, Cerebras, Fireworks, TogetherAI, AI21, Cloudflare, Baseten, Lepton, Novita, Writer, LlamaFile, Friendli, Portkey, Fal, Luma, Stability, OpenRouter, Z.AI
• 🌐 ASP.NET Core Ready - DI integration, health checks, SSE streaming middleware
• 🎯 Structured Output - Generate typed objects with JSON schema validation
• ⚡ Real-time Streaming - Server-sent events for chat and completions
• 🧪 Fully Tested - Comprehensive test suite with 100% pass rate
• 📖 Production Ready - Timeout enforcement, unified error handling, real health checks

Installation

Simplified Installation - One Package for Everything

# Single package includes all providers
dotnet add package AiSdk

# Optional: ASP.NET Core integration (if building web apps)
dotnet add package AiSdk.AspNetCore

That's it! All 33 providers are included in the single AiSdk package.

Quick Start

Console Application

using AiSdk.Abstractions;
using AiSdk.Providers.OpenAI;

// Create a language model
var provider = new OpenAIProvider(apiKey: Environment.GetEnvironmentVariable("OPENAI_API_KEY")!);
var model = provider.ChatModel("gpt-4");

// Generate text
var options = new LanguageModelCallOptions
{
    Messages = new List<Message>
    {
        new Message(MessageRole.User, "Tell me a joke about programming")
    }
};

var result = await model.GenerateAsync(options);
Console.WriteLine(result.Text);

// Stream text in real-time
await foreach (var chunk in model.StreamAsync(options))
{
    if (chunk.Type == ChunkType.TextDelta)
    {
        Console.Write(chunk.Delta);
    }
}

ASP.NET Core Application

using AiSdk.AspNetCore;
using AiSdk.Providers.Anthropic;

var builder = WebApplication.CreateBuilder(args);

// Register AI SDK with dependency injection
builder.Services.AddAiSdk(options =>
{
    options.DefaultProvider = "anthropic";
    options.Providers.Add("anthropic", new ProviderConfiguration
    {
        ApiKey = Environment.GetEnvironmentVariable("ANTHROPIC_API_KEY")!,
        DefaultModel = "claude-3-5-sonnet-20241022"
    });
});

var app = builder.Build();

// Enable SSE streaming middleware
app.UseAiSdkStreaming();

// Create chat endpoint with streaming
app.MapPost("/api/chat", async (ChatRequest request, ILanguageModel model) =>
{
    var options = new LanguageModelCallOptions
    {
        Messages = new List<Message>
        {
            new Message(MessageRole.User, request.Message)
        }
    };

    return Results.Stream(async stream =>
    {
        await foreach (var chunk in model.StreamAsync(options))
        {
            if (chunk.Type == ChunkType.TextDelta)
            {
                await stream.WriteAsync($"data: {chunk.Delta}\n\n");
                await stream.FlushAsync();
            }
        }
    }, "text/event-stream");
});

app.Run();

record ChatRequest(string Message);

Included Providers

All providers are included in the single AiSdk package - no need to install separate provider packages!

Language Model Providers (33 Complete / 19+ Planned)

Priority 1 - Major Cloud Providers (All Complete ✅)

Priority 2 - Fast Inference & Aggregators (All Complete ✅)

Priority 3 - Specialized & Ecosystem (All Complete ✅)

Priority 4 - Specialized Infrastructure (All Complete ✅)

Priority 5 - Gateways & Local Execution (All Complete ✅)

Planned - Additional Providers

Other Provider Types (Future)

• 📋 Audio Providers: Deepgram, ElevenLabs, AssemblyAI, Azure Speech
• 📋 Embedding Providers: OpenAI Embeddings, Cohere Embed, Voyage AI
• 📋 Image Generation: Stability AI, DALL-E, Midjourney API
• 📋 Video Generation: Runway, Luma AI

Project Structure

Monorepo Design - Single package, organized codebase:

ai-sdk.net/
├── src/
│   ├── AiSdk/                       # 📦 Main package (includes everything)
│   │   ├── Abstractions/            # Core interfaces, models, errors
│   │   ├── Core/                    # Utili3ies (streaming, JSON, HTTP)
│   │   ├── Providers/               # All 32 providers in one place
│   │   │   ├── OpenAI/              # OpenAI (GPT-4, GPT-3.5)
│   │   │   ├── Anthropic/           # Anthropic (Claude)
│   │   │   ├── Azure/               # Azure OpenAI
│   │   │   ├── Google/              # Google Gemini
│   │   │   ├── Groq/                # Groq (Llama 3.1, Mixtral)
│   │   │   ├── Mistral/             # Mistral AI
│   │   │   ├── Cohere/              # Cohere
│   │   │   ├── GoogleVertex/        # Google Vertex AI
│   │   │   ├── AmazonBedrock/       # Amazon Bedrock
│   │   │   ├── Replicate/           # Replicate (Llama 2, Mixtral)
│   │   │   ├── Perplexity/          # Perplexity (Sonar online search)
│   │   │   ├── DeepSeek/            # DeepSeek (Chat, Coder, Reasoner)
│   │   │   ├── ZAI/                 # Z.AI (GLM-4, CodeGeeX)
│   │   │   ├── OpenAICompatible/    # Universal OpenAI-compatible
│   │   │   ├── Vercel/              # Vercel AI Gateway
│   │   │   ├── XAI/                 # xAI (Grok)
│   │   │   ├── HuggingFace/         # HuggingFace Inference API
│   │   │   ├── Cerebras/            # Cerebras (ultra-fast)
│   │   │   ├── Fireworks/           # Fireworks AI
│   │   │   ├── TogetherAI/          # Together AI (200+ models)
│   │   │   ├── AI21/                # AI21 Labs (Jamba, Jurassic-2)
│   │   │   ├── Cloudflare/          # Cloudflare Workers AI
│   │   │   ├── Baseten/             # Baseten (Llama, Mistral, Mixtral)
│   │   │   ├── Lepton/              # Lepton AI (fast inference)
│   │   │   ├── Novita/              # Novita AI (Llama, Mistral, Qwen)
│   │   │   ├── Writer/              # Writer (Palmyra models)
│   │   │   ├── LlamaFile/           # LlamaFile (local execution)
│   │   │   ├── Friendli/            # Friendli AI (ultra-fast)
│   │   │   ├── Portkey/             # Portkey AI Gateway
│   │   │   ├── Fal/                 # Fal AI (multi-model)
│   │   │   ├── Luma/                # Luma AI (Dream Machine)
│   │   │   ├── Stability/           # Stability AI (StableLM)
│   │   │   └── OpenRouter/          # OpenRouter (100+ models)
│   │   ├── AiClient.cs              # High-level API
│   │   └── Models/                  # Shared models
│   └── AiSdk.AspNetCore/            # 📦 Optional web integration package
│       ├── Configuration/
│       ├── Extensions/
│       ├── HealthChecks/
│       └── Middleware/
├── tests/
│   ├── AiSdk.Abstractions.Tests/    # 4 tests
│   ├── AiSdk.Core.Tests/            # 118 tests
│   └── AiSdk.AspNetCore.Tests/      # 60 tests
└── examples/
    ├── GettingStartedExample/       # Basic usage patterns
    ├── StreamingExample/            # Real-time streaming
    ├── ToolCallingExample/          # Function/tool calling
    ├── StructuredOutputExample/     # Generate typed objects
    ├── MinimalApiExample/           # ASP.NET Core Minimal API
    ├── MvcExample/                  # ASP.NET Core MVC
    └── BlazorServerExample/         # Blazor Server with SignalR

Benefits of Monorepo Structure:

• 📦 Single package to install
• 🔄 Easier version management
• 🚀 All providers always compatible
• 📝 Simpler maintenance

Example Applications

All examples are fully functional and can be run directly:

Console Examples

# Basic usage patterns
cd examples/GettingStartedExample && dotnet run

# Real-time streaming
cd examples/StreamingExample && dotnet run

# Function calling with tools
cd examples/ToolCallingExample && dotnet run

# Structured output (typed objects)
cd examples/StructuredOutputExample && dotnet run

Web Examples

# Minimal API (modern REST API)
cd examples/MinimalApiExample && dotnet run
# Visit: http://localhost:5000/swagger

# MVC application (traditional web app)
cd examples/MvcExample && dotnet run
# Visit: http://localhost:5001

# Blazor Server (interactive SPA)
cd examples/BlazorServerExample && dotnet run
# Visit: http://localhost:5002

Using Different Providers

All providers are included - just switch the namespace:

// OpenAI
using AiSdk.Providers.OpenAI;
var provider = new OpenAIProvider(new OpenAIConfiguration { ApiKey = "..." });
var model = provider.GPT4();

// Anthropic (Claude)
using AiSdk.Providers.Anthropic;
var provider = new AnthropicProvider(new AnthropicConfiguration { ApiKey = "..." });
var model = provider.Claude35Sonnet();

// Google Gemini
using AiSdk.Providers.Google;
var provider = new GoogleProvider(new GoogleConfiguration { ApiKey = "..." });
var model = provider.Gemini15Pro();

// Groq (ultra-fast inference)
using AiSdk.Providers.Groq;
var provider = new GroqProvider(new GroqConfiguration { ApiKey = "..." });
var model = provider.Llama3_1_70B();

// Mistral
using AiSdk.Providers.Mistral;
var provider = new MistralProvider(new MistralConfiguration { ApiKey = "..." });
var model = provider.MistralLarge();

// Cohere
using AiSdk.Providers.Cohere;
var provider = new CohereProvider(new CohereConfiguration { ApiKey = "..." });
var model = provider.CommandRPlus();

// Azure OpenAI
using AiSdk.Providers.Azure;
var provider = new AzureOpenAIProvider(new AzureOpenAIConfiguration { ... });
var model = provider.ChatModel("deployment-name");

// Google Vertex AI (on GCP)
using AiSdk.Providers.GoogleVertex;
var provider = new GoogleVertexProvider(new GoogleVertexConfiguration { ... });
var model = provider.Gemini15Pro();

// Amazon Bedrock (multi-provider)
using AiSdk.Providers.AmazonBedrock;
var provider = new AmazonBedrockProvider(new AmazonBedrockConfiguration { ... });
var model = provider.Claude35Sonnet();  // or Llama3, TitanText, etc.

// Replicate
using AiSdk.Providers.Replicate;
var provider = new ReplicateProvider(new ReplicateConfiguration { ApiKey = "..." });
var model = provider.Llama2_70B();

// Perplexity (online search)
using AiSdk.Providers.Perplexity;
var provider = new PerplexityProvider(new PerplexityConfiguration { ApiKey = "..." });
var model = provider.SonarLargeOnline();

// DeepSeek (code & reasoning)
using AiSdk.Providers.DeepSeek;
var provider = new DeepSeekProvider(new DeepSeekConfiguration { ApiKey = "..." });
var model = provider.Coder();  // or Chat(), Reasoner()

// OpenAI-Compatible (Ollama, LocalAI, vLLM, LM Studio)
using AiSdk.Providers.OpenAICompatible;
var model = OpenAICompatibleProvider.ForOllama("llama2");
// or ForLocalAI(), ForVLLM(), ForLMStudio(), or custom endpoint

All providers implement the same ILanguageModel interface, so you can easily switch between them.

Advanced Features

Structured Output

Generate strongly-typed objects with automatic JSON schema validation:

public record Person(string Name, int Age, string City);

var options = new StreamObjectOptions<Person>
{
    Schema = JsonSchema.FromType<Person>(),
    Prompt = "Generate a person: John, 30 years old, lives in Seattle"
};

var person = await StreamObjectAsync(model, options);
Console.WriteLine($"{person.Name} is {person.Age} years old");

Tool/Function Calling

Define tools that the model can call:

var tools = new List<Tool>
{
    new Tool
    {
        Name = "get_weather",
        Description = "Get the current weather in a location",
        Parameters = JsonSchema.FromType<WeatherRequest>()
    }
};

var options = new LanguageModelCallOptions
{
    Messages = messages,
    Tools = tools
};

var result = await model.GenerateAsync(options);
if (result.ToolCalls?.Count > 0)
{
    foreach (var toolCall in result.ToolCalls)
    {
        // Handle tool execution
    }
}

Health Checks (ASP.NET Core)

Monitor AI service availability:

builder.Services.AddAiSdk(options =>
{
    options.EnableHealthChecks = true;
});

app.MapHealthChecks("/health");

Configuration (ASP.NET Core)

Configure via appsettings.json:

{
  "AiSdk": {
    "DefaultProvider": "anthropic",
    "TimeoutSeconds": 120,
    "EnableHealthChecks": true,
    "EnableTelemetry": true,
    "Providers": {
      "anthropic": {
        "ApiKey": "sk-ant-...",
        "DefaultModel": "claude-3-5-sonnet-20241022",
        "TimeoutSeconds": 90
      },
      "openai": {
        "ApiKey": "sk-...",
        "DefaultModel": "gpt-4",
        "BaseUrl": "https://api.openai.com/v1"
      }
    }
  }
}

Building from Source

# Clone repository
git clone https://github.com/ai-sdk-dotnet/ai-sdk.net.git
cd ai-sdk.net

# Build all projects
dotnet build

# Run tests (182 tests)
dotnet test

# Run specific example
cd examples/MinimalApiExample
dotnet run

Requirements

• .NET 10.0 SDK or later
• API keys for providers you want to use

Roadmap

Phase 5 (Complete):

• ✅ Groq provider (Llama 3.1, Mixtral, Gemma)
• ✅ Mistral provider (Mistral Large, Medium, Mixtral)
• ✅ Cohere provider (Command R+, Command R)
• ✅ Google Vertex AI provider
• ✅ Amazon Bedrock provider (multi-model aggregator)
• ✅ Replicate provider (Llama 2, Mixtral, Mistral)
• ✅ Perplexity provider (Sonar online search models)
• ✅ DeepSeek provider (Chat, Coder, Reasoner R1)
• ✅ OpenAI-Compatible provider (universal connector)

Next Up (Phase 6):

• Audio providers (Deepgram, ElevenLabs, AssemblyAI)
• Embedding support
• Image generation support
• Enhanced telemetry with OpenTelemetry

Package Information

Why Single Package?

• ✅ Simpler dependency management
• ✅ All providers always version-compatible
• ✅ Easier to maintain and update
• ✅ No "which package do I need?" confusion
• ✅ Better for monorepo/enterprise scenarios

License

This project is licensed under the MIT License - see the LICENSE.md file for details.