OpenAI-based Open Source tools for Kubernetes AIOps

AI nowadays is a hot topic everywhere, and the Kubernetes-powered DevOps world is no exception. It seems to be quite organic for software engineers who, in their nature, are huge automation enthusiasts. Thus, in light of all the hype that’s coming from ChatGPT, relevant projects for Kubernetes operators are beginning to pop up, too. Let’s see which Open Source tools, backed by OpenAI & ChatGPT, have recently surfaced to make the K8s operators’ life easier. Most of them are designed for terminal (CLI) usage.

Troubleshooting K8s with AI

1. k8sgpt

  • “A tool for scanning your Kubernetes clusters, diagnosing, and triaging issues in simple English”
  • Website:
  • GitHub:
  • GH stars: ~3100
  • First commit: Mar 21, 2023
  • ~700 commits, 32 releases, ~40 contributors
  • Language: Go

Launched by Alex Jones and marketed as “Giving Kubernetes Superpowers to everyone,” k8sgpt is the best-known, most prominent project of its kind.

It is a CLI tool with a primary command, k8sgpt analyze, designed to reveal the issues going on within your Kubernetes cluster. To achieve this, it uses so-called “analyzers”, which define the logic for each K8s object and the possible problems it may be encountering. E.g., an analyzer for Kubernetes Services will check whether a particular Service exists and has endpoints at all as well as whether its endpoints are Ready.

Identifying such an issue in and of itself is not a big deal, but the magic is hidden deeper. The magic happens when you ask k8sgpt to explain what you can do about your existing issues — this is performed by executing k8sgpt analyze --explain. The command will ask the AI for instructions for your specific case and display them for you. These instructions will include the actions to perform your troubleshooting, including specific kubectl commands you need to execute via simple copy & paste, thanks to the fact that your Kubernetes resources’ names are already in place.

An example of “k8gpt analyze –explain” output (source)

Speaking of mentioning your actual resources’ names, k8sgpt boasts an anonymization feature (--anonymize flag for the k8sgpt analyze command), which will prevent sensitive data from being sent to the AI system. Helpful, isn’t it? As of right now, it is not yet implemented in all the analyzers though.

Today, k8sgpt features built-in analyzers for numerous Kubernetes objects, including Nodes, Pods, PVCs, ReplicaSets, Services, Events, Ingresses, StatefulSets, Deployments, CronJobs, NetworkPolicies, and even HPA and PDB. It should not be too difficult to extend this set by creating your custom analyzers too.

Another benefit is that k8sgpt is not limited to a single AI system. Yes, OpenAI is the default AI provider giving you access to the well-known GPT-3.5-Turbo and GPT-4 language models. However, you can choose among other AI providers as well, which currently include:

  • Azure OpenAI;
  • Cohere (this one was added just recently, on July 20th);
  • LocalAI — a local model with OpenAI-compatible API (e.g., you can use it with llama.cpp and ggml — this will allow you to leverage the AI even in air-gapped environments);
  • FakeAI — for simulating AI system behavior without actually invoking it.

K8sgpt also features an API for integrations, which allows you to leverage external tools, invoking their capabilities to tackle your Kubernetes issues. The only currently implemented integration is for Trivy, the well-known Open Source security scanner. Upon enabling it via k8sgpt integration activate trivy (assuming that Trivy Operator is installed inside your cluster), you will have a new k8sgpt filter, called VulnerabilityReport. You can access it via k8sgpt analyze --filter VulnerabilityReport.

Last but not least k8sgpt feature is that you can install it as a Kubernetes operator inside your cluster. To do so, use the Helm chart provided here. After installing it and applying the k8sgpt configuration object (kind: K8sGPT), your cluster will be analyzed by the tool with the scan results stored in the Results objects. That means you will be able to see them by executing kubectl get results -o json | jq .

The k8sgpt operator architecture (source)

k8sgpt has generated impressive community interest. Focused on troubleshooting your K8s issues, it is packed with ready-to-use features designed just for that purpose. Moreover, it is also: a) flexible for leveraging different AI systems and b) extensible to benefit from custom analyzers and third-party tools integration.

2. Kubernetes ChatGPT bot

Created by Robusta, this project focuses on troubleshooting Kubernetes issues by integrating AI with your alerts displayed in Slack.

To leverage this bot, there are specific requirements you will have to follow:

  • have or be ready to install Robusta on top of Prometheus (VictoriaMetrics is also supported) and AlertManager;
  • use Slack.

If everything is in place, you will already have your monitoring alerts sent to Slack via an incoming webhook. This bot adds an “Ask ChatGPT” button to your alerts in Slack. Thus, clicking on it will query the AI (using your OpenAI API key) and bring you its response, instructing you on possible actions to mitigate the issue that caused this alert.

New Slack button powered by the K8s ChatGPT bot

As of right now, it’s as simple as that. However, the author suggests a possible further improvement on the way it works by supplying additional data — such as Pod logs and kubectl get events output — to the AI. If it piques your interest, third-party contributors are welcome on GitHub.

Kubectl AI-powered plugins

3. kubectl-ai

This project was launched the day before k8sgpt was born. However, its idea of applying the power of AI to Kubernetes is entirely different. Motivated by avoiding “finding and collecting random manifests when dev/testing things,” the author wanted to simplify Kubernetes resources’ manifests generation.

Installed as a kubectl plugin, kubectl-ai features the kubectl ai command. You can use it to obtain ready-to-use YAML manifests from AI based on your needs. Here’s a self-explanatory example from its README:

 kubectl ai "create an nginx deployment with 3 replicas"
✨ Attempting to apply the following manifest:
apiVersion: apps/v1
kind: Deployment
  name: nginx-deployment
	app: nginx
  replicas: 3
  	app: nginx
    	app: nginx
  	- name: nginx
    	image: nginx:1.7.9
    	- containerPort: 80
Use the arrow keys to navigate: ↓ ↑ → ←
? Would you like to apply this? [Reprompt/Apply/Don't Apply]:
+   Reprompt
  ▸ Apply
    Don't Apply

The “reprompt” option allows you to refine the resulting manifest by changing specific parameters. You can generate a few manifests at once, which makes sense for interrelated objects, such as Deployment and Service. When you’re happy with what’s suggested, you can apply it to your cluster with ease.

You can subsequently also modify your existing Kubernetes objects by asking kubectl ai to scale them or change other parameters.

As for the AIs, kubectl-ai supports OpenAI API, Azure OpenAI Service, and LocalAI for airgap cases (a recent addition merged on July 31st). While GPT-3.5-Turbo serves as its default language model, GPT-4 is supported as well.

4. kubectl-gpt

  • “A kubectl plugin to generate kubectl commands from natural language input by using GPT model”
  • GitHub:
  • GH stars: ~40
  • First commit: May 29, 2023
  • ~20 commits, 3 releases, 1 contributor
  • Language: Go

This plugin introduces the kubectl gpt command, whose sole mission is to make your wishes — i.e. requests stated in human language — come true in your Kubernetes cluster. Here are examples of what you can expect from this plugin as outlined in its documentation:

kubectl gpt "Print the creation time and pod name of all pods in all namespaces."
kubectl gpt "Print the memory limit and request of all pods"
kubectl gpt "Increase the replica count of the coredns deployment to 2"
kubectl gpt "Switch context to the kube-system namespace"

The outcome can be both informative output only and real actions that are affecting your K8s resources. Either way, it will execute a command, but first, it will display this command so you can see it and confirm that you’re happy to carry on with it. You can also disable these features (printing a generated command/asking for confirmation) if you wish.

Kubectl-gpt requires an OpenAI API key to work. GPT-3 only is supported, with GPT-3.5-Turbo enabled by default. Any human languages that are supported by OpenAI GPT API can be used.

This project is developed by a solitary enthusiast and hasn’t seen any updates since the end of May.

AIOps multi-tools for Kubernetes

All the projects described in this category were launched at about the same time. They also share similar ideas providing the user with various AI-assisted features while they are operating a Kubernetes cluster. All of them also feature similar stats: one or just a few contributors, roughly 100 stars, and dozens of commits. Let’s see what they offer and how they differ.

5. kopilot

Kopilot is the only one out of these 3 projects written in Go. It covers two functions, troubleshooting and auditing. So what do they do?

  1. Imagine you have a Pod that is stuck in Pending or CrashLoopBackOff. This is when the kopilot diagnose command will come in handy. It will reach AI for help and print you its answer with possible explanations as to why this has happened.
An example of “kopilot diagnose” output
  1. Not sure if your Deployment is good enough? The kopilot audit command, using a similar approach, will check it against the well-known best practices and possible security misconfigurations.

This tool will use your OpenAI API token and the human language of your choice for the answers. The README also hints that there will be an option to use other AI services in the future.

Sadly, the project hasn’t seen any commits since the beginning of April, raising obvious concerns.

6. kopylot

  • “An AI-Powered assistant for Kubernetes developers”
  • GitHub:
  • GH stars: ~70
  • First commit: Mar 28, 2023
  • ~70 commits, 5 releases, 2 contributors
  • Language: Python

This tool has similar audit and diagnose functions and goes one step further by providing the “chat” command. It brings you a very specific chatbot experience: you can ask in English for a particular action that will be transformed into a kubectl command. If the command it prints seems fine, you can confirm its execution. That’s what we’ve seen in kubectl-gpt.

An example of “kopylot chat” output

Kopylot also offers the “ctl” command, a simple wrapper for kubectl allowing you to execute any command directly, i.e. without any AI intervention. This feature seems to aim to make kopylot your best friend while working with Kubernetes instead of the good old kubectl, which is still always available, just in case.

Currently, kopylot supports the OpenAI API key only and can’t function with any other human languages. It relies on the text-davinci-003 GPT-3.5 model (it’s hardcoded), which is considered legacy. The support for using other LLM models is mentioned in the project’s roadmap, though.

The chances it will happen are dubious since its latest release is dated April 4th.

7. kube-copilot

Kubernetes Copilot brings the multi-tools feature set to the next level. In addition to Kubernetes troubleshooting, auditing, and perform-any-action functionality, this multi-tool can also generate manifests based on your prompt (like kubectl-ai does).

By the way, auditing in kube-copilot is more powerful than you might expect. While the tool has the “analyze” command to reveal possible issues in your K8s resources (the way kopilot audit does), it also features the “audit” command. The latter leverages the Trivy scanner to look specifically for security problems that your Pods might have.

Another thing to point out in kube-copilot is that it comes with two interfaces: CLI and web UI. The latter one is quite simple, yet perhaps it might still be a killer feature for some users.

kube-copilot web UI

And the last CLI feature to mention is that this tool is happy to…google for you right in the terminal. Well, perhaps there are some questions that might be better solved this way rather than in ChatGPT. It doesn’t sound like the most important thing for the K8s-related tool, in my opinion, though.

As for the AI support, kube-copilot works with your OpenAI API key or Azure OpenAI Service. It allows you to use both the GPT-3.5 and GPT-4 models.

Despite this tool being developed by a solitary individual, its commits’ history is the most consistent and therefore promising. There is neither a public roadmap nor any issues that will shed light on how the tool will evolve, though.


There are more OpenAI-based tools & services for Kubernetes, which I’d like to mention in this article as well.

Botkube by Kubeshop is a messaging bot (i.e. ChatOps) for monitoring and debugging Kubernetes clusters in Slack, Mattermost, Discord, or Microsoft Teams. Its recently added Doctor plugin connects with the AI in two different ways: 1) directly asking the chatbot any questions you have; 2) using the “Get Help” button, which appears just below the error events. The bot will reply with AI-generated recommendations regarding your question or specific issue. An OpenAI API key is required to enable this Botkube plugin.

An example of a chat with Botkube’s Doctor (source)

KubeGPT by metaKube is not a CLI tool but an online chat available in the web browser to talk with the AI about Kubernetes. Just like regular ChatGPT, it can answer general questions (e.g., Kubernetes architecture or some best practices) as well as generate specific YAML manifests for you.

Chat with KubeGPT by metaKube

Actually, it was okay to help me with other tech issues as well. For example, it didn’t mind giving information on Nomad. It even provided me with guidance on installing Ubuntu Linux. However, it anticipated that “its expertise lies in Kubernetes,” so, formally, it should not assist much with inappropriate requests. This service is currently in the beta phase.

MagicHappens is a PoC operator that is “meant for fun and experimentation only.” It defines a new CRD (kind: MagicHappens) that lets you describe tasks in a human language — e.g., “create such a namespace and such deployment there”. When the operator receives a YAML manifest with this description, it sends it to OpenAI to get a relevant YAML and applies this resulting manifest to your cluster. Note that there have been no new commits to this project since April.

Kube or Fake is a just-for-fun online service that gives you five Kubernetes terms generated by ChatGPT. Since some of them are real and others are fake, your task is to guess that correctly.


Finally, there is no surprise that various well-known tools have been recently embedding OpenAI-powered features too. There are actually quite a lot of them already, so I don’t even promise that I’ll list all of them. However, here are the examples I’ve seen so far — not necessarily Open Source, but at least directly related to well-known Open Source projects — presented in chronological order:

  • ARMO Platform, based on Kubescape, can generate custom controls based on your request specified in a human language and then processed by GPT-3. (February 2023)
  • KubeVela Workflow allows you to use OpenAI API for validating your Kubernetes resources, ensuring their satisfactory quality, etc. (April 2023)
  • Monokle by Kubeshop achieved an AI-assisted YAML resource creation, letting you leverage the AI to generate YAML manifests based on your prompt and validation policies. (June 2023)
  • Portainer introduced experimental support for ChatGPT in its Business Edition v2.18.3, which can provide ready-to-use answers on how to deploy an app. (July 2023)
  • Argo CD got an AI Assistant in the Akuity Platform. Powered by OpenAI API, it helps to detect certain issues and analyze logs, performs the actions you request it to, and answers your questions. (July 2023)


Has AIOps delivered us to the promised land yet? While there’s definitely room for improvement, the Open Source ecosystem already has viable options to offer Kubernetes administrators & users:

  • k8sgpt is the most successful project overall since it has attracted numerous contributors and users. Currently focused on troubleshooting, it’s very flexible and extensible, so I’m excited to see what its authors and growing community will bring in the foreseeable future.
  • ChatGPT bot by Robusta and Botkube by Kubeshop are good ChatOps-style options if you want to simplify processing your monitoring alerts in Slack by adding AI-generated thoughts on what’s going on and how to fix it.
  • Try kubectl-ai or kube-copilot if you need to automate your YAML manifests generation.

The whole AIOps thing is currently exploding. Many projects we can see today will disappear, while many new ones will emerge and become an integral part of the cloud-native ecosystem. I hope this article was helpful in providing an overview of the Open Source landscape we currently have.

P.S. This article was written in an old-fashioned way, completely manually, with no ChatGPT involved.

Comments 4

Your email address will not be published. Required fields are marked *

  1. Maria

    Great article! Thanks for the Botkube feature 😊

  2. Bob

    NONE of these tools are “ChatGPT-based”, there are OpenAI based, it’s different because ChatGPT is a chat implementation over OpenAI which store and keep your data to train it. please be rigourous on the naming because it’s totally none the same thing

    • Dmitry Shurupov

      Thanks a lot for your valid concern, Bob! I just made a few corrections to improve the respective phrases in the article.

      • bob

        You rocks. Your article is great, I’m now trying all these tools one by one 🙂