Skip to content

Home

k8s-servant k8s-servant

A read-only AI SRE for Kubernetes clusters. It looks at a cluster the way an experienced cluster administrator would: it finds what is wrong, explains why with evidence, says what to do about it, and remembers what it already told you. It never changes anything. You choose the model: Azure AI Foundry, OpenAI, Anthropic, a self-hosted model through Ollama, or any OpenAI-compatible server. Results go to the terminal and to an MS Teams channel.

Warning

Early-stage project. k8s-servant is young and has run on only a few clusters so far. Findings can be wrong or incomplete, and settings, Helm values and commands may change before v1.0. It cannot change your cluster, but verify its findings before you act on them, and please report what breaks.

  • Try it in 10 minutes

    A throwaway kind cluster with something broken in it, a model of your choice, and your first questions.

    Getting started

  • Choose a model

    One recipe per provider: settings for a laptop, a Helm values file, sign-in, and what doctor should print.

    Model providers

  • Run it in a cluster

    One pod from the Helm chart, with scheduled health checks, a morning report and Teams cards.

    Deploying to a cluster

  • Why it is safe

    Read-only RBAC without Secrets, an allow-listed kubectl, no shell, and what data goes to the model.

    Security model

What it does

  • Answers questions. "Are there unhealthy pods?", "why is X pending?", "is Flux healthy?". It investigates with read-only tools and answers with evidence: object, field, event, time.
  • Watches. Every hour it runs a health check and remembers the result, so it reports problems as new, still open or resolved instead of repeating them. When nothing changed, it does not even call the model.
  • Reports every morning. On weekday mornings it posts how the cluster did since the previous report, what came and went during the night, and what needs attention today.
  • Learns your platform. What is normal on your platform goes into an environment description and skills, plain Markdown delivered through Helm values (teaching the agent).

What it never does

  • It never creates, changes or deletes anything in the cluster. Asked to fix something, it describes the fix for you to apply.
  • It never reads Secrets: its RBAC has no access to them, the kubectl guard refuses them and the MCP server is configured to deny them.
  • It never runs a shell. kubectl runs with an allow-list of read-only verbs and flags.

How it works: the flow from a question to an answer. The source, issues and releases are on GitHub.