Hyground vs K8sGPT: a supported platform, not a cluster scanner

K8sGPT does one job well, and it is free. More than thirty analyzers walk your cluster objects, and what they find comes back in plain language. Point it at OpenAI, Azure, Cohere, Amazon Bedrock, Google Gemini or a local model, and with the local one nothing leaves your network. An operator handles continuous scanning and feeds Prometheus and Alertmanager, and an MCP server lets other agents drive it. The project is active. Version 0.4.38 shipped on 2026-09-01, with roughly a hundred commits in the last six months. Hyground is a different kind of thing. It correlates the cluster with your logs, metrics, runbooks and tickets into one investigation, and it comes with a vendor.

REVIEW NOTES FOR THE TEAM. Draft only. Delete this block before publishing.

Claims checked 2026-09-14 against the K8sGPT project README on GitHub and the repository's own release and commit history through the GitHub API. On our side against the product docs repo at commit d4604a6b and the current hyground.ai pricing and security pages. Only claims that need a second pair of eyes are listed.

• We described an active project as a dying one, and the numbers were wrong. Please read this one first. The old card said 36 commits from 10 contributors in six months and a last release of v0.4.33 on 2026-05-13. As of today the project has shipped v0.4.38 on 2026-09-01, has roughly a hundred commits in the last six months, over eight thousand stars, and was pushed to today. The staleness argument is gone and an FAQ says plainly that we got it wrong.

• A claim about a named person's employer is off the page. The old card stated where the project's founder works full time, to imply the project is a side project. It is unverifiable from here, it is about a person rather than a product, and it is the kind of line that gets screenshotted. It is gone and should not come back.

• K8sGPT supports local models, so bring-your-own-model is not a differentiator against them. Their README lists OpenAI, Azure, Cohere, Amazon Bedrock, Google Gemini and local models. The page now credits that directly and says the two products are closer on sovereignty than anything else on this site. That is a real concession and it needs your agreement.

• Three "no" claims were too broad and are narrowed. They have a log analyzer among their optional analyzers, their operator integrates with Prometheus and Alertmanager for continuous scanning, and they ship an MCP server so other agents can drive them. The page no longer says "no log search, no metric correlation"; it contrasts a deterministic analyzer pass plus one explanation call against a multi-source investigation.

• Our own compliance line was out of date and understated us. The old copy said an ISO 27001 attestation was "due in the next three months". We have been ISO/IEC 27001:2022 certified since 2026-08-24, and the page now says so.

• K8sGPT gets the table's one competitor tick, on cost and licence. Free, Apache 2.0, no vendor in the supply chain. The page also says in the fits section and an FAQ that plenty of teams will never need more than K8sGPT. That is deliberate and it is a positioning call.

• The analyzer count is now "more than thirty". Their README lists fifteen enabled by default and about twenty optional. The old "roughly thirty" was close but the page states it as a floor rather than an estimate.

• Open decision. The six cards are prose paragraphs; the Dash0 layout uses bullet lists. Same question as the rest of the family.

• Nothing a visitor sees changes until this is published. The live site is still served from Strapi.