Kagent is a framework for building, deploying, and running AI applications on Kubernetes. After you deploy Kagent, use declarative APIs to create agents and MCP Servers and connect multiple LLMs.
Basic agent concepts
An agent interacts with users in natural language, using an LLM to generate responses and perform operations on the user's behalf.
Each agent consists of the following components:
-
LLM: Understands and responds to user requests. Examples include Qwen and ChatGPT.
-
Agent instruction: A system prompt that defines the agent's role, guidelines, available actions, and response format. Example: "You are a Kubernetes O&M engineer. You can help users manage their Kubernetes resources."
-
Tool: A function the agent uses to interact with its environment, such as an MCP Server (MCP protocol) or another agent (A2A protocol). For example:
-
An ACK MCP Server supports operations such as retrieving cluster information, running cluster diagnostics, and viewing Prometheus metrics.
-
A custom agent handles specific tasks, such as retrieving product information.
-
Core features
Kagent is a Kubernetes-native framework for agent deployment and orchestration:
-
Deploy an available agent in an ACK cluster with a YAML file.
-
Deploy an MCP Server in an ACK cluster with a YAML file and connect it to an agent.
-
Integrate with a self-developed MCP Server to create a custom agent.
-
Orchestrate multiple agents with a YAML file for multi-agent architectures.
-
Centrally orchestrate self-developed agents.
Core APIs
Kagent uses Kubernetes CustomResourceDefinitions (CRDs) as its core APIs:
-
ModelConfig: Configures AI model access, with parameters such as URL, model name, and API key for providers such as Qwen or OpenAI. -
RemoteMCPServer: Registers an HTTP-based MCP Server for use in agents. Supports servers deployed outside the cluster or as in-cluster Services. -
Agent: Defines an agent that combines an AI model, instructions, and tools. The model processes requests from humans or other agents based on the configured prompt and calls MCP Servers to perform tasks. Tools include MCP Servers and A2A-compatible agents.
Kagent architecture
Kagent consists of three main parts:
-
controller: Processes Kagent APIs and transforms them into agent applications in the cluster.
-
App/Engine: The core Kagent component, a Python application built on ADK. When an agent resource is created, the controller translates it into an Agent App in real time. Handles the conversation loop and supports the A2A protocol.
-
UI: A default web UI to create, manage, and interact with agents in a browser.
Usage notes
The community installation template has extensive RBAC permissions, but ACK restricts them. By default, the template only has permissions in the kagent namespace, such as creating and deleting Deployment and Service resources and reading Secret resources. You can create resources such as Agent and ModelConfig only in the kagent namespace. To use other namespaces, add a ClusterRole for Kagent roles.
Kagent APIs are in Alpha. Before upgrading, check the target version for incompatible changes.