Onboard MCP Tools

Give a crew new capabilities by onboarding MCP servers.

A crew gains new abilities by gaining new tools, and tools come from MCP servers. You can declare them directly, or let Kubemoot’s autonomic onboarding acquire them when the crew hits a capability gap. This guide covers both.

Declare a tool server directly

When you already know the MCP server you want, declare an MCPServer and let an MCPGateway discover it. The operator handles the deployment, health probes, and - for stdio servers - the auto-injected mcp-bridge sidecar.

apiVersion: kubemoot.ai/v1alpha1
kind: MCPServer
metadata:
  name: kubernetes-mcp
  labels:
    app.kubernetes.io/part-of: my-crew
spec:
  image: mcp/kubernetes:latest
  transport: stdio
  command: ["node", "dist/index.js"]

The gateway selects servers by label, aggregates their tools, and exposes them to agents. Then scope the tools onto a Tooler with spec.enabledTools, keeping the set small (~10-15 tools) for reliable tool-calling. See MCP Tools for the chain and the MCPServer / MCPGateway references for the specs.

Let the crew propose a tool for itself

When a discussion finds that no existing agent can answer, the coordinator signals a capability gap. An agent running in onboarding mode reacts to that signal:

  1. Detects the gap from the coordinator’s signal.
  2. Proposes an MCP server from its model’s own knowledge (the call carries no tools).
  3. Replies with a proposed MCPServer manifest when the user consents. You review and apply it.

Everything after the proposal is manual today: you apply the manifest, create the Tooler Agent that uses the new server’s tools, and optionally a RAGSource for its documentation. The operator does not create the Tooler for you. Because every step produces ordinary Kubernetes resources, you can inspect, version, and remove them like anything else.

Onboarding mode is an environment variable (KUBEMOOT_ONBOARDING_MODE=true) on an Agent you declare. The operator chart’s internalAgents values deploy the kubernetes-mcp and fetch-mcp servers (github-mcp is opt-in) and a quality policy; an MCPCatalog is rendered only when onboardingAgent.enabled is set. The onboarding and RTFM agents are off by default and the chart creates no Agent for them. For the flow, configuration, and troubleshooting, see the Onboarding Guide.

Read the requester’s date in a tool

A tool server can run in another time zone than the person asking. A tool whose answer depends on the date (“recent”, “next”, “as of”) should use the requester’s date, which every tools/call carries in params._meta when the client sent one:

{"method": "tools/call",
 "params": {"name": "find_released", "arguments": {"topic": "kubernetes"},
            "_meta": {"timeZone": "America/Los_Angeles",
                      "localDateTime": "2026-10-24T09:15:00",
                      "requestDate": "2026-10-24"}}}

Read it in this order, and ignore a value you cannot parse:

  1. requestDate, as the date.
  2. timeZone, with the date now in that zone.
  3. The tool’s own default zone, which the crew chart configures. A request without _meta behaves this way.

State the date and zone the tool used in its result, so a reader can see which “today” applied. The date matters more than the time of day. Never take the date from a tool argument the model fills in. The Time MCP server answers in its server’s zone; use it only when the requester’s date does not matter.

Which path to use

  • Declare directly when you know the tool and want it deterministically present - the normal way to equip a crew you’re authoring.
  • Autonomic onboarding when you want the crew to grow to meet questions it wasn’t pre-built for.