// the find
skyhook-io/radar
The missing open source Kubernetes UI. Topology, event timeline, and service traffic — plus resource browsing and Helm management.
Radar is a Kubernetes UI that runs as a single binary on your laptop, talks directly to the cluster API, and requires zero cluster-side installation. It covers an unusually wide surface: topology graphs, Helm management, GitOps for both FluxCD and ArgoCD, RBAC visibility, cluster audit, and a built-in MCP server so AI assistants can query your cluster through token-optimized summaries instead of raw YAML. At 2,300 stars and active daily commits, it's past the 'toy project' threshold.
Zero cluster-side footprint at this feature depth is genuinely rare — K9s has it but is terminal-only; Lens historically required agents or a cloud account. The image filesystem viewer is a standout: browse container image contents without pulling locally or execing in, using the cluster's own ImagePullSecrets, with disk-based layer caching. GitOps support goes deeper than a resource table — drift detection, stuck-loop detection, and mutations that refuse on terminating resources are the kind of operational detail most UIs skip. The MCP server returns pre-processed, deduplicated data to AI assistants rather than raw YAML, which actually reduces context waste instead of just wrapping kubectl.
Traffic visualization requires Hubble (Cilium), Caretta, or Istio — plain CNIs get nothing, which means a lot of clusters have a dead tab. Timeline storage defaults to in-memory, so you lose all event history when the process exits; SQLite mode exists but requires an explicit flag, which will catch people off guard. The local-first story breaks down for teams: shared access means deploying the in-cluster Helm chart plus setting up proxy or OIDC auth, which is not trivial and partially undermines the zero-install pitch. MCP is labeled beta and is the most differentiating feature — the thing that separates this from a nicer K9s is also the least proven part in production.