Semantic Kernel

Microsoft's Semantic Kernel gives .NET, Python, and Java developers one programming model for calling language models, exposing their own code as tools, and assembling agents. You register AI services and plugins on a kernel object; plugins can be native functions, prompt templates, OpenAPI descriptions, or Model Context Protocol servers, and the model invokes them through function calling. Layered on top are agent classes, multi-agent orchestration, a process framework for long-running business workflows, and vector store connectors for Azure AI Search, Elasticsearch, Chroma, and others. Model connectors include OpenAI, Azure OpenAI, Hugging Face, and NVIDIA NIM, and local models work through Ollama, LM Studio, or ONNX. It needs Python 3.10 or later, .NET 10, or JDK 17, with packages on PyPI and NuGet. The MIT license permits commercial use. Status matters: Microsoft names Agent Framework as the successor and is putting most new features there, while Semantic Kernel 1.x continues to receive bug fixes, security patches, and regular releases. Best for teams with existing Semantic Kernel code or close Azure ties.