Skip to main content
Short answers to the questions customers actually hit, each linking to the full page. If your question isn’t here, support would rather you ask than guess.

Connecting

For standard Kubernetes API-state evidence, nothing. ChangeGuard Fleet connects centrally using customer-authorized, read-only access — see Connect your EKS fleet. ChangeGuard Edge is an optional in-cluster install for signals that genuinely require local presence (runtime detection, node benchmarks) or environments Fleet cannot reach centrally.
No. The AWS connection authorizes discovery — listing and describing your EKS clusters with a role you create. Kubernetes read access is a separate, per-cluster grant you make explicitly through the generated authorization workflow, and you can revoke it per cluster at any time. See Connect your EKS fleet.
Environments → Connected environments shows every environment with its status, evidence source (Fleet, Edge or both), judgment readiness, execution readiness, latest Permit and mode. Environments → AWS Account Center shows each connected AWS account and whether that connection is healthy, with one guided action per problem. See Connected Environments and AWS Account Center.
Only through a separate, per-environment identity that you create and can revoke: a role in your account with no AWS permissions, mapped to a Kubernetes group that holds get, list and patch on workloads in exactly the namespaces you chose — and even that identity does nothing without a short-lived, single-use Permit for one approved change. Enabling execution approves nothing. See Enable execution.
No. The Cloud Executor acts on a reachable EKS cluster with nothing installed. Edge is an optional evidence collector; it never acts, and it is not a prerequisite for execution or for Auto. A Local Executor exists for topologies or policies that require local presence. See Fleet, Edge and the Executor.
Fleet reverses the direction: ChangeGuard AI connects to your Kubernetes API endpoint centrally, so the cluster itself needs no outbound path to ChangeGuard AI for standard evidence. The direct-egress requirement below applies only to the optional Edge install.

Installing ChangeGuard Edge (optional)

Both CG_API_KEY and CG_CLUSTER_NAME are required — the installer refuses to run without them, and the Helm chart refuses to render without agent.clusterName. The correct commands are on Install ChangeGuard AI.
Your cluster connects within about a minute of install (up to ~2 minutes on a first-ever install while images pull), and the first CSC Score follows within ~10 seconds of connecting. If it’s been longer, the fix table covers the three causes — it’s almost always egress or the API key.
Not today. The cluster needs direct outbound HTTPS to api.changeguard.ai:443. This is tracked on Known Limitations & Roadmap.
Yes — mirror the images and repoint every one of them with --set global.imageRegistry=<your-registry>. The registry list is on Before you install.
One component: the Falco DaemonSet (runtime syscall detection), on by default with security scanning. Everything else runs unprivileged and read-only. What it mounts, why, and how to disable it: Permissions & RBAC.
Helm uninstall plus one manual step (deleting the CRD) — exact commands on Install ChangeGuard AI. Removing already-sent data is per-cluster in-app; full-tenant erasure goes through support.

Using ChangeGuard AI

The one verdict vocabulary, everywhere: SHIP = sufficient evidence supports proceeding. HOLD = ChangeGuard AI cannot establish that proceeding is safe enough — including when important evidence is missing or too stale to trust (it says so rather than pretending the evidence is available). BLOCK = affirmative evidence indicates the change should not proceed. A verdict is a judgment, not an authorization: it grants nothing and, in the product, blocks nothing. In CI, the shipped templates fail the job on BLOCK by default and never on HOLD — see Use the Verdict in CI.
A real score of 0 renders as a red 0/100; “Calculating your first score” means the first snapshot hasn’t arrived yet. If it turns into “Score unavailable,” the collector isn’t pushing — use Retry and the connection checks.
Since the November 2026 release they are generally available on the Professional and Enterprise plans, on by default, and nothing is in Early Access. If you don’t see them, check three things in order: your plan (the 14-day evaluation and Starter don’t carry them, and the product says so in plain words), Settings → Engineering assistance (an admin may have turned a capability off, and the Opinion must be on for Reasoning and Experience), and, for the Advisor, whether the environment has live evidence (it stays silent without Fleet or Edge evidence and reports why in advice_status). Details: Engineering intelligence and Engineering Advisor.
Defaults are requests 25m/64Mi, limits 500m/512Mi (chart ≥ 5.3.6). Override any subset: --set agent.collector.resources.limits.memory=1Gi — see Operations.
Create the new key in-app, update the changeguard-credentials Secret, restart the collector, then revoke the old key — steps on Operations. There’s no self-serve rotation API yet; after a suspected exposure, also confirm invalidation with support.
Only where you ask it to. A verdict authorizes nothing, and a judgment in the product blocks nothing. In CI, the shipped templates fail the job on BLOCK by default and never on HOLD; set enforce_block: false to run purely advisory. Execution is the other direction — ChangeGuard AI applying a fix — and that happens only in environments where you enabled an Executor, under a Permit, after a person approves (Advise, the default) or your policy allows (Auto). See Advise and Auto and ChangeGuard Permit.

Troubleshooting

The two canonical fix tables:
  • Install and first connection — Connect your first cluster: repeated 401s, egress timeouts, an environment never reaching Ready in Connected environments, operator CrashLoopBackOff.
  • Day-2 operations — Operations handbook: collector crash loops, stale clusters, missing features, version skew.
Still stuck? support@changeguard.ai — include the collector logs (kubectl logs -n changeguard-agent deploy/changeguard-collector --tail=50); they contain no secrets.