How we compare
An honest comparison with OpenCost, Kubecost, Grafana Loki, Grafana Pyroscope, Datadog and Cast AI — where they win, where KubeHero differs, with sources.
Every tool on this page does part of what KubeHero does, and several do their part better or with more polish. KubeHero's claim is narrower and checkable: all of these signals, in one open install you run yourself, with automation a human arms. If a cell below is wrong or out of date, open an issue and we'll fix it.
Checked 2026-09-28
Based on each vendor's public documentation on that date. ✓ = built in · ◐ = partial (see the note) · ✗ = not offered · — = nothing to guard (the tool doesn't automate changes).
The matrix
| KubeHero | OpenCost | Kubecost | Grafana Loki | Grafana Pyroscope | Datadog | Cast AI | |
|---|---|---|---|---|---|---|---|
| Cost allocation (idle, shared, labels) | ✓ | ✓ | ✓ | ✗ | ✗ | ✓ | ✓ |
| Logs (store + LogQL-class query) | ✓ | ✗ | ✗ | ✓ | ✗ | ✓ | ✗ |
| Continuous profiling (incl. no-code eBPF) | ✓ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ |
| Network cost per workload (cross-zone, egress) | ✓ eBPF | ◐ ³ | ✓ ³ | ✗ | ✗ | ✓ ⁵ | ✓ ⁶ |
| Rightsizing that applies | ✓ | ✗ | ✓ ² | ✗ | ✗ | ✓ ⁵ | ✓ |
| Human-armed guardrails (armed · bounded · audited · undo) | ✓ | — | ◐ ² | — | — | ◐ ⁵ | ◐ ⁶ |
| Built-in AI investigation with cited evidence | ✓ | ✗ | ✗ | ✗ ⁴ | ✗ ⁴ | ✓ ⁵ | ✗ |
| MCP server | ✓ | ✓ ¹ | ✓ ² | ◐ ⁴ | ◐ ⁴ | ✓ ⁵ | ✓ ⁶ |
| Self-hosted, data stays in your cluster | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✗ |
| License | Apache-2.0 + BSL-1.1 ⁷ | Apache-2.0 | proprietary (free tier) | AGPL-3.0 | AGPL-3.0 | proprietary | proprietary |
| Price | free, self-hosted | free | free tier + enterprise | free OSS; Grafana Cloud paid | free OSS; Grafana Cloud paid | usage-based SaaS | quote-based SaaS |
- OpenCost ships a built-in MCP server from v1.118 — docs.
- Kubecost 3.0 added automated container request sizing (the free tier covers up to 250 cores) with suspendable, audited resize runs; an MCP server ships by default from v3.3 — Kubecost 3.0, mcp-kubecost.
- OpenCost and Kubecost network costs come from a separate network-costs DaemonSet that classifies egress as cross-zone, cross-region or internet; there's no service map — network cost configuration.
- Loki and Pyroscope are AGPL-3.0. MCP access goes through the separate mcp-grafana server; Grafana's AI assistant is a Grafana Cloud feature, not part of the OSS projects.
- Datadog: Kubernetes Autoscaling applies workload rightsizing (in-app or via GitOps); data-transfer cost allocation comes from Cloud Network Monitoring (eBPF) plus Cloud Cost Management; Bits AI SRE and the MCP server are generally available.
- Cast AI: network cost monitoring (egress, cross-AZ), policy-driven workload autoscaling, and an MCP server; a SaaS control plane with an in-cluster agent.
- The CLI, collector, cost model and protos are Apache-2.0; the control plane, operator, advisor, pricing engine and dashboard are BSL-1.1 — source-available and free to self-host, but not OSI-approved open source.
Tool by tool
OpenCost
What it gets right. The CNCF reference model for Kubernetes cost allocation: free, open, vendor-neutral, and now with a built-in MCP server. Kubecost's core is built on it.
Where KubeHero differs. OpenCost is an allocation engine and API; KubeHero implements the same allocation API (/allocation/compute) and adds logs, profiles, an eBPF service map, alerting, rightsizing that applies and an investigation agent on the same data.
Pick OpenCost if you want only the allocation model and will build the rest. Pick KubeHero if you want allocation joined to the signals that explain it. Your OpenCost-based scripts keep working — point them at KubeHero.
Kubecost
What it gets right. A mature, polished FinOps product for Kubernetes: allocation, network costs, savings insights, and — since 3.0 — automated request sizing with audited runs, plus an MCP server. Strong enterprise features.
Where KubeHero differs. Kubecost is cost-first; it doesn't store logs or profiles. KubeHero's automation is armed per policy by a human and enforced by a Kubernetes operator from CRDs you can review in Git. And KubeHero is free to self-host at any size.
Pick Kubecost if cost and savings reporting is the whole job and you value its maturity. Pick KubeHero if you want cost, logs, profiles and network in one place.
Grafana Loki
What it gets right. Battle-tested, horizontally scalable log storage on object storage, with LogQL and the Grafana ecosystem around it.
Where KubeHero differs. KubeHero implements LogQL and Loki's push and query APIs on ClickHouse, next to cost and profiles, and prices log volume per team. It supports the common LogQL surface, not all of it — see the support matrix. For very large log volumes, Loki's object-storage architecture is purpose-built.
Pick Loki if logs are the main job at very large scale. Pick KubeHero if you want logs correlated with cost, profiles and the network — and Grafana keeps working via the Loki data source.
Grafana Pyroscope
What it gets right. A dedicated continuous-profiling database with broad SDK support, eBPF profiling through Grafana Alloy, and excellent flamegraph tooling.
Where KubeHero differs. KubeHero's eBPF profiler is built into the collector you already run, it accepts Pyroscope's /ingest and scrapes with Pyroscope's annotations, and it prices every function with the workload's real CPU spend.
Pick Pyroscope if you want the deepest profiling feature set. Pick KubeHero if you want profiles with dollar figures, next to everything else.
Datadog (Kubernetes + Cloud Cost Management)
What it gets right. The broadest commercial platform here: metrics, logs, APM and profiling, network monitoring, cost management with data-transfer allocation, Kubernetes autoscaling that applies recommendations, a GA AI SRE agent and MCP server — managed for you.
Where KubeHero differs. KubeHero is self-hosted and free: your telemetry never leaves your infrastructure, which matters for regulated and air-gapped environments and for the size of the bill. KubeHero doesn't do APM or traces.
Pick Datadog if you want a managed, all-in-one SaaS including APM and budget for it. Pick KubeHero if data residency, cost or open source matter more.
Cast AI
What it gets right. Aggressive, automated optimization — node selection, spot automation, workload autoscaling — with cost and network cost monitoring and an MCP server.
Where KubeHero differs. Cast AI's strength is autonomous optimization run from its SaaS control plane. KubeHero is observability first, self-hosted, and its automation is deliberately conservative: human-armed, bounded, reversible, and expressed as CRDs in your cluster.
Pick Cast AI if you want hands-off optimization and are comfortable with a SaaS control plane. Pick KubeHero if you want every change to be yours to arm and undo.
What KubeHero doesn't replace
- Your metrics TSDB — keep Prometheus or Mimir; KubeHero exports to it.
- Your tracing backend — v0.3 doesn't store traces.
- Your cloud bill — allocation is an estimate reconciled to node prices; reconcile to the invoice in your FinOps tool (the FOCUS export helps).