kagent
by solo-dot-io, now a Cloud Native Computing Foundation (CNCF) Sandbox project
An open-source, Kubernetes-native framework for running AI agents as cluster resources. Agents are declared
as Kubernetes custom resources, connected to tools through MCP servers, and can call each other over A2A.
Aimed at DevOps and platform engineers.
Overview
kagent treats an AI agent as a workload, the same way Kubernetes treats a deployment. An agent is defined in
YAML, applied with the usual cluster tooling, and run as a pod that the kagent controller manages. Its
primary audience is DevOps and platform engineers who want agents to live next to the systems they operate
rather than in a separate agent platform.
Core model
- Agents as custom resources — an
Agentresource picks an LLM, attaches tools and selects the target
cluster. The controller reconciles it into a running pod and reports status conditions such as Accepted and Ready. - Tools through MCP — tools are exposed by MCP servers, and a
RemoteMCPServerreference connects an agent to
them. Coverage cites REST, gRPC, databases and internal APIs as connectable through a declarative spec. - Agent-to-agent — agents can discover and invoke each other over the A2A protocol, enabling delegation
and multi-agent workflows. - Kubernetes governance — agents and tools are managed with GitOps, kubectl and RBAC. Kubernetes admission
controllers apply to them as they do to other workloads. - Observability — traces, metrics and logs for prompts, tool calls and token use, exported through
OpenTelemetry, Prometheus and structured logging. - Model providers — OpenAI, Anthropic, xAI, Google Gemini, Azure OpenAI, AWS Bedrock, Vertex AI, Ollama and
Hugging Face, per the project site. - Interface — a web UI for choosing agents and chatting with them, including approving pending actions,
and a CLI installed withkagent install.
Installation
Installed as a Helm chart from OCI registries. The chart version used in Solo.io’s October 2026 walkthrough is 0.10.2.
The walkthrough runs it on a local kind cluster with a local Ollama model.
Guardrails
Solo.io’s October 2026 walkthrough treats the prompt as a request rather than a guarantee. It recommends three controls
that do not depend on the prompt:
- Tool allowlist — the model sees only the tools listed on the agent
- Approval gate —
requireApprovalpauses a run until a user approves or rejects an action - RBAC — the Kubernetes API server refuses any action the agent’s service account is not allowed to take
The walkthrough’s default configuration uses an unsecured authentication mode and a built-in database password, and
it is intended for a local lab only.
Project status
- Licence: Apache 2.0
- CNCF status: Sandbox, accepted 22 May 2025 (per the CNCF project page)
- Origin: first announced at KubeCon EU 2025, with the intention to donate it to the CNCF
- Community: about 2,700 total contributors from 765 organisations and about 2,700 GitHub stars, per the CNCF
project page
Related
- solo-dot-io — the company that created kagent
- mcp-agent-gateway-vendors — MCP and agent gateway vendors; a governance layer in front of agents is the
next step the walkthrough recommends