The Realm and the Crew
Kubemoot has about twenty resource kinds. They fall into two layers, plus a small cluster level above both:
- The realm is what crews draw on: the working context and the limits a crew works within. Models, tools, knowledge, and where notifications go.
- The crew is the crew itself: who sits at the table and how it behaves.
The cluster level
KubemootConfig and MootArchetype are cluster-scoped. They set how Kubemoot runs and which consensus archetypes exist, for every crew in the cluster.
The realm
| Part | Kinds | What it gives a crew |
|---|---|---|
| Models | ModelServer, Model, EmbeddingModel | The catalog of models the cluster can run, and the servers that serve them. The cluster owner keeps ModelServers and Models in the operator’s namespace |
| Tools | MCPServer, MCPGateway, MCPCatalog, MCPQualityPolicy, MCPServerReport | The tools agents may call, the gateway they reach them through, and the rules for which community servers are let in. MCPServerReport is generated, not written: it records how each server fared in trials, as a status report beside the quality policy that reads it |
| Knowledge | RAGSource | Indexed documents an agent can search |
| Delivery | NotificationSink | Where a concern raised in a discussion is sent |
These are namespaced kinds, and where one lives depends on who needs it. The model catalog is shared, so it sits with the operator. A crew’s own tools, knowledge, and notification sink can ship with the crew, in the crew’s namespace, and serve that crew only.
The crew
| Part | Kinds | What it is |
|---|---|---|
| The crew itself | Crew, Agent | Who sits at the table and the role each plays. See Crews and agents |
| Its rules and know-how | PromptModule, Skill | The ADL rules that govern an agent and the skills it can load |
| Its evidence | CrewFitness, CrewFitnessSuite | Scenarios that score the crew against known answers. See Fitness |
On the line: CrewSchedulingPolicy
A CrewSchedulingPolicy belongs to a crew, but it says how that crew draws on the
realm: for example, quality over speed, or which capabilities are hard needs and which
are wishes. It is the crew resource that selects from the catalog.
The point: crews ask, the realm fills
An agent declares the capability it needs, never a model, a provider, or an endpoint. The scheduler places each call on what the realm offers at that moment (see Models and scheduling).
That is why a crew is portable. The same crew runs on a small cluster with two accelerators, on a cluster with a large pool, or against a catalog that changes underneath it, with no change to the crew: moving a model to another accelerator changes the realm, not the crew’s namespace.
It also divides the work. A platform team curates the realm: which servers run, which models the catalog holds, which tool sources are allowed. Crew authors write crews and ask for capabilities. Neither edits the other’s resources.