Radar crossed 3,000 GitHub stars. Nobody stars a Kubernetes tool from a landing page: each one of those followed someone pointing it at a cluster. That's the part we're grateful for. brew install skyhook-io/tap/radar github.com/skyhook-io/r...

A 200 from the Kubernetes API-server proxy does not prove your application's real traffic path works. NetworkPolicy may treat that path differently. Radar reports: "Reached via API server - not live traffic." How Reachability works: radarhq.io/blog/kuberne...

If you approve every proposal an agent makes without reading it, you do not have an approval gate. You have an autonomous agent and a ritual. Roy Libman on where the real line sits: radarhq.io/blog/ai-agen...

Your cluster deleted the events from 02:14 about an hour after they happened. That's the default TTL, and on EKS, GKE or AKS you can't raise it. Worse: the config change that actually broke you never emitted an event in the first place.

our CEO made the strongest case he could that AI agents will make our own kubernetes dashboard obsolete, then published it. the counterexample: the companies furthest into agents keep shipping UI. radarhq.io/blog/dashboa...

If you gave an AI agent a broken Kubernetes cluster, would you hand it kubectl - or the resource graph and the timeline of what changed? We benchmarked both. Same model, 52 faults. The connected-context path used 76% fewer tool calls, ran in half the time, and diagnosed slightly better.

Radar MCP vs kubectl: A Kubernetes Agent Benchmark on 52 Faults

We benchmarked Radar's Kubernetes MCP server against raw kubectl on 52 live-cluster faults. Radar improved diagnosis quality and cut tool calls by 76%.

radarhq.io

We pointed an AI agent at a broken cluster through Radar's MCP server: root cause in half the time, at a third of the tool calls, vs the same agent on raw kubectl. Method and numbers in the post. (v1.8 also added Prometheus querying over MCP, so the agent reads live metrics too.)

When you get paged, do you think "the checkout app is down" or "a deployment in namespace prod-7 is unhealthy"? Radar v1.8 bets on the first: it groups your cluster's resources into the applications they form. One Go binary, open source, no account. github.com/skyhook-io/radar