vmware-harden

VMware Harden (Compliance & Baseline)

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-Harden under the MIT license.

AI-native VMware compliance scanner — built-in CIS / vSphere SCG / 等保 2.0 三级 / PCI-DSS / BSI IT-Grundschutz / EU NIS2 baselines, drift detection, LLM-driven remediation advice, and a web dashboard.

Companion skills: vmware-aiops (inventory + collectors data source; host/VM remediation target), vmware-monitor (read-only inspection), vmware-storage (datastore remediation target), vmware-nsx (segment/gateway evidence), vmware-nsx-security (DFW evidence + remediation target), vmware-aria (metrics correlation), vmware-avi (load balancer evidence), vmware-vks (Tanzu Kubernetes evidence), vmware-pilot (remediation execution with approval gates), vmware-policy (audit log). See references/cross-skill-workflows.md for end-to-end remediation flows that span pilot + sibling skills.

What This Skill Does

Quick Install

uv tool install vmware-harden
vmware-harden baseline list

For first-time use, ensure a vmware-aiops target is configured (harden uses aiops collectors) and optionally set ANTHROPIC_API_KEY for live remediation advice.

When to Use This Skill

Use vmware-harden when the user needs to:

• Run a compliance scan against CIS / vSphere SCG / vSphere 9 STIG-aligned / 等保 2.0 三级 / PCI-DSS / BSI IT-Grundschutz / EU NIS2
• Author or import a custom YAML baseline (e.g., internal corporate baseline)
• Detect drift between two scans of the same target
• Get AI-suggested remediation steps for a violation (advice only — does not execute)
• Browse a web dashboard of compliance posture across multiple targets

Do NOT use this skill when:

• The task is general vCenter/ESXi monitoring or read-only inspection → use vmware-monitor
• The task is VM lifecycle, snapshots, or guest operations → use vmware-aiops
• The user wants to actually execute a remediation (set advanced setting, change DFW rule, etc.) → use vmware-pilot (multi-step approval-gated workflow)
• The task is purely NSX networking/segments → use vmware-nsx
• The user wants continuous, fleet-wide compliance enforcement + automated remediation across the estate → use VCF Operations SPM/ACC (UI). VCF Operations 9.1 Automated Configuration Compliance / Security Posture Management is UI- and schedule-driven (paid Salt engine) and exposes no public compliance API. harden is the complementary, API-scriptable, DuckDB-persisted, cross-target point-in-time scanner for CI and agent workflows; it does not replace SPM/ACC. See references/stig-content-sync.md.

Related Skills — Skill Routing

Common Workflows

1. First-time scan with 等保 2.0 三级

1. Install: uv tool install vmware-harden

2. Verify aiops is configured: vmware-aiops doctor — harden reuses aiops connection for the vCenter collector

3. List baselines: vmware-harden baseline list — confirm dengbao-2.0-level3-vmware is present

4. Scan: vmware-harden scan --baseline dengbao-2.0-level3-vmware --target prod-vcenter

5. Report: vmware-harden report --format json > violations.json (or vmware-harden web for the rendered dashboard). The JSON is an object — {"violations": [...], "coverage": {...}} — read coverage before reporting a result; an empty violations list only means nothing was found among the rules that could be evaluated.

   Failure branch: If you see ConnectError: vmware-aiops target not found, the aiops side is not configured. Run vmware-aiops init first; harden cannot scan without a working collector.

2. Custom baseline import + scan

1. Author YAML under ~/.vmware-harden/baselines/my-corp.yaml (see references for schema)

2. Validate: vmware-harden baseline validate ~/.vmware-harden/baselines/my-corp.yaml

3. Import: vmware-harden baseline import ~/.vmware-harden/baselines/my-corp.yaml

4. Scan: vmware-harden scan --baseline my-corp --target prod-vcenter

   Failure branch: baseline validate failure usually means a check.path references a node type the collectors do not produce (e.g. nsx.gateway.* when no NSX collector ran). See references/cli-reference.md for valid node paths and the baseline schema.

3. Drift investigation

1. Run scan today: vmware-harden scan --target prod-vcenter --baseline cis-vmware-esxi-8.0-subset

2. Run scan again next week (or after a change window): same command

3. View drift: vmware-harden drift (renders the latest snapshot vs its prior snapshot for the same target)

4. Get advice on critical drift: vmware-harden advise --violation-id or vmware-harden advise --all-critical (uses ANTHROPIC_API_KEY; falls back to mock template if unset)

5. Open web view: vmware-harden web --port 8080 then navigate to /drift

   Failure branch: If vmware-harden drift reports No drift detected since previous snapshot, both scans likely ran against the same state. Ensure two scans actually completed against the same --target; the Twin DB at ~/.vmware-harden/twin.duckdb must contain at least two snapshots for that target.

Usage Mode

MCP Tools (8 — 8 read, 0 write)

All 8 tools are read-only with respect to vSphere/NSX. Writes to the local Twin DuckDB are scan-internal and do not modify any VMware resource. Actual remediation execution is intentionally deferred to vmware-pilot (approval-gated).

List results are enveloped. list_baselines, get_baseline_rules, and list_drift_events return {items, returned, limit, total, truncated, hint} rather than a bare list, so completeness is stated rather than inferred — read the rows from items, and treat truncated: true as "there is more, raise limit". Because the twin is a local DuckDB, total is a real count, not an estimate: a page that exactly fills limit is still reported truncated: false when it is genuinely the whole set. list_violations keeps its own older {violations, total, limit, offset, has_more} envelope with the same guarantee.

An empty violation list is not a compliance verdict. A rule can only judge configuration some collector actually gathers; rules whose data is not collected are not executed and are reported as undetermined, never as passing. list_violations and scan_target therefore also return coverage ({evaluated, undetermined, total, tracked, complete, undetermined_rules}) and a one-sentence note. Read it before summarising: report "no violations among the N of M rules that could be evaluated" when complete is false, and never call an estate compliant or clean on a partial scan. When tracked is false the snapshot predates coverage tracking — re-scan rather than assume.

CLI Quick Reference

vmware-harden baseline list
vmware-harden baseline import 
vmware-harden baseline validate 
vmware-harden scan --baseline  --target 
vmware-harden report [--format text|json]
vmware-harden drift [--format text|json]
vmware-harden advise (--violation-id  | --all-critical)
vmware-harden web [--host 127.0.0.1] [--port 8080]

Full CLI reference: see references/cli-reference.md Full capabilities table with response token estimates: see references/capabilities.md

Troubleshooting

"vmware-aiops target not found" / collectors return empty

Harden does not connect to vCenter directly — it relies on vmware-aiops collectors. Run vmware-aiops doctor and confirm the --target name matches an aiops target.

ANTHROPIC_API_KEY not set — advice looks generic

The advisor falls back to a deterministic mock template when no API key is present. Set export ANTHROPIC_API_KEY=... in your shell or in ~/.vmware-harden/.env for live LLM-driven suggestions.

uvx reports "UnknownIssuer" behind a corporate TLS proxy

Don't use uvx for the MCP server in this environment. Use the entry point installed by uv tool install:

{
  "command": "vmware-harden",
  "args": ["mcp"]
}

This avoids uvx re-resolving PyPI through the corporate MitM proxy. The legacy vmware-harden-mcp console script still works and is equivalent. As a workaround, UV_NATIVE_TLS=true lets uv use the system CA store. See CLAUDE.md 踩坑 #25.

"Twin DB not found" on first MCP call

Run at least one scan first: vmware-harden scan --baseline cis-vmware-esxi-8.0-subset --target . The DuckDB file is created on first scan at ~/.vmware-harden/twin.duckdb (override with VMWARE_HARDEN_DB).

等保 baseline reports most rules as not evaluated

Two different causes, and the report distinguishes them. A rule whose attribute no collector produces is recorded undetermined with the reason naming that attribute — see coverage.undetermined_rules; those are collector work, tracked in RELEASE_NOTES. Separately, the 等保 baseline spans several collectors (vCenter advanced settings + NSX DFW), so if only the vCenter collector ran, the DFW rules have no nodes to match. Run a scan with all collectors installed, or pick a baseline whose applies_to matches what you have.

Web dashboard shows 0 violations even after a scan

Verify the dashboard is reading the same DuckDB. If VMWARE_HARDEN_DB is set in your shell but not in the systemd/launchd unit running vmware-harden web, the web server reads the default ~/.vmware-harden/twin.duckdb while your scans wrote elsewhere.

Audit & Safety

1. Source code: github.com/vmware-skills/VMware-Harden — MIT license, publicly auditable.
2. Config / state files: custom baselines in ~/.vmware-harden/baselines/*.yaml; Twin DuckDB at ~/.vmware-harden/twin.duckdb. No passwords are stored — all credentials live in the upstream skill (~/.vmware-aiops/.env).
3. Webhook data scope: none. Harden makes no outbound network calls other than (a) optional Anthropic API requests when ANTHROPIC_API_KEY is set for advisor suggestions, and (b) the local web dashboard bound to 127.0.0.1 by default.
4. TLS verification: harden does not connect to vCenter/NSX directly — TLS handling is delegated to vmware-aiops. The advisor's HTTPS calls to api.anthropic.com use system TLS verification (no opt-out).
5. Prompt injection protection: advisor LLM context is built exclusively from typed Twin queries (rule id, severity, evidence dict) — no free-text user input is forwarded. Evidence text passes through _sanitize() (truncate ≤500 chars, strip C0/C1 control characters).
6. Least privilege: all 8 MCP tools are read-only. Remediation execution is intentionally not exposed — agents that need to apply a fix must invoke vmware-pilot, which provides approval gates and audit logging.

All MCP operations are audited via the @vmware_tool decorator (vmware-policy dependency) to ~/.vmware/audit.db. View with vmware-audit log --last 20.

Environment scoping: policy rules apply per environment, and skills that connect to a VMware estate declare environment: per target in their config.yaml. Harden has no such config — it reads through vmware-aiops and writes only its local Twin DB — so it reports a constant local. scan_target is its only state-changing tool, and the state it changes is the snapshot in that local DB; its vCenter interaction is read-only collection. No harden tool mutates a remote VMware estate, so there is no production change for an environment-scoped rule to protect.

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

License

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