This page is what happened when we ran it. The project itself — code, maintenance, community — has its own review: full repo review →
Kubernetes
The Kubernetes MCP server, mcp-server-kubernetes by Flux159, hands an AI agent your kubectl. It is a community project, not a CNCF or vendor release, and it wraps get, apply, patch, delete, logs, describe, scale, rollout, exec, port-forward and Helm install into MCP tools that run against whatever kubeconfig context the process can see.
How we started it
npx -y mcp-server-kubernetesUse it if you want an agent doing read-heavy cluster triage, and only if you have pointed it at a context where the worst possible command is survivable. The tools are thorough and the descriptions are honest about what they do, which is exactly the problem: kubectl_delete, kubectl_patch and exec_in_pod are one confident hallucination away from a bad afternoon. It boots clean with no env vars and no auth dance, so the only thing standing between the model and production is your kubeconfig hygiene. That is a thin wall.
Use it if
Skip it if
The bill for your context window
| Tool | Est. tokens | Described? | Required params |
|---|---|---|---|
| kubectl_create | ~995 | yes (82 chars) | 0 |
| node_management | ~383 | yes (67 chars) | 1 |
| kubectl_delete | ~380 | yes (83 chars) | 0 |
| kubectl_logs | ~354 | yes (66 chars) | 3 |
| install_helm_chart | ~339 | yes (83 chars) | 3 |
| kubectl_patch | ~324 | yes (90 chars) | 2 |
| kubectl_get | ~322 | yes (81 chars) | 1 |
| kubectl_rollout | ~320 | yes (75 chars) | 4 |
Its tools/list is 23,854 bytes, about 5,964 tokens by the chars/4 estimate, for 23 tools. That is heavy for a directory whose median server lands near two thousand, and the weight is distributed oddly: kubectl_create alone accounts for roughly a sixth of the payload because its schema enumerates resource kinds, while cleanup, ping and stop_port_forward cost almost nothing. The description text is not the culprit; most tool descriptions are one line. The JSON schemas are. If your client lets you disable tools, dropping the Helm trio and kubectl_create buys back real room without touching the triage workflow you probably came for.
Setup reality
There is nothing to configure, which is the whole story. We started it with npx and it printed "Starting Kubernetes MCP server v4.1.4, handling commands..." with telemetry disabled and no environment variables requested; the harness recorded zero env hints. Auth is your existing kubeconfig, so the server is exactly as privileged as the shell that launched it. First boot took 26,252 ms because npx was resolving the package cold, and the install log is noisy with deprecation warnings from transitive deps. Restarts are far quicker. If you want safety, the work is outside the server: a dedicated context, a service account with narrow RBAC, and kubectl_context pinned before anything else runs.
Questions people ask
Does the Kubernetes MCP server need any API keys or environment variables?
No. We booted it with a clean environment and it recorded no env hints; it authenticates by reading your kubeconfig the same way kubectl does. That also means it inherits your permissions, so the security question is which context you leave selected, not which secret you set.
Can I make it read-only?
Not from the server side in the build we measured; all 23 tools are listed unconditionally, including delete, patch and exec_in_pod. The README describes a non-destructive mode via environment flag, so check the current docs. Otherwise, enforce it with RBAC on a dedicated service account.
Is that startup time the server's fault?
That was npm, not the server. Our first run resolved and installed the package from scratch under npx, which dominated the 26,252 ms we recorded; the server itself announces readiness immediately after. A warm cache or a global install makes startup unremarkable.
What does it actually expose besides tools?
Alongside the tools it advertises 5 resources and 1 prompt, which is more than most servers in this directory bother with. The resources cover cluster objects your client can read directly rather than through a tool call, which is cheaper when it works.
Other cloud servers we started
How this page is made: the server is spawned as an unprivileged user with a clean environment and no credentials, then asked for its tools, resources and prompts over stdio. Token figures are estimates at four characters per token, not a tokenizer count. One run, one machine. Corrections: contact the desk.