DevOps AI Toolkit (dot-ai)

What it does

dot-ai, also called the DevOps AI Toolkit, is a Kubernetes-centred engine that you deploy into your own cluster with Helm. A coding assistant then connects to it over HTTP. Its tools take a plain-language request and turn it into cluster work: they answer questions about what is running, suggest and generate manifests for a new deployment (installing third-party Helm charts from ArtifactHub when nothing in the cluster fits), plan and apply day-2 changes such as scaling, updates, rollbacks and deletions after a dry run, look for the root cause of a failing workload, and estimate the blast radius of an operation before you run it.

The engine sends each request to a language model you configure (Anthropic Sonnet by default; Google, Amazon Bedrock, Kimi, Alibaba, GitHub Copilot, OpenAI, xAI or an OpenAI-compatible endpoint can be used instead). It keeps organisational data in a bundled Qdrant vector store: scanned cluster capabilities, deployment patterns, policy intents (which can be turned into Kyverno policies) and an ingested documentation knowledge base it can search. It also serves shared prompts as slash commands and generates repository governance files such as LICENSE, CONTRIBUTING and workflows.

The v2.4.0 source registers 9 MCP tools in src/interfaces/mcp.ts. The same tools are also available as REST endpoints, and there is a separate CLI for the same engine.

What it needs

A Kubernetes cluster (1.19+) with kubectl access, Helm 3 and an ingress controller (or Gateway API on 1.26+). Install the dot-ai-controller chart and the dot-ai chart from oci://ghcr.io/vfarcic. You need a static token (DOT_AI_AUTH_TOKEN) and an API key for the chosen model provider (Anthropic by default). Models need a 200K+ token context window. Embeddings come from OpenAI, Google or Bedrock, or from the in-cluster localEmbeddings option, which is amd64 only. OAuth sign-in needs HTTPS and Dex enabled.

Limits

The server does not run as a local stdio process; it runs only as an HTTP service in a cluster. Without a model API key it starts, but only prompts and the REST gateway work, not the AI tools. recommend, operate and remediate throw an error unless the agentic-tools plugin is loaded (mcp.ts). recommend and manageOrgData give poor results until capability scanning has run. Git pushes from recommend go to github.com only by default (gitops.allowedRepoHosts). The docs call the Copilot provider unofficial and say it may break without notice, and say custom endpoints (Ollama, vLLM, LocalAI) are not yet tested. DeepSeek is not supported. The README badge marks the project beta.

Who it suits

It suits platform teams who already run Kubernetes and want an assistant to query, deploy into and change their clusters in line with their own patterns and policies, and who are willing to host a Helm release, a vector store and a paid model key. It is a heavy choice for someone who only wants to read cluster state from a laptop, because nothing runs without a cluster deployment. The tools can apply and delete resources, so it fits teams that will review dry-run plans before approving them. It is a community project by Viktor Farcic, not a Kubernetes or CNCF release.