Squad.NET 0.1.0

Squad.NET

Squad.NET is a .NET-native framework for building typed, observable multi-agent workflows with sensible local defaults and provider escape hatches.

Current release target: 0.1.0.

What Works In v0.1

• Local offline runs with FakeChatModel
• Sequential, parallel, and hierarchical task execution with bounded manager replanning
• Per-agent model routing with a named model registry
• Provider-neutral native tool calls and prompt-style tool calls
• Structured output with Throw, ReturnRaw, and ReturnDefault policies
• OpenAI-compatible /v1/chat/completions provider via Squad.OpenAI
• OpenRouter provider via Squad.OpenRouter
• AWS Bedrock Runtime Converse provider via Squad.Bedrock
• Ollama local/network provider via Squad.Ollama
• LM Studio local/network provider via Squad.LMStudio
• ASP.NET Core hosting library plus runnable hosting sample
• Microsoft.Extensions.DependencyInjection integration

Quick Start

No API key is needed for the first run:

using Squad.Abstractions;
using Squad.Core;

var squad = SquadBuilder.Create("HelloSquad")
    .AddAgent(AgentDef.Create("Assistant", "Helpful assistant", "Answer clearly"))
    .AddTask(TaskDef.Create("Answer", "Answer the user's question"))
    .WithModel(new FakeChatModel("local", "Hello from Squad.NET."))
    .Build();

var result = await squad.RunAsync(SquadInput.FromText("Hello"));
Console.WriteLine(result.OutputText);

Run the bundled sample:

dotnet run --project samples/MinimalHelloWorld

Per-Agent Models

Register named models once, then opt agents into them only when needed. Agents without a model use the default.

var squad = SquadBuilder.Create("ContentSquad")
    .AddAgent(AgentDef.Create("Researcher", "Research", "Find facts").WithModel("research"))
    .AddAgent(AgentDef.Create("Writer", "Writer", "Write clearly").WithModel("writer"))
    .AddTask(TaskDef.Create("Research", "Research the topic").WithAgent("Researcher"))
    .AddTask(TaskDef.Create("Write", "Write the final answer").WithAgent("Writer"))
    .WithModel("research", researchModel)
    .WithModel("writer", writerModel)
    .WithDefaultModel("research")
    .Build();

Unknown model names fail fast with an actionable message that names the agent and the missing registration.

Hierarchical Replanning

Hierarchical squads use the manager agent to create a plan. By default, if a planned task fails, the manager gets one chance to revise the remaining plan while completed successful tasks are preserved.

var squad = SquadBuilder.Create("ContentSquad")
    .WithProcess(Process.Hierarchical)
    .WithManagerAgent("Researcher")
    .WithHierarchicalOptions(o => o.WithMaxPlanRevisions(1))
    .Build();

Disable replanning when you want strict fail-fast behavior:

.WithHierarchicalOptions(o => o.DisableReplanning())

Providers

Provider-specific boilerplates are available in docs/providers.

OpenAI-compatible endpoint:

using Squad.OpenAI;

var model = new OpenAIChatModel(new OpenAIChatModelOptions
{
    ModelId = "gpt-4.1-mini",
    ApiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!,
    BaseUrl = "https://api.openai.com/v1"
});

OpenRouter:

using Squad.OpenRouter;

var model = new OpenRouterChatModel(
    "openai/gpt-oss-120b:free",
    Environment.GetEnvironmentVariable("OPENROUTER_API_KEY")!);

AWS Bedrock Converse:

using Squad.Bedrock;

var model = new BedrockChatModel(new BedrockChatModelOptions
{
    ModelId = "anthropic.claude-3-haiku-20240307-v1:0",
    Region = "us-east-1"
});

Bedrock uses explicit options, environment variables, or the shared AWS credentials/config files.

Ollama on the same machine:

using Squad.Ollama;

var model = new OllamaChatModel("llama3.2");

Ollama on another device on your network:

var model = new OllamaChatModel("llama3.2", "http://192.168.1.20:11434");

LM Studio:

using Squad.LMStudio;

var model = new LMStudioChatModel("qwen2.5-coder-7b");

Hosting

Squad.Hosting is a reusable library. The runnable ASP.NET Core app lives in samples/MVP6.HostingDemo.

dotnet run --project samples/MVP6.HostingDemo

Execute:

curl -X POST http://localhost:5000/api/squad/execute \
  -H "Content-Type: application/json" \
  -d "{\"squadName\":\"DemoSquad\",\"input\":\"Hello\"}"

Both hosted endpoints are POST endpoints and use the same JSON body:

{
  "squadName": "DemoSquad",
  "input": "Hello"
}

execute returns a SquadResult JSON response. stream returns Server-Sent Events.

You can choose the route prefix:

app.MapSquadApi("/ai/squads");

Stream:

curl -N -X POST http://localhost:5000/api/squad/stream \
  -H "Content-Type: application/json" \
  -d "{\"squadName\":\"DemoSquad\",\"input\":\"Hello\"}"

Build And Test

dotnet restore Squad.NET.sln
dotnet build Squad.NET.sln -c Release --no-restore
dotnet test Squad.NET.sln -c Release --no-build
dotnet pack src/Squad.NET/Squad.NET.csproj -c Release --no-build
docker build -t squad-hosting-sample .

Project Layout

src/
  Squad.Abstractions/
  Squad.Core/
  Squad.OpenAI/
  Squad.OpenRouter/
  Squad.Bedrock/
  Squad.Ollama/
  Squad.LMStudio/
  Squad.DependencyInjection/
  Squad.Hosting/
  Squad.NET/
samples/
  MinimalHelloWorld/
  MVP1.OpenRouterDemo/
  MVP2.ToolsDemo/
  MVP3.StructuredOutputDemo/
  MVP4.ProcessesDemo/
  MVP5.ObservabilityDemo/
  MVP6.HostingDemo/
  MVP7.LocalProvidersDemo/
tests/

License

MIT. See LICENSE.