VMware VKS

Disclaimer: This is a community-maintained open-source project and is not affiliated with, endorsed by, or sponsored by VMware, Inc. or Broadcom Inc. "VMware" and "vSphere" are trademarks of Broadcom. Source code is publicly auditable at github.com/vmware-skills/VMware-VKS under the MIT license.

AI-powered VMware vSphere Kubernetes Service (VKS) management — 23 MCP tools.

Requires vSphere 8.x+ with Workload Management enabled. Companion skills: vmware-aiops (VM lifecycle), vmware-monitor (monitoring), vmware-storage (storage), vmware-nsx (NSX networking), vmware-nsx-security (DFW/firewall), vmware-aria (metrics/alerts/capacity), vmware-avi (AVI/ALB/AKO), vmware-harden (compliance baselines). | vmware-pilot (workflow orchestration) | vmware-policy (audit/policy)

What This Skill Does

Quick Install

uv tool install vmware-vks==3.1.0
vmware-vks check

When to Use This Skill

• Check if vSphere environment supports VKS
• Create, update, or delete Supervisor Namespaces with resource quotas
• Deploy, scale, upgrade, or delete TKC (TanzuKubernetesCluster) clusters
• Get kubeconfig for Supervisor or TKC clusters
• Check Harbor registry info or storage usage

Use companion skills for:

• VM lifecycle, deployment → vmware-aiops
• Inventory, health, alarms → vmware-monitor
• iSCSI, vSAN, datastore → vmware-storage
• Load balancing, AVI/ALB, AKO, Ingress → vmware-avi

Related Skills — Skill Routing

Common Workflows

Deploy a New TKC Cluster

Pre-flight (judgment):

• Supervisor must be vSphere 8.x+ with WCP enabled — supervisor check returns pass/fail. If fail, no amount of TKC commands will work; resolve at vSphere/WCP layer first.
• K8s version: pick a TKR version that's still supported by VMware (not EOL). New clusters on EOL versions look fine until you need a CVE patch and there isn't one.
• VM class sizing: best-effort-* for dev, guaranteed-* for prod. A best-effort worker can be evicted under host pressure — production workloads need guaranteed.
• Storage policy: must already exist in vCenter. list_supervisor_storage_policies first and pass the returned policy ID (not the display name); creating a TKC against a missing policy fails after CP boot, leaving partial state.
• Control-plane count: 1 for dev, 3 for prod (HA). Cannot upgrade from 1→3 without recreating; choose right the first time.
• Namespace quota: TKC consumes CP + worker × (cpu, memory) from namespace quota. If quota is too tight, workers fail to schedule with no obvious error.
• TKC API version: auto-detected at runtime via the K8s discovery API (prefers cluster.x-k8s.io/v1 when the Supervisor serves it, falls back to v1beta1 on vSphere 8.0). No manual selection needed; advanced callers can override via the api_version parameter on generate_tkc_yaml().

Steps:

1. vmware-vks supervisor check --target prod → must pass
2. vmware-vks tkc versions -n <ns> → pick a non-EOL TKR
3. (If new namespace) vmware-vks namespace create dev --storage-policy <policy> --cpu <enough-for-cp+workers> (a dry run by default), then the same with --apply
4. vmware-vks tkc create dev-cluster -n dev --version <tkr> --control-plane 1 --workers 3 --vm-class best-effort-large (a dry run by default), then the same with --apply
5. Wait for phase=running (typically 10-15 min); do not assume success on apply return
6. Only if the user asked for cluster access: vmware-vks kubeconfig get dev-cluster -n dev -o ./kubeconfig — writes an owner-only file; report the path, never paste the token into the agent context

Scale Workers for Load Testing

Judgment: scaling is fast but reverse-scaling is destructive — workers are deleted, in-flight pods lost. Treat scale-down like a delete.

1. tkc get dev-cluster -n dev → record current worker count and any pending pods
2. Scale-up: tkc scale dev-cluster -n dev --workers 6 → safe, additive operation
3. Verify new workers reach Ready in kubectl get nodes before sending traffic
4. Scale-down: drain pods first via kubectl drain on the to-be-deleted nodes, THEN tkc scale --workers 3. Skipping drain causes pod restarts on remaining nodes — measurable user impact.
5. Confirm namespace quota leftover supports the new size — quota is enforced at scheduling, not at scale request

Namespace Resource Management

Judgment: quota changes are atomic but consequences are not. Reducing quota below current usage doesn't evict pods — they keep running, but no new pods schedule, looking like a "namespace is broken" symptom.

1. namespace list → see all namespaces and their phase
2. storage -n dev → check current CPU/memory/storage usage; never reduce quota below current usage + 20% headroom
3. namespace update dev --cpu <new> --memory <new> --dry-run → preview, then real
4. Validate by attempting a small pod scale-up; if it pends with Insufficient cpu, quota is still the bottleneck

Architecture

User (Natural Language)
  ↓
AI Agent (Claude Code / Goose / Cursor)
  ↓ reads SKILL.md
  ↓
vmware-vks CLI  ─── or ───  vmware-vks MCP Server (stdio)
  │
  ├─ Layer 1: pyVmomi → vCenter REST API
  │   Supervisor status, storage policies, Namespace CRUD, VM classes, Harbor
  │
  └─ Layer 2: kubernetes client → Supervisor K8s API endpoint
      TKC CR apply / get / delete  (cluster.x-k8s.io/v1beta1)
      Kubeconfig bearer token from POST /wcp/login (Supervisor JWT)
  ↓
vCenter Server 8.x+ (Workload Management enabled)
  ↓
Supervisor Cluster → vSphere Namespaces → TanzuKubernetesCluster

Usage Mode

MCP Tools (23 — 14 read, 9 write)

All accept optional target parameter to specify a named vCenter.

list_namespaces, list_supervisor_storage_policies and list_vm_classes return the family list envelope — {items, returned, limit, total, truncated, hint} — rather than a bare array. Read the rows from items; truncated says whether the listing is complete, so it never has to be guessed from the row count. These three read their collection in one un-paged call, so total is the real count and truncated is always false.

create_namespace, create_tkc_cluster, delete_namespace, delete_tkc_cluster take confirm (default False). Without it they change nothing and return blast_radius: the spec or YAML plan to be created, or what a delete destroys (TKC clusters, VMs and PVCs in the namespace; node counts and running workloads of a cluster). Show it to the user; pass confirm=True only after they decide on what they saw.

confirm=True is refused while a blocker stands — TKC clusters inside a namespace, running workloads in a cluster (unless force=True), a name already taken — or when any part of the blast radius could not be read. confirmed and dry_run are deprecated aliases until the next minor release.

Credential access: issue_supervisor_kubeconfig and issue_tkc_kubeconfig are not reads. The kubeconfig embeds a Supervisor bearer token (JWT from /wcp/login) that acts as the configured vCenter account until it expires — typically hours, and not revoked when vmware-vks exits. So:

• Run them only when the user explicitly asks for a kubeconfig; never auto-run them or fetch one as a side step. Both are annotated readOnlyHint: false (MCP clients ask first) and risk_level: medium (so an operator's min_risk_level: medium deny rule covers them).
• Always pass output_path (MCP) or -o <path> (CLI) and report only the path. The file is created owner-only (0600) — also when it replaces an existing file — and a symlink target is refused. Delete it when no longer needed.
• The audit row records who fetched which kubeconfig, but the returned kubeconfig is redacted (sensitive_result=True).

Full capability details and safety features: see references/capabilities.md

CLI Quick Reference

# Supervisor
vmware-vks check [--config <path>]
vmware-vks preflight-auth [--target <name>]   # live-validate POST /wcp/login (issue #13)
vmware-vks supervisor status <cluster-id> [--target <name>]
vmware-vks supervisor storage-policies [--target <name>]

# Namespace
vmware-vks namespace list [--target <name>]
vmware-vks namespace get <name> [--target <name>]
vmware-vks namespace create <name> --cluster <id> [--cpu <n>] [--memory <mb>] [--storage-policy <name>] [--apply]
vmware-vks namespace update <name> [--cpu <n>] [--memory <mb>] [--target <name>]
vmware-vks namespace delete <name> [--target <name>]

# TKC Clusters
vmware-vks tkc list [-n <namespace>] [--target <name>]
vmware-vks tkc create <name> -n <ns> [--version <v>] [--workers <n>] [--vm-class <name>] [--apply]
vmware-vks tkc scale <name> -n <ns> --workers <n> [--pool <name>] [--target <name>]
vmware-vks tkc upgrade <name> -n <ns> --version <v> [--target <name>]
vmware-vks tkc delete <name> -n <ns> [--skip-workload-check] [--target <name>]

# Kubeconfig
vmware-vks kubeconfig supervisor -n <namespace> [-o <path>] [--target <name>]
vmware-vks kubeconfig get <cluster-name> -n <namespace> [-o <path>] [--target <name>]

# Harbor & Storage
vmware-vks harbor [--target <name>]
vmware-vks storage -n <namespace> [--target <name>]

Full CLI reference with all flags and interactive creation: see references/cli-reference.md

Troubleshooting

"VKS not compatible" error

Workload Management must be enabled in vCenter. Check: vCenter UI → Workload Management. Requires vSphere 8.x+ with Enterprise Plus or VCF license.

Namespace creation fails with "storage policy not found"

List policies first: vmware-vks supervisor storage-policies, then pass the Policy ID column value (not the display name) as --storage-policy.

TKC cluster stuck in "Creating" phase

Check Supervisor events in vCenter. Common causes: insufficient resources on ESXi hosts, network issues with NSX-T, or storage policy not available on target datastore.

Every REST tool returns 401

vCenter keeps two independent session stores, and this skill uses both. Namespace, storage-policy and Supervisor-status tools call the vSphere Automation REST API under /api, which authenticates with a session id from POST https://<vcenter>/api/session (HTTP Basic on that one call, then the id in a vmware-api-session-id header). The pyVmomi SOAP session under /sdk is a different store and its key is rejected there — sending it produced a 401 on every REST tool. A 401 is refreshed automatically once; if it persists, check whether a proxy between you and vCenter strips the vmware-api-session-id header. A 403, not a 401, is what an account short of Workload Management permissions gets.

Validating Supervisor auth (POST /wcp/login)

Supervisor/TKC Kubernetes auth uses a JWT obtained from POST https://<supervisor>/wcp/login (HTTP Basic → JSON session_id bearer token), not the pyVmomi SOAP session key, and not the /api/session id above either — three separate credentials. To validate this end-to-end against your real Supervisor, run:

vmware-vks preflight-auth [--target <name>]

It performs the real login (no mocks) and reports, per target: vCenter reachable → /wcp/login HTTP status → parseable session_id → does the JWT authenticate a trivial Supervisor K8s API call. A healthy result is all four steps ✓ PASS ending in target '<name>': /wcp/login auth flow validated end-to-end. (exit code 0). The login goes to the Supervisor API endpoint vCenter publishes, or to supervisor_host when that target sets it (vCenter itself answers 404 for /wcp/login on 9.1). On failure each step prints a teaching message — e.g. a 404 on /wcp/login means the wrong host (set supervisor_host) or an endpoint path that differs on your Supervisor version, a 401 on the K8s probe means session_id is not the bearer token on your version. It never tracebacks — every failure is status output.

Kubeconfig retrieval fails

Supervisor API endpoint must be reachable from the machine running vmware-vks. Check firewall rules for port 6443.

Scale operation has no effect

Verify the cluster is in "Running" phase before scaling. Clusters in "Creating" or "Updating" phase reject scale operations.

Delete namespace rejected unexpectedly

The namespace delete guard prevents deletion when TKC clusters exist inside. Delete all TKC clusters in the namespace first, then retry.

Prerequisites

• vSphere 8.x+ with Workload Management enabled
• Enterprise Plus or VCF license
• NSX-T (recommended) or VDS + HAProxy networking
• Supervisor Cluster configured and running

Setup

uv tool install vmware-vks==3.1.0
mkdir -p ~/.vmware-vks
vmware-vks init

All tools are automatically audited via vmware-policy. Audit logs: vmware-audit log --last 20

Full setup guide, security details, and AI platform compatibility: see references/setup-guide.md

Audit & Safety

Operations are audited via vmware-policy:

• Every MCP tool call, and every CLI command that reaches vCenter or the Supervisor (@guarded writes and credential reads, @audited reads), is logged to ~/.vmware/audit.db (SQLite). The seven namespace/TKC write operations are also mirrored to ~/.vmware-vks/audit.log (JSON Lines)
• Policy rules enforced via ~/.vmware/rules.yaml (deny rules, maintenance windows, risk levels)
• Risk classification: each tool tagged as low/medium/high/critical
• View recent operations: vmware-audit log --last 20
• View denied operations: vmware-audit log --status denied

Local files (sensitive, keep owner-only): ~/.vmware-vks/.env holds per-target passwords (b64-obfuscated, not encrypted; created 0600, vmware-vks doctor flags a wider mode). ~/.vmware/audit.db (+ WAL/SHM, 0600 in a 0700 directory) keeps operation history — resource names, parameters, results with credentials redacted — and rotates at 100 MB, keeping 5 archives. ~/.vmware-vks/audit.log (0600) is never rotated or pruned. Exported kubeconfigs are 0600 and are never cleaned up by the skill. Delete or rotate these yourself per your retention policy.

In-memory kubeconfig (v1.5.18+): for its own API calls the skill builds the Supervisor/TKC kubeconfig as a Python dict and loads it with load_kube_config_from_dict(); the bearer token (also cached in process memory, up to 8 h) is never written to a temp file. Only an explicit export (output_path / -o) puts it on disk.

vmware-policy is automatically installed as a dependency — no manual setup needed.

License

MIT — github.com/vmware-skills/VMware-VKS