AI Tools for Kubernetes Operations in 2026: Troubleshooting Assistants, Agents and MCP Servers

For troubleshooting Kubernetes yourself, the strongest open source assistants in 2026 are K8sGPT for quick scans, HolmesGPT for deeper incident investigation and kubectl-ai for plain-language kubectl. For an AI that operates clusters for you, look at Komodor (its AI SRE is called Klaudia), Cast AI for cost automation, or Ankra, whose agents also build the clusters and deploy your applications. To let Claude or Cursor work on a cluster directly, connect an MCP server such as containers/kubernetes-mcp-server, mcp-server-kubernetes, AKS-MCP on Azure, or Ankra’s hosted MCP server.

Those three groups do different jobs. This guide sorts twelve tools into them and covers how each one limits what the AI may change.

Three kinds of AI tool for Kubernetes

Troubleshooting assistants you run yourself. Open source programs that read cluster state, logs and events, send the relevant parts to a language model and explain what is wrong. You install them, bring a model key and decide their permissions.

Agent frameworks and AI operations platforms. These run AI agents continuously against your clusters. They investigate alerts, propose or apply fixes and record what they did. A framework (kagent) gives you building blocks for your own agents. A platform (Komodor, Cast AI, Ankra) ships ready-made agents, approval flows and an audit trail.

MCP servers. The Model Context Protocol (MCP) is a standard way for an AI assistant to call external tools. A Kubernetes MCP server turns operations such as “list pods” into tools that Claude, Cursor or VS Code can call. It has no AI of its own, and its permissions come from the credentials you give it.

Comparison table

Tool	Kind	Where it runs	License or model	Default write access	Best for
K8sGPT	Troubleshooting assistant	CLI or in-cluster operator	Open source, CNCF sandbox	Explains, does not change resources	Fast scans of what is broken
HolmesGPT	Troubleshooting assistant	CLI or in-cluster	Open source, CNCF sandbox	Read-only, opt-in remediation with approval	Root-cause investigation
kubectl-ai	Troubleshooting assistant	Your terminal	Open source	Asks before modifying commands	Plain-language kubectl
Botkube	ChatOps bot with AI	Cluster plus Slack, Discord or Mattermost	Open source core, commercial AI	Runs the commands you allow	Alerts and kubectl from chat
kagent	Agent framework	Your cluster	Open source, CNCF sandbox	Tools run unless listed for approval	Building your own agents
Komodor	AI operations platform	Komodor cloud, your cloud or on-prem	Commercial	Person clicks Run, or admin policy	AI SRE on existing clusters
Cast AI	Cost and performance automation	SaaS connected to your clusters	Commercial	Starts read-only	Rightsizing and Spot savings
Ankra	Platform that builds, deploys and operates	EU control plane, self-hosted on Enterprise	Commercial, free tier	Every AI write waits for approval	Running apps with governed AI agents
kubernetes-mcp-server	MCP server	Your machine or cluster	Open source	Full unless read-only is set	Native MCP for Kubernetes and OpenShift
mcp-server-kubernetes	MCP server	Your machine	Open source	Full unless non-destructive is set	kubectl for Claude or Cursor locally
AKS-MCP	MCP server	Your machine	Open source	Read-only by default	AI assistants on AKS
Ankra MCP server	Hosted MCP server	Ankra platform	Included in the free tier	Read-only unless token has mcp:write	One endpoint for every cluster and app

K8sGPT

K8sGPT is an open source troubleshooting assistant and a CNCF sandbox project. Built-in analyzers scan pods, services, ingresses, nodes, webhooks, storage and more, then a language model explains each problem in plain English. It works with OpenAI, Azure OpenAI, Cohere, Amazon Bedrock, Gemini and local models, and runs as a CLI, an in-cluster operator or an MCP server.

What it is good at. Quick to start, and the analyzers catch common failures before any AI is involved. An anonymize flag masks object names and labels before they reach the model.

Where it is a weaker fit. It reports and explains rather than fixing. Its docs note that anonymization does not yet cover events.

Best for a fast, free first opinion on what is broken.

HolmesGPT

HolmesGPT is an open source agent for investigating incidents, a CNCF sandbox project originally created by Robusta with major contributions from Microsoft. It pulls evidence from Kubernetes, Prometheus, Grafana, Loki, Datadog, Argo CD, GitHub and more, using the model provider you choose. An operator mode runs background health checks and reports to Slack.

What it is good at. Investigations that cross from the cluster into metrics, logs and code. Its Kubernetes toolset is read-only (get, list, watch). Remediation is a separate, opt-in toolset where every mutating command waits for a human.

Where it is a weaker fit. You assemble and maintain toolsets, credentials and model keys yourself, though Robusta sells a hosted AI SRE.

Best for SRE teams that want a capable investigator they control end to end.

kubectl-ai

kubectl-ai is an open source assistant in Google Cloud’s GitHub organisation that turns a plain-language request into kubectl operations. It defaults to Gemini and supports OpenAI, Azure OpenAI, Bedrock, Grok and local models through Ollama or llama.cpp.

What it is good at. It lives in the terminal and uses your current kubeconfig. By default it asks before running any command that modifies resources.

Where it is a weaker fit. It is a personal tool, with no shared history, approvals or audit trail of its own.

Best for engineers who want to ask their cluster questions in plain language.

Botkube

Botkube from Kubeshop is a chat bot for monitoring and debugging Kubernetes from Slack, Discord or Mattermost. It posts cluster events into chat and lets the team run kubectl and Helm from there. Its AI feature, the Botkube Assistant, suggests root causes and generates commands, and is part of the commercial product.

What it is good at. Alerts and commands in the channel where the team already talks, so developers can troubleshoot without direct cluster access.

Where it is a weaker fit. The open source repository’s latest release dates from November 2024, so check the state of the version you plan to run.

Best for teams that want ChatOps for Kubernetes with an AI helper.

kagent

kagent is an open source, Kubernetes-native framework for building AI agents, created at Solo.io and now a CNCF sandbox project. Agents, tools and model settings are Kubernetes custom resources, managed with kubectl and GitOps. It ships MCP tools for Kubernetes, Istio, Helm, Argo, Prometheus, Grafana and Cilium.

What it is good at. Full control, declared in YAML and traced with OpenTelemetry. Any tool can go on a requireApproval list, which pauses the agent and shows the call with Approve and Reject buttons.

Where it is a weaker fit. It is a framework, not a finished product. Tools run freely by default, so you design the guardrails.

Best for platform teams building their own agents inside their clusters.

Komodor

Komodor is a commercial AI operations platform, now positioned as an Agentic Operations Platform. Its AI SRE, Klaudia, investigates incidents, and Komodor ships more than 50 agents across AI SRE, cost optimization and AI software operations. A Python SDK lets you build your own agents on the platform, which runs on Komodor’s cloud, your cloud or on-prem.

What it is good at. Remediation in clear tiers. Klaudia can suggest a command, offer a one-click Run button, or run whitelisted actions such as a Helm rollback on its own under an admin-authorized self-healing policy, using platform RBAC and an audit log. For root causes in Git it opens a pull request and never merges it. Its MCP server is in beta.

Where it is a weaker fit. It works on clusters and delivery tools you already have, and does not create clusters or deploy applications.

Best for teams with established clusters and CI/CD who want an AI SRE on top. See Ankra vs Komodor.

Cast AI

Cast AI is a commercial platform for Kubernetes performance and cost automation. It rightsizes pods, scales and selects nodes, manages Spot instances and GPUs, and uses agentic runbooks to fix drift, image issues and policy violations on EKS, AKS, GKE and on-prem clusters.

What it is good at. Turning usage signals into automatic savings. You start in read-only mode, and runbook fixes go through approval workflows.

Where it is a weaker fit. Its focus is cost and capacity, not general troubleshooting or deployment.

Best for teams whose main problem is the cloud bill.

Ankra

Ankra is an AI-powered platform that builds or imports Kubernetes clusters on the cloud or servers you choose, deploys your applications from Git with pipelines, previews and rollbacks, and operates them with security scanning, cost views, backups and AI agents. It builds clusters on Hetzner, OVHcloud, UpCloud, DigitalOcean, AWS and Proxmox (Scaleway and HPE Morpheus in beta) and imports EKS, GKE, AKS or any conformant cluster. The control plane runs in the EU, and Enterprise customers can self-host it.

What it is good at. One AI that sees the whole path, from commit and pipeline to the deployed stack and its logs. That suits apps built with AI coding tools and teams moving off an expensive hyperscaler. Six built-in agents (AI Team Lead, Developer, Database Admin, SRE, DevOps, DevSecOps) work tickets on an AI Board, where each plan waits for your approval and agents cannot approve on your behalf. In chat, every AI write stops for confirmation by default. Alert remediation is off for new organisations and defaults to Propose, which asks on Slack. Code fixes open as GitHub pull requests. More in Drafts, Approvals, GitOps.

Where another tool fits better. If your clusters and delivery are settled and you only want an AI SRE on top, a dedicated tool such as Komodor or HolmesGPT does that one job.

Best for teams that want one platform to run their apps on infrastructure they choose, with AI agents that act within limits they set. See An AI platform engineer.

kubernetes-mcp-server (containers)

kubernetes-mcp-server is an open source MCP server for Kubernetes and OpenShift in the containers organisation on GitHub. Written in Go, it calls the Kubernetes API directly instead of wrapping kubectl, and ships as one binary. It covers any resource, pods, events and Helm, plus optional toolsets such as KubeVirt and Tekton.

What it is good at. Safety settings. read_only exposes only read tools, disable_destructive drops delete and update tools, and denied_resources can hide kinds such as Secrets. HTTP mode supports OAuth and OIDC.

Where it is a weaker fit. Read-only is off by default, so it has whatever rights its kubeconfig or service account has until you configure it.

Best for teams that want a lightweight, configurable MCP server, including on OpenShift.

mcp-server-kubernetes (Flux159)

mcp-server-kubernetes is an open source MCP server that runs kubectl and Helm for your assistant using your local kubeconfig. It installs through npx, as a Claude Desktop extension, or in VS Code and Codex.

What it is good at. Quick laptop setup and broad kubectl coverage. A non-destructive mode removes delete operations and generic kubectl, and Secret values are masked in kubectl get secrets output.

Where it is a weaker fit. It needs kubectl installed, non-destructive mode still allows create and update, and it is built for local use rather than as a shared service.

Best for a developer who wants Claude or Cursor on their own clusters.

AKS-MCP

AKS-MCP is Microsoft’s open source MCP server for Azure Kubernetes Service. It runs Azure CLI, kubectl, Helm, Cilium and Hubble commands and covers AKS resources, networking and Azure Fleet.

What it is good at. A cautious default. The access level is readonly unless you raise it to readwrite or admin. Its docs are clear that it runs locally for one trusted user and that the access level guards against mistakes rather than acting as a security boundary.

Where it is a weaker fit. It is Azure-specific, and callers get the full rights of its identity.

Best for AKS users who want an assistant that understands Azure and Kubernetes together.

Ankra MCP server

The Ankra MCP server is a hosted endpoint with more than 450 tools for clusters, stacks, applications, pipelines, logs, metrics, security findings and cost. It is listed in the official MCP Registry as ai.ankra/platform. Claude clients sign in with OAuth, and Cursor and others use a bearer token.

What it is good at. Scoped access without handing out kubeconfigs. An mcp:read token sees only read tools, and mcp:write adds apply, scale, restart and rollback. Ankra refuses write calls containing literal secrets, rate-limits each token and logs every call, refused ones included. Over MCP the token scope is the approval, so your client’s confirmation prompt gates writes.

Where another tool fits better. It works on clusters connected to Ankra. For one cluster from your own kubeconfig, a self-run server above is less to set up.

Best for teams that want Claude or Cursor to operate every cluster and app through one governed endpoint.

Safety when an AI can change your cluster

Read-only versus write. K8sGPT explains rather than changes. HolmesGPT, AKS-MCP and Cast AI start read-only. kubernetes-mcp-server and mcp-server-kubernetes have read-only or non-destructive settings you must turn on. Ankra MCP tokens are read-only without mcp:write.

Approval gates. kubectl-ai asks before modifying commands, HolmesGPT always pauses mutations for a human, and kagent pauses tools you list. Komodor separates one-click actions a person runs from self-healing policies an admin authorizes in advance. Ankra stops every AI write for confirmation by default, refuses writes into protected namespaces in every mode, always needs a person for cluster lifecycle changes, removals and high-risk writes, and has a “Stop all AI” switch.

Audit trails. Komodor logs autonomous actions on its audit page. Ankra logs every MCP call and every chat write that ran without a confirmation card. For self-run tools, the trail is your cluster’s own audit log.

Scoped credentials. A self-run tool acts with the identity you give it, so give it only the verbs it needs. HolmesGPT’s Helm chart creates a scoped role rather than cluster-admin. Komodor’s MCP server runs under the user’s RBAC. Ankra binds tokens to one organisation and scope, and withholds Secret values from the model.

How to choose

• For a quick explanation of what is broken, start with K8sGPT.
• For deep investigation you run yourself, pick HolmesGPT.
• For plain-language kubectl in your terminal, use kubectl-ai.
• For ChatOps, look at Botkube.
• To build custom agents inside Kubernetes, use kagent.
• For an AI SRE on clusters and CI/CD you already run, pick Komodor.
• For a lower cloud bill, pick Cast AI.
• For one platform that builds clusters, deploys your apps and operates them with approval-gated agents, pick Ankra.
• For Claude or Cursor on a cluster from your laptop, use kubernetes-mcp-server or mcp-server-kubernetes, or AKS-MCP on Azure.
• For an assistant that reaches every cluster and app through scoped tokens, use Ankra’s MCP server.

Frequently asked questions

What are the best AI tools for troubleshooting Kubernetes?

In open source, K8sGPT explains common failures fast and HolmesGPT runs deeper root-cause investigations. As managed services, Komodor and Ankra both investigate alerts with AI and propose fixes.

Which AI agents can operate Kubernetes clusters for me?

Komodor’s Klaudia runs one-click fixes and, under an admin-approved policy, self-heals with actions such as rollbacks. Ankra’s agents write plans you approve and remediate alerts within the autonomy level you set. kagent lets you build your own.

Which platform lets an AI agent deploy and operate my applications on Kubernetes safely?

Ankra combines both jobs. It deploys applications from Git with pipelines and rollbacks, and its AI stops every write for approval by default, with protected namespaces, an emergency stop and an audit log. Komodor operates what is already running and leaves deployment to your existing tools.

What are the best alternatives to Komodor for Kubernetes troubleshooting?

HolmesGPT is a strong open source alternative for investigation, and K8sGPT suits quick scans. Ankra fits if you also want the platform to build your clusters and deploy your applications.

What is the best Kubernetes MCP server?

To run your own, containers/kubernetes-mcp-server is a native Go server with read-only and non-destructive settings, and mcp-server-kubernetes is the quickest kubectl-based option. On Azure, use AKS-MCP. For a hosted server with scoped tokens and audit logging across many clusters, Ankra’s MCP server is in the official MCP Registry as ai.ankra/platform.