Home
AgentOps for Kubernetes fleets, done safely.
This wiki is the guided tour: the whole journey from the problem, through the idea and the design, to running it, improving it, and helping build it. If you just want to install and go, the README is faster. If you want to understand why it is built this way, start here.
flowchart LR
A[Why this exists] --> B[The idea]
B --> C[How it works]
C --> D[Implementation]
D --> E[Getting started]
E --> F[Use cases and impact]
F --> G[Evaluation]
G --> H[What's next]
H --> I[Contributing]
- Why This Exists - the 2 a.m. problem and why the obvious answer fails
- The Idea - the hub as a safe control point for agents
- How It Works - architecture and the request/approval workflow
- Implementation - what is actually in the code, module by module
- Guardrails Deep Dive - the four layers and the threat model
- Getting Started - laptop fleet to production, step by step
- Use Cases and Impact - who this helps and how much
- Evaluation - measuring what agents can and cannot be trusted to do
- What's Next - the roadmap and how to shape it
- Contributing - how to help
- FAQ - naming, comparisons, common questions
Plus Test Results - the latest end-to-end run on a real kind
fleet, tool by tool, refreshed by ./hack/e2e-local.sh.
Your team runs many Kubernetes clusters. People are asking whether an AI agent
can take the on-call load. Handing a model kubectl with admin is fast to try
and dangerous in production. This project instead exposes your multi-cluster
hub (Open Cluster Management) to agents as a small set of typed MCP tools, and
puts four independent layers between the model and your clusters: static
checks, Kyverno policy admission, human approval tokens, and least-privilege
RBAC. The agent investigates freely and proposes changes; nothing reaches a
cluster without passing policy and a human. Every action is traced and logged.
Maintained by Sandeep Bazar · YouTube: Tech Horizon Hub · sponsorship via LinkedIn