Build .NET AI agents by swapping model names, not architectures

What it does

LLM Tornado is a .NET SDK for constructing AI agents and workflows. It wraps roughly thirty commercial and local LLM providers—along with vector stores like PgVector and Pinecone—in a single strongly-typed layer that has no external dependency on the vendors’ own SDKs. You define an agent graph once using an Orchestrator, Runner, and Advancer model, then change the model string to jump between backends.

The interesting bit

Instead of forcing you to hand-author JSON schemas for tool calling, the library lets you register plain C# delegates as agent tools. It also speaks emerging inter-agent protocols—MCP and A2A—out of the box, which suggests the author is trying to turn the library into connective tissue rather than just a chat wrapper.

Key highlights

• Built-in connectors for 30+ providers (OpenAI, Anthropic, Google, DeepSeek, Groq, etc.) and local engines (Ollama, vLLM, LocalAI), all strongly typed and maintained independently of first-party SDKs.
• Agent orchestration with handoffs, parallel execution, and Mermaid graph export via Orchestrator/Runner/Advancer primitives.
• Multimodal support across text, images, video, documents, URLs, and audio inputs/outputs.
• Native MCP and A2A protocol support, plus vector database connectors for Chroma, PgVector, Pinecone, Faiss, and QDrant.
• Interoperability with Microsoft.Extensions.AI and Semantic Kernel for shops already in that ecosystem.

Verdict

Worth a look if you are building non-trivial agent workflows in .NET and want to avoid provider lock-in. If you only need a thin HTTP client for a single LLM API, it is likely more surface area than you need.