HolmesGPT vs K8sGPT: Kubernetes checks and incident investigation
Compare K8sGPT analyzers and HolmesGPT toolsets for Kubernetes checks, incident investigation, access control, and model usage.
An ImagePullBackOff gives the on-call engineer a concrete place to start. The Pod status and recent events can already tell you whether to check the image name, the registry, or credentials. A checkout service whose latency rose from 200 milliseconds to three seconds is a different investigation. Every Pod might still be ready, and the slow part might be a database call or a downstream dependency.
K8sGPT is a good fit for the first job when an analyzer already checks the symptom. HolmesGPT is more useful when the next question depends on what you just learned. The cost difference starts with the evidence each workflow needs. This comparison uses K8sGPT CLI 0.4.41, K8sGPT Operator 0.2.29, and HolmesGPT 0.46.0 as reference points.
Where each tool starts
| Decision | K8sGPT | HolmesGPT |
|---|---|---|
| Starting point | Run selected analyzers against Kubernetes resources | Ask an investigation question and use configured tools to gather evidence |
| Model dependency | The analysis can run without a model. --explain can send eligible findings to a configured backend |
Configure a supported model provider for the investigation, along with toolsets and access |
| Kubernetes identity | CLI uses kubeconfig. An in-cluster Operator uses its configured identity | CLI uses kubeconfig. In-cluster core and log tools use the configured ServiceAccount |
| Recurring work | The K8sGPT Operator reconciles a K8sGPT resource and writes findings as Result resources | Holmes Operator can schedule or trigger a question-based health check |
| Main operating work | Keep analyzer coverage, scope, and result capture current | Maintain data source access, tool inputs and outputs, and limits on investigation scope |
A CLI scan for a known namespace can start with k8sgpt analyze --namespace=checkout --filter=Pod,Service --output=json. The namespace and filter choose what to analyze. The kubeconfig identity still determines what the CLI can read. Keep permission and retrieval errors visible so that an empty findings list does not look like a clean check. K8sGPT CLI · Filters and integrations
Make recurring checks narrow enough to interpret
For known Pod problems or a Service with no ready endpoints, check which analyzer covers the symptom and what resources it needs. The output is useful when you can say which namespaces, analyzers, and time the run represents. If you need to check on a schedule, the in-cluster K8sGPT Operator exposes findings as Result custom resources. Retain the observations you will need later in your existing monitoring or incident history. K8sGPT Operator · Custom analyzers
When you want a plain-language explanation, enable --explain with the chosen backend configured. Review what the enabled analyzers collect before sending findings to that backend. --anonymize masks selected data. In 0.4.41 the CLI flag description specifically excludes events from that masking. A more capable model can explain a finding more clearly, while coverage still depends on the analyzers. K8sGPT privacy notes · Pinned CLI flags
Holmes Operator also has scheduled health checks. Its schedule runs a question against a model and the configured tools, so the work and cost can change with the answers the tools return. A recurring analyzer run and a recurring investigation can both be useful. They take different inputs and produce different evidence. Holmes Operator scheduled checks · Triggered checks
Give cross-system investigations the right evidence
Suppose the latency alert fired at 10:02, a scale-up completed at 10:05, and you opened the cluster at 10:10. The current replica count shows the state after recovery. To investigate the earlier incident, you need the relevant request traces or latency series, the deployment timeline, and identities that let you join them. An earlier deploy is a lead worth checking before treating it as a cause.
HolmesGPT can follow those leads when the sources are connected. The built-in Kubernetes Core and Logs toolsets can read resources and Pod logs under the identity they run with. For a CLI run, check the kubeconfig. With the standalone 0.46.0 Helm chart's default values, the built-in role grants read access across the cluster. Scope the grants to the namespaces you need and verify the rendered bindings. Read-only access can still expose operational data. HolmesGPT Kubernetes toolsets · Namespace-scoped access · Pinned Helm values · Pinned role template
Connect each metrics, logs, tracing, or other backend separately and check that the tool can see the time window and service you need. Pick a model supported by your provider configuration. Record where its inputs and tool outputs go. Keep an inaccessible data source visible as a gap in the investigation. For a first deployment focused on evidence, use read permissions and leave any write-capable extensions out of that workflow. HolmesGPT toolsets · Model providers
Decide what you are prepared to operate
A scoped K8sGPT scan is straightforward to size when the namespace and analyzers are fixed. Repeated explanations still incur model usage. With HolmesGPT, the number of tool queries, lookback window, and amount of returned context vary by question. Budget for the connections and their maintenance as well as for model calls. Grouping repeated alerts into an incident can avoid asking the same investigation to run several times.
If the evidence is already in Guance, the first step may be simpler. From a supported trace or infrastructure page, Obsy Copilot can use the selected object, time range, and filters within available permissions. Teams that need a continuing workflow can configure tools and tasks in Obsy Agent Teams. External systems still need connections and permissions, and plan and credit terms matter for repeated runs. Guance Obsy AI
Choose K8sGPT when most of the work is a defined set of cluster checks. Choose HolmesGPT when an investigation regularly crosses Kubernetes, telemetry, and other systems and someone can maintain those connections. The comparison is based on official documentation and pinned source. The example is illustrative, and root-cause accuracy would require a separate controlled test.
Contact us
Join the community
to join the community
Try Guance
Start online and pay only for what you use.
Get startedChoose a Guance plan
