Skip to main content

Enterprise Tier

Project and data plane management

Isolated projects are created here, the data planes that serve them are deployed, each data plane is labelled with the projects it serves, and projects are moved between data planes without disruption. Models must be enabled in the organization catalog before they can be granted to a project; that catalog work is covered in Provision models and providers.


Three objects underpin these guides; the full treatment is in Key concepts and Planes and core components.

  • A project is the logical boundary that owns models, API keys, Model Context Protocol (MCP) servers and profiles, routing and policy, budgets, and members. It is also the isolation boundary: each request resolves to its project from its API key, and a project sees and serves only its own configuration.
  • A data plane is the request-path deployment, running in a Kubernetes cluster, that contains the Controller and the AI Gateway proxy. Since 0.7.0 it lists the hostnames it serves in its own Helm values (its declared hostnames) and reports them to the management plane. One management plane can manage many data planes.
  • A project label connects the two. A platform administrator labels a data plane with each project it serves, and the data plane then serves the configuration of every project labelled on it. One data plane, and one hostname, can serve several projects, and one project can be served from several data planes. A project has no callable address until at least one data plane is labelled with it.

Data planes before 0.7.0 cannot declare hostnames. They keep the address stored for them through the management plane and serve projects through gateways, as described in each guide's section for earlier data planes.