KAgent
Kubernetes-native framework for deploying AI agents and MCP servers as declarative cluster resources.
About
KAgent brings the Kubernetes operating model to AI agents: you declare an Agent, a Team, or a Tool as a custom resource and a controller reconciles it, the same way Deployments manage pods. It was built by Solo.io on top of Microsoft's AutoGen 0.4, speaks MCP natively for tool integrations, and ships with plugins for Argo, Helm, Istio, Prometheus, and kubectl. In May 2025 it was accepted into the CNCF as a Sandbox project.
The pitch is aimed squarely at platform and DevOps teams: agents that live inside cluster RBAC, get standard ServiceAccount identity, and can be observed with the same dashboards as everything else. The kmcp subproject adds CRDs for managing MCP servers declaratively too. It is early — the API surface is still shifting around v0.9.x and the community is small next to general-purpose frameworks — but for teams already running Kubernetes in production, it is the most native way to operate agents there.
What it can do
- Agents, Teams, and Tools as Kubernetes custom resources
- Native MCP server support plus kmcp CRDs for declarative MCP management
- Pre-built plugins for Prometheus, Helm, ArgoCD, Istio, and kubectl
- Standard Kubernetes RBAC and ServiceAccount identity for agents
- Controller-managed lifecycle, scaling, and retries
- Works with models served by KServe/llm-d or external APIs
Setup
- Install a Kubernetes cluster (kind, k3s, or your existing cluster) and the
kagentCLI. - Deploy the kagent controller:
kagent install. - Write an Agent custom resource with your model provider, tools, and system prompt.
- Apply it with
kubectl apply -f agent.yaml— the controller spins up the agent. - Optionally declare MCP servers via the kmcp CRDs and compose agents into Teams for multi-agent workflows.
Prerequisites
A Kubernetes cluster; a model provider API key (OpenAI, Anthropic, or an in-cluster model endpoint); kubectl.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: CNCF Sandbox listing confirms acceptance on May 22, 2025; GitHub API shows 4,028 stars, Apache-2.0 license, pushed Oct 15 2026; Solo.io's announcement and a KubeCon Europe tutorial corroborate the Agent/Team/Tool CRD model and the AutoGen 0.4 base. Strengths: first genuinely Kubernetes-native agent framework, MCP-native tooling, agents inherit cluster RBAC and observability. Docked for: (1) Sandbox maturity — API surface still shifting around v0.9.x; (2) Kubernetes-only worldview, little value outside clusters; (3) modest community next to general-purpose frameworks. License: Apache-2.0.
★ 4,028 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 10.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: starScore from 4,028 stars on kagent-dev/kagent (log scale vs 298k peer max)
- · Maintenance 10: pushed 2026-10-15 (recency 10); active CNCF Sandbox project with steady commits
- · Openness 10: Apache-2.0 (0.7*10) + official docs verified at kagent.dev (0.3*10)
- · Manual 8: CNCF Sandbox status and CRD model corroborated by CNCF listing, Solo.io announcement, and KubeCon tutorial; docked for Sandbox maturity (API still shifting), Kubernetes-only scope, small community
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.