Skip to content

Security and data

This page is for the people who decide whether a workspace may run in their organisation: the security team, data protection, and the administrator who answers their questions. It describes how the workspace separates users and their code, how it knows who someone is, where it keeps data, and which connections leave your network. The other pages go into detail where they are linked.

Separation

A workspace runs in one Kubernetes namespace of its own. On a cluster that hosts other workloads, the workspace's network policies admit traffic from that namespace and from kube-system only, so nothing else on the cluster reaches the workspace's services.

Inside the namespace, the workspace tells two kinds of pods apart: its own services (ShinyProxy, the deploy service, the registry, Redis, the package repository) and the pods that run code your users wrote. Those are IDE sessions, apps, APIs, builds and pipeline runs. Network policies apply to the second kind:

  • Sessions and apps accept connections from ShinyProxy only. An API accepts them on the port of its authentication proxy only.
  • User code may reach the internet, its own namespace and kube-system. It may not reach any private address (10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 100.64.0.0/10, link-local), so it cannot reach the cluster's nodes, other namespaces or your internal network. You open the internal networks your users need, such as a database or a package mirror, with user_networks in velonde.env.
  • On a single host, user code cannot reach the host's own public addresses either, and a firewall rule on the host keeps pods away from link-local addresses, where a cloud provider serves its metadata.

Identity

People log in through your identity provider: OIDC, LDAP or SAML, whichever you configure in the overlay. The user id it returns names the user's home volume and owns the apps they deploy, so it must never change for a person. Your provider's groups decide who is an admin and who may open which app.

The velonde command in a session needs no password or token. The deploy service knows the user from the session the request comes from, which ShinyProxy labelled with the user and their groups when it started it. As a consequence, no credential for the workspace lies in a user's home directory, and nothing that a program or the coding agent could copy out of a session works from anywhere else. The deploy service accepts requests from inside the cluster only.

ShinyProxy places the user's access token of your identity provider in every container it starts. The workspace removes it from sessions and apps, so user code and apps do not hold it.

Deployed APIs and MCP servers identify their callers in three ways, described in Apps, APIs, MCP servers, pipelines: a session or app inside the workspace by the pod it calls from, a caller outside the workspace by a token of your identity provider, and any caller by the API's own key.

User code

IDE sessions, apps, APIs and pipeline runs run as an unprivileged user, without added capabilities and with the container runtime's default seccomp profile. Apps are built on the workspace's own base images. A workspace accepts a user's own Dockerfile only when its admin allows it, and then only from the groups and on the base images the admin names, never on the workspace's IDE images. An app or a pipeline built from a user's own Dockerfile may run as root inside its container, without the capabilities that would let it act on the node's network. An API from a user's own Dockerfile must run as a numbered user other than root. Admission policies in the cluster check every pod that ShinyProxy and the deploy service create. They refuse host namespaces, host paths and privileged containers, and admit added capabilities for builds only.

Builds are the exception. The workspace builds an app's image with BuildKit in a pod with a user namespace of its own. Inside it BuildKit runs as root with the capabilities SYS_ADMIN, NET_ADMIN and SYS_PTRACE and without seccomp and AppArmor profiles, which it needs to build an image. Outside the user namespace these are an unprivileged user's rights. A build step, which runs the user's code, runs in BuildKit's own sandbox on a network of its own inside the pod. On a cluster where such pods must stay apart from other workloads, set build_pool and give builds nodes of their own.

Users and their code never hold a credential for the workspace's registry. Nodes pull images with a credential the cluster keeps. With the workspace's own registry, builds push with an account that may write app images only.

Secrets

The secrets of a workspace (the passwords of Redis and the registry, the client secret of your identity provider, the mail password) live in the overlay's secrets.env on your host and in Kubernetes Secrets in the workspace's namespace. They are never part of a release or of anything we receive.

An app's own secrets come from its .Renviron or .env file or from velonde env set, and the workspace stores them as a Kubernetes Secret of that app. They are not built into the app's image: the upload leaves .Renviron and every .env file out. velonde.yaml holds no secrets.

Where data is kept

What Where How long
Home directories One volume per user, on the cluster's default storage class Until an admin deletes it
Pipeline stores One volume per pipeline Until velonde delete deletes the pipeline
App images The workspace's registry The last ten versions of each app
Logs of apps, sessions and builds Loki, with the monitoring stack 90 days
Audit trail Loki, with the monitoring stack One year
Backups Your S3-compatible bucket, with backups=true As the bucket's lifecycle rules say

Data at rest is encrypted if the storage class and the bucket encrypt it. The workspace adds no encryption of its own. The audit trail records who deployed, changed, deleted or read what.

Connections that leave your network

The workspace sends us nothing of yours. The installer contacts our distribution registry to install or upgrade the workspace and once a day for a renewed license, as described in Accounts and the trial. The cluster pulls the images of third-party components from Docker Hub and public.ecr.aws, listed in Installation on Kubernetes. Builds and sessions reach the package repositories named in velonde.env, or your mirrors of them.

By default, user code (sessions, apps, APIs, builds and pipeline runs) may connect to any address on the internet. With egress=proxy in velonde.env it connects only through a proxy in the workspace's namespace, to the hosts on an allow-list, on ports 80 and 443. The proxy logs every connection with its host name, the allowed and the refused ones. The platform's list holds what installing packages, the coding agents and builds need, and you add to it in the overlay.

Claude Code and Posit Assistant send what the user works on to Anthropic, with the user's own account. Claude Code and Posit Assistant describes what goes where and how to restrict the sign-in to your organisation's accounts. A workspace without AI has neither.

Service meshes

Keep the workspace's namespace out of a service mesh: do not let Istio, Linkerd or a similar mesh inject its proxy into the workspace's pods. The workspace recognises a session by the address its requests come from, and a mesh proxy that forwards them from inside the pod replaces that address with its own. velonde commands and calls to APIs from inside the workspace then fail with "requests must come from a workspace IDE session" or a missing identity.