← Back to all projects

semantic-kernel

An agent SDK with unusual depth, still maintained, though new capability now lands in its successor Agent Framework.

OfficialMIT
Stars
28.7k
Forks
4.8k
Open issues
337
Last commit
30 Oct 2026

What is semantic-kernel?

Semantic Kernel puts a kernel at the centre of the application: a container you register model connections and plugins into, which the agents you build then draw on. Mark ordinary methods as kernel functions and the model can call them, import a whole API from an OpenAPI description, wrap filters around a call to log, cache, redact or stop it, and hand a conversation between several specialised agents. The trade-off is direction rather than quality. Microsoft has named Agent Framework the successor and says the majority of new features will be built there, so this codebase keeps getting fixes and will see some existing features reach general availability, but few new ideas. It suits a team extending something already written against it far better than a project starting from nothing today.

What can you do with semantic-kernel?

  • One kernel holds the model connections and the tools — You register chat model connections and plugins into a kernel, and the agents you build draw on it. Moving from OpenAI to Azure OpenAI, Ollama or Hugging Face is a change to that registration rather than to the agent code around it.
  • Any method in your code becomes a tool the model can call — Tag a method as a kernel function and its name, description and parameter types become the tool description sent to the model. A whole API can be imported from an OpenAPI description; importing tools from an MCP server, and publishing your own functions as an MCP server, are documented for Python only, with .NET served by samples in the repository while its documentation still says it is coming.
  • Three places to intercept what the model does — A prompt render filter runs before a prompt is assembled, a function invocation filter runs around any kernel function call, and an auto function invocation filter runs inside the automatic tool-calling loop. The documented uses are redacting personal data, caching answers, retrying against a different model, and stopping the loop once the answer is good enough. Filters hang off the kernel, so calling a chat completion service directly without passing the kernel skips them.
  • Five named patterns for several agents working together — Concurrent broadcasts one task to every agent, Sequential passes each result to the next, Handoff moves control between agents as the context demands, Group Chat runs a conversation under a manager, and Magentic widens that group chat to open-ended tasks. All five share one construct-and-invoke interface, so changing how agents collaborate does not mean rewriting them.
  • Conversation history stays local or lives in a hosted service — A thread carries the state of one conversation, either held in your own process or kept server-side by a service such as the Azure AI Agent service inside a Microsoft Foundry project. A service-backed agent accepts only its own matching thread type and raises an error immediately if you pair it with another.
  • The vector search half can be used on its own — Annotate a data model to declare its key, its filterable fields and its vectors, and roughly fifteen databases — Azure AI Search, Postgres, Qdrant, Redis, Weaviate, Pinecone, an in-memory store and more — sit behind one API. The documentation states these implementations do not depend on the core Semantic Kernel stack, and in .NET the abstractions ship as a separate package. It also lists a maintainer per connector: Oracle, Elastic and Couchbase maintain their own rather than Microsoft.

Before you choose semantic-kernel

  • Microsoft names Agent Framework the successor: Semantic Kernel v1.x is promised critical bug and security fixes for at least a year past Agent Framework's April 2026 general availability, and new work goes there.
  • Multi-agent orchestration and the Process Framework are still marked experimental, and in .NET ship only as prerelease packages; the earlier group-chat API is no longer maintained, though a migration guide is provided.

Star history

17 Aug to 30 Oct · +263

28.5k28.7k

Frequently asked questions

Is semantic-kernel free for commercial use?

semantic-kernel is released under the MIT licence — OSI-approved open source, which permits commercial use.

How can semantic-kernel be deployed?

semantic-kernel is available as Self-hosted / Runs locally.

semantic-kernel
Ask AI
GitHub