Skip to content

Production install

The development install bundles Postgres, Redis/Valkey, and an object store for convenience. A production install brings those as managed/external services and turns on the hardening dials.

  • Kubernetes v1.31+ with etcd encryption at rest enabled.
  • Knative Serving (for the serving execution model); KEDA if you use queue-depth scaling.
  • A managed PostgreSQL with the pgvector extension, for control-plane state: runs, versions, datasets, cost and knowledge chunks.
  • A TLS certificate for the token service, valid for ctxmesh-token-service.ctxmesh.svc. In production it serves the credential API over TLS only, and does not start without one.
  • At least one model provider key in your secret backend (or use the mock provider first).

The State Layer (Valkey, for session memory) runs in the cluster on a persistent volume by default; to use a managed Redis or Valkey instead, add --set statelayer.externalAddr=<host>:<port>. This release has no setting for an external object store, and the production profile does not deploy the bundled one, so knowledge-base document upload does not work on a production install.

The production settings ship inside the chart, so pull it first:

Terminal window
helm pull oci://ghcr.io/ctxmesh/charts/ctxmesh --version 0.1.0-beta.8 --untar
kubectl create namespace ctxmesh
kubectl -n ctxmesh create secret tls token-service-tls --cert=token-service.crt --key=token-service.key
helm install ctxmesh ./ctxmesh \
--namespace ctxmesh \
--set profile=production \
-f ctxmesh/values-production.yaml \
--set postgres.externalDsn='postgres://user:pass@db.example:5432/ctxmesh?sslmode=require' \
--set bff.mcp.grantHmacKey="$(openssl rand -hex 32)" \
--set bff.mcp.credentialNamespace=ctxmesh-credentials

bff.mcp.grantHmacKey salts the hashed user identities on MCP grants; keep it, because changing it later means everyone reconnects their accounts. bff.mcp.credentialNamespace is the namespace the chart creates to hold those grants, readable only by the credential components.

The profile=production dial flips the platform from “integrated dev stores” to “bring-your-own,” and turns on the availability guards (see High availability). The render fails if the database, the grant key or the credential namespace is missing.

To serve the console on your own domain, put it behind your ingress and set bff.consoleURL to its https:// origin, so MCP sign-ins return to it.

  • The CRDs and the controller (operator).
  • The gateway (model routing + budgets).
  • The console / BFF.
  • A run worker (durable run execution) — required in production.
  1. Verify the control plane is Available with 0 restarts.
  2. Apply your first ModelRoute + SecretBinding.
  3. Bind RBAC personas per namespace.
  4. Wire your trace backend and review the security posture.

High availability · Upgrade & versioning · Helm values · Secrets