How it works

A lightweight collector in your cluster. A control plane that reasons about change.

ChangeGuard installs as a Kubernetes operator that runs a small read-only collector. The collector pushes cluster state over HTTPS to the ChangeGuard control plane, which correlates changes, investigates incidents, scores deploys, and — only within the autonomy you grant — recommends or applies fixes.

Detailed engineering diagrams — deployment, data flow, communication paths, and security boundaries — are in the documentation.

The intelligence architecture

From evidence to judgment

Each layer builds on the one below it. Facts come from your cluster; the platform produces the intelligence — nothing customer-visible is fabricated.

  1. 1 · EvidenceCluster snapshots, change records with diffs, scanner findings, events — the raw facts, continuously collected.
  2. 2 · Change Intelligence GAEvery failing workload correlated to the change that shipped it. The substrate everything else reasons over.
  3. 3 · Opinion EAAn owned position on the incident — belief, confidence, tradeoffs — with every claim tied to evidence.
  4. 4 · Experience EAInstitutional memory: similar past incidents and their outcomes, recalled only when they change the decision.
  5. 5 · Reasoning EAHonest ✓/⚠ factors, why-not-the-obvious-move, and what would change its mind — shown, not hidden.
  6. 6 · Advisor EAThe read a senior engineer would give before the change proceeds — and silence when there’s nothing material.
  7. 7 · Verification GANo fix is called successful until real workload health confirms it. One audited rollback on failure.
  8. 8 · LearningOutcomes — including when advice was ignored and things still worked — feed back into experience.
The flow

From change to decision

1

Collect

A read-only collector reads cluster state and receives your changes from CI/GitOps, pushing them to the control plane over HTTPS.

2

Correlate

Every failing workload is linked to the change that shipped it; the CSC Score reflects current posture.

3

Explain & advise

Incidents get a cited root cause and a proposed fix; pre-flight checks return a safe-to-ship verdict.

4

Act (within policy)

Only at Approve/Auto, and only inside your execution policy, does ChangeGuard apply a fix — then verify it.

Boundaries

What leaves your cluster — and what never does

Never leaves

  • Your source code (repos are read in-cluster for the optional code graph; only the graph is sent)
  • The plaintext values of your Secrets
  • Execution authority — fixes run in your cluster, scoped by RBAC you grant

Sent to the control plane

  • Cluster snapshots (workloads, events, posture)
  • Change metadata and diffs from your CI/GitOps
  • Scanner findings (CVEs, CIS, SBOM) and logs you ingest
Egress is a single direct HTTPS connection to api.changeguard.ai:443, TLS 1.2+. Outbound HTTP proxy is not currently supported — the collector needs direct egress. AI analysis runs in the control plane on Amazon Bedrock; it does not run in your cluster (an optional self-hosted inference mode is on the roadmap).

Install in minutes

One Helm command installs the operator and collector. No PersistentVolumes, no cluster-admin, no changes to your workloads.

Read the install guideStart free