Access Tiers
Status: Available.
msh local classifies every intercepted request into one of four risk tiers. The classification happens in real-time with sub-millisecond latency and drives the policy engine's allow / escalate / deny decision.
OBSERVE
Read-only operations.
HTTP Methods: GET, HEAD, OPTIONS
Examples across supported CLIs:
kubectl get pods·kubectl describe node·kubectl logs deploy/apiaws ec2 describe-instances·gcloud compute instances listgh repo view·terraform show·helm list
The default policy allows observe-tier requests without a prompt. Operators can override this per-namespace, per-project, or per-app in policy.yaml.
MODIFY
Standard write operations.
HTTP Methods: POST, PUT, PATCH
Examples:
kubectl apply -f manifest.yaml·kubectl patch deploymentaws s3 cp·gcloud compute instances createfly deploy·terraform apply
Modify-tier requests escalate by default and block until a human approves them via the desktop app or msh local approval allow <id>.
SENSITIVE
Irreversible or high-impact operations.
HTTP Methods: DELETE, plus any request that touches a high-value resource regardless of method. For kubectl, reading Secrets is also elevated to sensitive because Secrets carry credentials.
Examples:
kubectl delete pod/x·kubectl get secretaws s3 rm·gcloud projects deletefly apps destroy
Sensitive-tier requests escalate by default.
BREAK-GLASS
Highest-risk operations. The default policy escalates the tier for human approval, while specific catastrophic operations are hard-denied. Operators can set other rules in policy.yaml.
For kubectl: dangerous subresources (exec, attach, portforward, proxy, debug) and mutations on privilege-escalation resources (RBAC roles/bindings, webhooks, CRDs, namespaces, PersistentVolumes).
For other CLIs: IAM mutations, credential issuance, infrastructure teardown.
Examples:
kubectl exec -it pod -- sh·kubectl port-forwardaws iam create-user·gcloud iam service-accounts createfly machines api-proxy
Fail-closed behaviour
Unknown commands, unrecognised CLI tools, and classifiers that cannot determine a tier all fail closed — the request is treated as escalate or deny, never silently allowed. This prevents a novel or unexpected command from bypassing enforcement by accident.
The generic: block in policy.yaml covers any host on the inspect list that has no dedicated classifier. Its tier assignment starts with the HTTP method: GET/HEAD/OPTIONS → observe, POST/PUT/PATCH → modify,DELETE → sensitive. Paths containing destructive-looking prefixes (delete, destroy, drop, purge, terminate, evict) are escalated to at least sensitive regardless of method.
Supported CLI classifiers
The following CLI tools have dedicated classifiers with command-level tier resolution:
| CLI tool | Inspect protocol key | CLI aliases recognised |
|---|---|---|
| kubectl | kubectl | kubectl, kustomize |
| Google Cloud CLI | gcloud | gcloud |
| AWS CLI | aws | aws |
| GitHub CLI | github | gh |
| Terraform / OpenTofu | (CLI only) | terraform, tofu |
| Helm | (CLI only) | helm |
| Cloudflare Wrangler | cloudflare | wrangler, flarectl |
| Fly.io | fly | flyctl, fly |
Any host not matched by a dedicated classifier falls through to the generic: block (HTTP-method tier, raised to sensitive for destructive path segments).
Use
msh local classify cli <argv>ormsh local classify http <method> <url>to see exactly which tier the daemon would assign to any specific request before putting it in front of a real workload.