Since the release of OpenAI’s ChatGPT, everything suddenly became “AI this, AI that.” From LLMs writing essays to agents managing tasks, it felt like every day brought a new buzzword. Somewhere along the way, I stumbled upon MCP servers and AI agents — and I got curious.

As a DevOps engineer, I’ve spent my fair share of time fighting YAMLs, forgetting kubectl flags, and Googling how to fetch logs from a broken pod at 2 AM. So I thought — what if AI could help here? Not just write code, but actually help run and manage Kubernetes.

About me

I'm Akash Jaiswal, a passionate engineer who wants to try and learn anything that I find interesting. One day I am reading about how compilers work and the other day I am Googling how to center a div.

I am an aviation geek and like discussing politics. I like traveling to new places, reading books, doomscrolling Twitter, and writing on Quora.

Yeah, trying my hands into photography as well :)

Connect with me on GitHub, LinkedIn, or Twitter.

The Idea: KAI

That’s when the idea hit me — can I build something that sits in my cluster and understands what I want in plain English? No YAMLs. No man pages. No “how do I scale a deployment” Stack Overflow searches.

So I started tinkering. I combined a local LLM, the MCP protocol, and my love for hacking around Kubernetes. The result? KAI — a small AI agent that translates human intent into kubectl magic.

This is the first part of KAI’s story. It’s still early, but I wanted to share the journey and get feedback. Maybe someone else out there is tired of remembering how to exec into a pod too :)

So sit back, relax and enjoy. Don't forget to check out the project and give your suggestions.

Key Features

KAI sits on your cluster, ready to:

• Translate your plain English instructions into exact kubectl commands

• Provide clear explanations and learning opportunities

🗣️ Natural Language Understanding

> KAI, find pods that aren't running in the default namespace
Analyzing request... ✓
Command: kubectl get pods --namespace default --field-selector status.phase!=Running

🛡️ Safety First Approach

> KAI, delete all pods
⚠️ Warning: This is a destructive operation!
Suggested safer alternative:
kubectl get pods --all-namespaces  # Review pods first
Use --namespace flag to target specific namespace

📚 Learning Opportunities

> [EXP] create a nodeport service
# NodePort Service Creation
A NodePort service exposes the application on a static port on each cluster node.
* Port range: 30000-32767
* Accessible from outside cluster
* Requires cluster node IP

Example command:
kubectl create service nodeport nginx --tcp=80:80

🔄 Real-time Execution

> Scale frontend deployment to 5 replicas
Validating... ✓
Current replicas: 3
Scaling... ⏳
Success! ✓
New replica count: 5

Screenshots

Technical Deep Dive

How It Works

sequenceDiagram
    participant User
    participant KAI
    participant LLM
    participant K8s

    User->>KAI: "Scale the frontend to 3 replicas"
    KAI->>LLM: Analyze request & generate command
    LLM-->>KAI: kubectl scale deployment frontend --replicas=3
    KAI->>KAI: Validate operation
    KAI->>K8s: Execute command
    K8s-->>KAI: Operation result
    KAI-->>User: Success confirmation & explanation

The user sends a natural language input like “scale the frontend to 3 replicas.” KAI sends that to a local LLM (via MCP protocol), which returns the corresponding kubectl command. KAI validates the operation, runs it on the cluster, and returns a success message.

Under the Hood

Right now, the architecture is intentionally simple:

• Go backend: Handles request parsing, validation, and kubectl execution

• Ollama LLM: Runs locally, no external APIs involved

• MCP Protocol: Used for consistent communication between KAI and LLM

• Kubeconfig-based access: Reads from the local kubeconfig to run commands inside the cluster

Validation is basic for now (e.g., blocking dangerous commands like global pod deletes), but I plan to add more safety checks and dry-run support soon.

References & Inspiration

• AI agent for k8s troubleshooting

• Go MCP SDK

Getting Started

1. Clone and setup:

git clone https://github.com/jaiakash/k8s-ai-agent
cd k8s-ai-agent
go mod download

2. Start the server:

go run server/server.go --kubeconfig=$HOME/.kube/config

3. In another terminal, start the client:

go run client/client.go

Future Roadmap

Short-term Goals

• [ ] Cluster state awareness

• [ ] Validate operations for safety

• [ ] Execute commands with appropriate permissions

• [ ] Learning from past interactions

Long-term Vision

• [ ] Integration with popular CNCF tools

• [ ] Advanced troubleshooting capabilities (https://github.com/k8sgpt-ai/k8sgpt)

Contributing

KAI is open source and welcomes contributions! Whether it's improving documentation, adding features, or reporting bugs, every contribution helps make Kubernetes more accessible.

Found this interesting? ⭐ Star the repo and share your thoughts!