Claude Code SDK for C#

A .NET SDK for interacting with Claude through the Claude Code CLI, providing both one-shot queries and interactive client sessions with full Microsoft Agent Framework (MAF) integration.

Example: d-code.

Status & Version

Note

The Microsoft Agent Framework is still in preview status, so the APIs for ClaudeCodeSdk.MAF may change between versions.

Features

• ✅ One-shot Queries - Simple request/response pattern via ClaudeQuery.QueryAsync
• ✅ Interactive Client - Bidirectional communication with ClaudeSdkClient
• ✅ Streaming Support - Real-time response streaming via IAsyncEnumerable<T>
• ✅ Microsoft Agent Framework Integration - Use Claude as an AIAgent (ClaudeCodeSdk.MAF)
• ✅ Session Management - Multi-turn conversations with thread support and automatic session lifecycle management
• ✅ Thread Persistence - Serialize/deserialize conversation threads for storage
• ✅ Tool Integration - Full support for Claude Code tools and MCP servers
• ✅ Thinking Blocks - Extended reasoning with configurable thinking tokens
• ✅ Usage Tracking - Token usage and cost monitoring
• ✅ .NET 10.0 Ready - Built on the latest .NET platform (compatible with .NET 8.0+)
• ✅ Modern Async/Await - Full IAsyncDisposable support with proper resource management

Installation

Core SDK

Install via NuGet:

dotnet add package ClaudeCodeSdk

Microsoft Agent Framework Integration (Optional)

For Microsoft Agent Framework support:

dotnet add package ClaudeCodeSdk.MAF

Prerequisites

• .NET 10.0 SDK
• Claude Code CLI: Install globally via npm:

  npm install -g @anthropic-ai/claude-code

• API Key: Set ANTHROPIC_AUTH_TOKEN environment variable with your Anthropic API key

Quick Start

One-shot Query

using ClaudeCodeSdk;

var messages = ClaudeQuery.QueryAsync("What is the capital of France?");
await foreach (var message in messages)
{
    Console.WriteLine(message.Content);
}

Interactive Client

using ClaudeCodeSdk;

using var client = new ClaudeSdkClient();
await client.ConnectAsync();

await client.SendMessageAsync("Hello Claude!");

await foreach (var message in client.ReceiveMessagesAsync())
{
    if (message is AssistantMessage assistantMsg)
    {
        Console.WriteLine($"Claude: {assistantMsg.Content}");
    }
}

Microsoft Agent Framework Integration

using ClaudeCodeSdk.MAF;
using Microsoft.Extensions.AI;

// Create Claude as an AIAgent
await using var agent = new ClaudeCodeAIAgent();

// Simple query
var response = await agent.RunAsync("Explain async/await in C#");
Console.WriteLine(response.Text);

// Multi-turn conversation with thread
var thread = agent.GetNewThread();
var response1 = await agent.RunAsync(
    [new ChatMessage(ChatRole.User, "What is dependency injection?")],
    thread: thread
);

// Context is automatically preserved across turns
var response2 = await agent.RunAsync(
    [new ChatMessage(ChatRole.User, "Show me an example in C#")],
    thread: thread
);

// Streaming with real-time updates
await foreach (var update in agent.RunStreamingAsync("Tell me a story", thread: thread))
{
    if (update.Contents != null)
    {
        foreach (var content in update.Contents)
        {
            if (content is TextContent text)
            {
                Console.Write(text.Text);
            }
        }
    }
}

See ClaudeCodeSdk.MAF README for complete MAF integration documentation.

Architecture

The SDK implements a simplified dual-pattern architecture for Claude Code interactions:

Core Components

ClaudeProcess - Unified subprocess manager

• Direct subprocess communication with Claude Code CLI
• JSON-based message protocol with strongly-typed parsing
• Automatic CLI discovery and process lifecycle management
• Shared by both ClaudeQuery and ClaudeSdkClient

ClaudeQuery - One-shot query API

• Fire-and-forget pattern for simple queries
• Streams responses as IAsyncEnumerable<IMessage>
• Automatically handles connection lifecycle
• Ideal for single-request scenarios

ClaudeSdkClient - Interactive client API

• Long-lived bidirectional communication
• Manual connection control via ConnectAsync/DisconnectAsync
• Session management with multi-turn conversations
• Interrupt support and resource cleanup

Message Types

All messages implement IMessage:

• AssistantMessage - Claude's responses with content blocks
• UserMessage - User input
• SystemMessage - System notifications and metadata
• ResultMessage - End-of-conversation marker with cost/usage data

Content Blocks

All content blocks implement IContentBlock:

• TextBlock - Plain text content
• ThinkingBlock - Claude's reasoning (when extended thinking is enabled)
• ToolUseBlock - Tool invocations
• ToolResultBlock - Tool execution results
• ErrorContentBlock - Error information

Exception Hierarchy

Custom exceptions inherit from ClaudeSDKException:

• CLINotFoundException - Claude Code CLI not found
• CLIConnectionException - Transport connection issues
• ProcessException - Subprocess execution failures
• CLIJsonDecodeException - Message parsing errors
• MessageParseException - Type conversion failures

Microsoft Agent Framework Integration

The MAF integration (ClaudeCodeSdk.MAF) provides:

ClaudeCodeAIAgent

• Full AIAgent implementation from Microsoft.Agents.AI
• Streaming and non-streaming execution modes
• Thread-based conversation management
• Automatic session persistence via ClaudeSdkClientManager
• System prompt extraction and configuration

ClaudeCodeAgentThread

• Thread serialization/deserialization for persistence
• Session ID management for conversation continuity
• Compatible with MAF's AIConversationState

ClaudeSdkClientManager

• Automatic client lifecycle management
• Disposes old clients when switching sessions
• Thread-safe with proper async resource management
• Optimizes resource usage across multiple threads

Configuration

Environment Variables

The SDK automatically configures these environment variables for the Claude Code CLI:

• ANTHROPIC_AUTH_TOKEN - API authentication (from ClaudeCodeOptions.ApiKey or environment)
• ANTHROPIC_BASE_URL - Custom API endpoint (from ClaudeCodeOptions.BaseUrl)
• CLAUDE_CODE_ENTRYPOINT - SDK identifier (always "sdk-csharp")

ClaudeCodeOptions

Key configuration options:

var options = new ClaudeCodeOptions
{
    ApiKey = "sk-ant-...",              // Anthropic API key
    BaseUrl = "https://api.anthropic.com", // Custom API endpoint
    MaxThinkingTokens = 10000,          // Extended thinking budget
    SystemPrompt = "You are a helpful assistant",
    Model = "claude-sonnet-4",
    PermissionMode = PermissionMode.Auto, // Tool approval mode
    WorkingDirectory = "/path/to/project",
    MaxTurns = 10,
    EnvironmentVariables = new Dictionary<string, string?>
    {
        { "HTTP_PROXY", "http://proxy:1080" }
    }
};

Development

Building

# Build the entire solution
dotnet build

# Build specific project
dotnet build src/ClaudeCodeSdk/ClaudeCodeSdk.csproj

Testing

# Run all tests
dotnet test

# Run with verbose output
dotnet test --verbosity normal

# Run specific test
dotnet test --filter "FullyQualifiedName~ExceptionsTests"

Running Examples

# Run all examples
dotnet run --project examples/ClaudeCodeSdk.Examples.csproj

# Run specific example class
dotnet run --project examples/ClaudeCodeSdk.Examples.csproj -- --example QuickStart

Packaging NuGet Packages

# Pack both SDK and MAF packages
dotnet pack src/ClaudeCodeSdk/ClaudeCodeSdk.csproj -c Release
dotnet pack src/ClaudeCodeSdk.MAF/ClaudeCodeSdk.MAF.csproj -c Release

# Pack with symbols
dotnet pack src/ClaudeCodeSdk/ClaudeCodeSdk.csproj -c Release -p:IncludeSymbols=true

Examples

The examples/ folder contains complete working examples:

• QuickStartExamples - Basic one-shot queries and interactive client usage
• StreamingExamples - Real-time response streaming patterns
• MafExample - Microsoft Agent Framework integration examples

Running Examples

Make sure you have:

1. Installed Claude Code CLI: npm install -g @anthropic-ai/claude-code
2. Set ANTHROPIC_AUTH_TOKEN environment variable

Then run:

dotnet run --project examples/ClaudeCodeSdk.Examples.csproj

Key Behaviors

Message Streaming and Termination

• ClaudeProcess.ReceiveAsync() automatically terminates when receiving a ResultMessage
• Both ClaudeQuery and ClaudeSdkClient rely on this automatic termination
• ClaudeSdkClient.ReceiveResponseAsync() provides convenience method that yields until ResultMessage

JSON Serialization

• Uses snake_case_lower naming policy via JsonUtil for Claude Code CLI compatibility
• Consistent serialization across all message exchanges

Resource Management

• All process-managing classes implement IAsyncDisposable
• ClaudeProcess handles subprocess lifecycle (start, kill, cleanup)
• Use await using for automatic cleanup

MAF Session Management

• ClaudeSdkClientManager automatically handles client creation/disposal when switching threads
• Thread session IDs map to Claude Code's Resume parameter for conversation continuity
• Session state persists via the thread's SessionId

Troubleshooting

"Claude Code CLI not found"

Ensure Claude Code CLI is installed globally:

npm install -g @anthropic-ai/claude-code

Authentication Errors

Set your API key:

# macOS/Linux
export ANTHROPIC_AUTH_TOKEN="your-api-key"

# Windows PowerShell
$env:ANTHROPIC_AUTH_TOKEN="your-api-key"

# Windows Command Prompt
set ANTHROPIC_AUTH_TOKEN=your-api-key

Or pass it via options:

var options = new ClaudeCodeOptions { ApiKey = "your-api-key" };

Process Lifecycle Issues

Always dispose of SDK objects properly:

await using var client = new ClaudeSdkClient();
await client.ConnectAsync();
// ... use client
// Automatic disposal on scope exit

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

MIT License - see LICENSE.txt for details

Links

• Claude Code Documentation
• Claude API Documentation
• Microsoft Agent Framework
• GitHub Repository