Ananke — AI Agent Orchestration Framework for .NET

Ananke is a vendor-agnostic, production-ready .NET framework for building AI agents and automated multi-step pipelines. It provides typed workflow orchestration, LLM tool calling, multi-provider AI model support, long-term memory (RAG + empirical learning), agentic design patterns, human-in-the-loop approval, distributed coordination, an external skill catalog, a design-time CLI (nnke), and OpenTelemetry observability — all with idiomatic C# and no external services required for testing.

Supports OpenAI (GPT-4.1, o-series), Anthropic Claude, Google Gemini, and any OpenAI-compatible endpoint including Ollama, Azure OpenAI, LM Studio, Groq, Deepseek, and Together AI.

Beyond orchestration, Ananke makes agents smarter over time: they accumulate knowledge from documents, recognize patterns across interactions, and build reusable skills and heuristics — turning raw LLM capability into compounding operational intelligence.

From a streaming chat agent to state-machine-coordinated multi-service pipelines: dotnet add package Ananke and start building.

Why Ananke?

The AI agent ecosystem is predominantly Python. For .NET teams shipping to production, that means either adopting a foreign ecosystem or assembling production infrastructure from scratch. Even within mature frameworks, capabilities that .NET developers take for granted — strong typing, real parallelism, dependency injection, structured concurrency — often require additional libraries, schema definitions, or workarounds.

The landscape is also moving fast. Frameworks rebrand, merge, or shift direction between releases, making it risky to couple production systems to a single vendor's roadmap.

Ananke starts from a different question:

What does the infrastructure need to look like so that building any agentic system — at any scale — is straightforward for .NET developers?

The answer is a typed, testable, composable foundation where the infrastructure comes first and LLM providers are pluggable.

Capabilities

🔀 Workflow Orchestration

Fluent graph-as-code builder · conditional & LLM-driven routing · fork/join parallelism · nested sub-workflows · human-in-the-loop interrupts · typed IAsyncEnumerable event streaming

🤖 AI Agents

AgentJob with tool calling + structured output · token-level streaming · multimodal messages (text, image, audio) · ChatSessionEvent async stream · multi-provider (OpenAI, Anthropic, Google Gemini + any OpenAI-compatible endpoint) · capability-based model routing · production decorators (429 retry with OTel, LLM response caching)

🔄 Agentic Patterns

AgenticPattern.ReviewCritique<T>() — generator → critic → loop until approved · AgenticPattern.IterativeRefinement<T>() — single-agent refinement loop · pre-wired builders for recognized design patterns on top of Workflow<T> primitives

🧠 Long-Term Memory

Document ingestion pipeline (extract → chunk → embed → store) · vector-indexed semantic search · knowledge catalog with LLM-enriched metadata and time-decay reranking · empirical memory (patterns, skills, heuristics) · episode store with temporal trajectories · Monte Carlo reward propagation · tag importance tracking · skill package export/import — portable bundles of learned knowledge with quality gates and trust scaling

🛑 Human-in-the-Loop

Pause execution at any step · checkpoint full workflow state · resume with optional human input · interrupt stack for state machines

🏭 State Machine

Simplified IStateMachine<S,T> with interrupt stack for in-process scenarios · production AbstractStateMachine with RedLock coordination · composable middleware pipeline · guard conditions · circuit breaking (fault/reset)

🌐 ASP.NET Core

SSE streaming · state-machine-driven chat sessions · in-memory session management · provider configuration helpers

🔌 MCP Server

Expose any workflow or tool kit as an MCP server capability

🎨 Design Tooling

Plain-text DSL for workflow topology · runtime binding · Mermaid diagram export

🛠️ nnke Design CLI

Scaffold new workflow projects · validate .ananke.yml topology files · export Mermaid diagrams · inspect project health · browse docs from the terminal · nnke mcp exposes all capabilities as MCP tools for AI coding assistants (GitHub Copilot, Claude, etc.)

🗄️ Infrastructure

Checkpointing (InMemory / File) · distributed locking (Redis) · MQTT pub/sub · OpenTelemetry tracing

🎓 External Skill Catalog

Discover CLI tools from the OpenClaw/ClawHub registry · local JSON cache for offline search · ToolKit.AddFromCatalogAsync() for natural-language skill discovery · local skill voting and reliability scoring · runs Python (PyPI) tools via uvx, Node.js tools via npx, Docker containers, or any shell command — no Python installation required

🧑‍💻 Developer Experience

Idiomatic C# (async/await, DI, generics) · full in-memory test mode for every infrastructure contract · 18 focused NuGet packages

Quick Example

Build a typed multi-step pipeline in minutes:

using Ananke.Orchestration;

record ContentState
{
    public string Topic { get; init; } = "";
    public string Draft  { get; init; } = "";
    public string Final  { get; init; } = "";
}

var workflow = new Workflow<ContentState>("content-pipeline")
    .Job("draft",  async (state, ct) => state with { Draft = $"Draft on: {state.Topic}" })
    .Job("polish", async (state, ct) => state with { Final = state.Draft.ToUpperInvariant() })
    .Chain("draft", "polish")
    .Then("polish", Workflow.End);

var result = await workflow.RunAsync(new ContentState { Topic = "AI in .NET" });
Console.WriteLine(result.State.Final);  // "DRAFT ON: AI IN .NET"
Console.WriteLine(result.Status);        // Completed

Drop an LLM agent into any job with the same API — add Ananke.Orchestration.OpenAI and replace the delegate with an AgentJob. See Getting Started for a full walkthrough.

Getting Started

Install the meta-package to get everything:

dotnet add package Ananke

Or install only what you need:

dotnet add package Ananke.Orchestration          # core: workflows, agents, knowledge pipeline
dotnet add package Ananke.Orchestration.OpenAI     # OpenAI chat + embeddings
dotnet add package Ananke.Documents              # PDF + Markdown extraction for knowledge ingestion
dotnet add package Ananke.OpenTelemetry            # distributed tracing

Then explore the demos to see each capability in action.

Demos

Each demo is a self-contained project you can run to see a specific capability end to end.

Packages

Package	Description	NuGet
Ananke	Meta-package — install once, get everything
Ananke.Abstractions	Shared interfaces and contracts (IDistributedLock, IChannelReader/Writer, etc.)
Ananke.StateMachine	Distributed FSM engine with middleware pipeline
Ananke.Orchestration	Workflow builder, runner, agents, checkpointing
Ananke.Orchestration.OpenAI	OpenAI provider (IStreamingAgentModel)
Ananke.Orchestration.Anthropic	Anthropic / Claude provider (IStreamingAgentModel)
Ananke.Orchestration.Google	Google Gemini provider (IStreamingAgentModel)
Ananke.MCP	Expose workflows and tools as MCP server capabilities
Ananke.A2A	Agent-to-Agent (A2A) protocol — call remote agents as IAgentModel, expose workflows as A2A endpoints
Ananke.Learning	Empirical memory, offline learning, episode store, Monte Carlo reward propagation, skill package export/import
Ananke.Skills	External skill catalog — discover and run CLI tools from the OpenClaw registry via ToolKit
Ananke.Documents	Document extractors for the knowledge pipeline (PDF, Markdown)
Ananke.Qdrant	Qdrant vector database provider for IKnowledgeStore, IKnowledgeCatalog, and IEmpiricalMemory
Ananke.Redis	Distributed lock and key-value store via Redis
Ananke.MQTT	Pub/sub channels via MQTTnet
Ananke.OpenTelemetry	One-liner OTLP tracing export
Ananke.AspNetCore	SSE streaming, provider configuration, and session management for ASP.NET Core
Ananke.Design	YAML manifest import and Mermaid diagram export

Documentation & Guides

The full documentation hub and progressive learning path are at docs/learning.md and the Feature Index.

Core Guides

Reference

Frequently Asked Questions

Common questions are answered in the FAQ. Quick answers:

• Which LLM providers are supported? OpenAI (GPT-4.1, o-series), Anthropic Claude, Google Gemini, and any OpenAI-compatible endpoint — Ollama, Azure OpenAI, LM Studio, Groq, Deepseek, Together AI, vLLM.
• Can I test without a real LLM or external services? Yes — every infrastructure contract ships with an in-memory implementation. Integration tests run in milliseconds with no API keys.
• Is Ananke only for chat agents? No — it supports batch pipelines, state-machine workflows, distributed multi-service coordination, document-ingestion pipelines, and agentic design patterns alongside interactive chat.
• Can agents learn and improve over time? Yes — IEmpiricalMemory accumulates patterns, skills, and heuristics from interactions. IOfflineLearner runs background sweeps that decay stale beliefs, explore hypotheses, and promote stable knowledge to the permanent store.
• Does Ananke support MCP (Model Context Protocol)? Yes — expose any ToolKit or Workflow as an MCP server (compatible with VS Code Copilot, Claude Desktop, and any MCP client) and consume tools from external MCP servers.
• Does it support the A2A agent protocol? Yes — call remote A2A agents as drop-in IAgentModel implementations and expose Ananke workflows as A2A endpoints.
• Can learned knowledge be transferred between agents? Yes — ISkillPackager exports empirical entries and linked episodes as a portable JSON package (quality gates: min confidence, min strength, min observations). Import into any other agent with configurable trust scaling. Tag importance weights are bundled so the receiving agent inherits feature correlations too.
• What is the OpenClaw skill catalog? Ananke.Skills connects to the OpenClaw/ClawHub registry of CLI-based tools. One call to toolkit.AddFromCatalogAsync("airbnb search lodging") discovers, caches, and resolves matching tools as ToolDefinition entries that any agent can call.
• What is nnke? A .NET CLI tool (dotnet tool install -g nnke) for design-time workflow tasks: scaffold projects, validate topology files, export Mermaid diagrams, inspect project health, and browse docs. nnke mcp runs as an MCP server so AI coding tools (GitHub Copilot, Claude) can call these capabilities directly.
• What is the difference between a Workflow and a State Machine? A workflow runs a directed pipeline end to end (best for task pipelines, document processing, batch jobs). A state machine models long-lived entities with stable states and event-driven transitions (best for conversation sessions, order lifecycle, device management). Both compose: state machines can invoke workflows, and workflows can interact with state machines.

→ Full FAQ →

Philosophy

Ananke takes its name from the Greek primordial goddess of necessity — the force that fixed the laws of the cosmos before creation could begin. Before time could flow and matter could form, something unchanging had to exist first.

Software is no different. Before agents can act, before workflows can run, the rules must be stable.

→ Read the full backstory and philosophy

License

Licensed under the Apache 2.0 License.