Klustercost vs KubeLedger

Klustercost gives Kubernetes users cost visibility by workload, monitoring pods and assigning prices to nodes. Its Go-based observers use Kubernetes informers and Prometheus to track pods, nodes and namespaces in real time. Pod and node data is stored in PostgreSQL every 10 minutes, and nodes are enriched with hourly virtual-machine costs from Azure retail pricing APIs.

Users can ask questions in plain English and receive natural-language answers with structured JSON; the site describes these AI queries as read-only. Access options include WhatsApp, Teams, Slack, Power BI, an HTTP API and MCP protocol. A Helm chart deploys the monitor, pricing service, database and optional AI layer. The project is open source under GPL-3.0 and its plan is free. Setup requires a running Kubernetes cluster, Helm 3.x, reachable Prometheus, Metrics Server and a StorageClass. Azure cost enrichment is supported; AWS and GCP support are planned. Natural-language queries through the MCP assistant require an OpenAI API key.

KubeLedger tracks Kubernetes CPU, memory, and GPU usage by namespace for usage accounting and cost allocation. Its accounting includes node overhead and resources that cannot be allocated, while comparisons between usage and requests over time can help identify over-provisioning and plan capacity. Built-in dashboards show hourly CPU and memory trends for seven days, daily accounting for 14 days, and monthly accounting for 12 months. Charts and datasets can be exported as PNG, CSV, or JSON, and integrations include a Prometheus exporter and an official Grafana dashboard. The site says KubeLedger retains a full year of analytics with a fixed footprint under 100 MB. Installation guidance covers Kubernetes and OpenShift through Kustomize or Helm. KubeLedger requires read-only access to Kubernetes API endpoints, including the Metrics API. Internal use is free at free; source code is available, but the product cannot be sold as a managed service. An optional MCP server offers ten read-only tools, but is disabled by default and lacks authentication at the tool layer.