It runs in your account. All of it.
Not a hosted control plane with an agent phoning home — the build plane, the workspaces, the registry reads, and the proxy all run on your own cluster, in your own AWS account, reading the Prometheus and ECR you already have. Keyless by default: the pod's IRSA role, not a static credential in a secret store.
Some estates can't rent their build plane.
Data residency, procurement, an existing account structure, or simply not wanting a second copy of the source.
The usual answer is a hosted control plane with an agent inside your network — which moves the boundary rather than respecting it. Metadata still leaves, the vendor still holds the state, and the interesting question becomes what exactly is in the telemetry.
The alternative is to make self-hosting the normal deployment rather than the enterprise exception. The platform is a set of Kubernetes resources and it runs where your cluster runs, reading the systems you already operate — your ECR, your Prometheus, your forge — through the pod's own IAM role.
Specifically, and without hedging.
Worth being precise, because the phrase is used loosely.
What runs in your account
- Remote build execution: your Buildbarn cluster, your workers, your cache — the build plane is not shared with anyone.
- Workspaces: DevWorkspace CRs and their persistent volumes on your cluster, reached through your own wildcard domain and session.
- Registry reads: your ECR, read through the pod's IRSA role. There is no static registry credential anywhere in the deployment.
- Metrics: read from the Prometheus your remote execution already exposes. Nothing is shipped out for a vendor dashboard, and with no Prometheus the panels are empty rather than fabricated.
- Your forge: one contract covers GitHub and self-hosted GitLab, so keeping source in-house doesn't cost a second integration.
Where the network edge actually is.
Two capabilities reach outside by design. Naming them is more useful than a blanket claim.
The chat plane runs its model turn on Amazon Bedrock, through your account and the pod's IRSA role rather than a static provider key — but it is a call to a managed model service, and it leaves your VPC. The grounding tool queries Wikidata, which is a public knowledge base, read-only and cached by query hash.
Everything else — build execution, workspaces, registry reads, metrics, the proxy — stays inside. If your constraint is strict egress control, those two are the surfaces to write policy about, and both are optional plugins rather than load-bearing platform components. We would rather name them than claim an air gap the deployment doesn't have.
Your account, your cluster, your forge.
The alternative runs in a vendor's cloud and reaches into yours.
Keyless by default: the pod's own IRSA role reads your ECR. There is no static registry credential in the deployment to leak or rotate.
Backed by shipping plugins.
Each claim on this page is anchored to a plugin that provides it. The site build resolves these cross-references against the live feature catalog and fails if one of them doesn't exist — so this list can't quietly outlive what the platform ships.
Dev environments are DevWorkspace CRs on your own cluster rather than a vendor's managed service — inspectable, template-driven, and wired to your remote execution.
plugin detail →Reads your own ECR keyless through the pod's IRSA role — repos, tags, image config and the real layer filesystem, with no static registry credential to store or rotate.
plugin detail →Reads the Prometheus your remote execution already exposes rather than shipping metrics anywhere; with no Prometheus the panels go empty instead of inventing a number.
plugin detail →Your wildcard and your session gate the whole workspace fleet from one Ingress — no per-workspace DNS record or certificate is ever minted.
plugin detail →One contract spans GitHub and self-hosted GitLab, so an estate that keeps its forge in-house isn't a second integration.
plugin detail →Adjacent problems.
If the source is the asset, then every tool that touches it is a disclosure decision — including the ones you turned on to go faster.
Read more →When every change has to survive an auditor, evidence collection stops being a project and starts being a property of how work already happens.
Read more →Prove it's safe to merge.
Hosted, or launched into your own AWS account. Start free — no card required.