MCP Lifecycle Operator

Kubernetes operator for managed Model Context Protocol server workloads.

Agentic Friendly

Component Category

AI tools / MCP Server lifecycle

Component Description

MCP Lifecycle Operator reconciles managed Model Context Protocol server workloads from declarative Kubernetes resources. It creates and updates the server deployment, service, and reported runtime status used by the BSQAI API and portal.

Why It Is Used

In BullSequana AI, MCP Lifecycle Operator gives managed MCP Servers a dedicated controller and lifecycle API. This separates workload reconciliation from the BSQAI API while keeping server creation, updates, health, and deletion observable.

Learn More

Deployment notes

BullSequana AI deploys MCP Lifecycle Operator v0.3.0. The operator runs in mcp-lifecycle-operator-system. A tenant organization's Managed MCP Servers run in its tenant namespace, tenant-<slug>, so they count against the tenant's compute quota and are removed with the tenant. Organizations without a tenant, such as the Global organization, use the shared mcp-servers namespace. In both, a NetworkPolicy admits only the MCP Gateway, the BSQAI API and the operator to server pods. It isolates them only on clusters whose network-policy engine enforces NetworkPolicies; without one it is declared but inert. With the platform's network policies on, which is the default, a server's outbound traffic is limited to cluster DNS and to the addresses in NETWORK_POLICY_EXTERNAL_EGRESS_CIDRS on the ports in MCP_LIFECYCLE_OPERATOR_NETWORK_POLICY_SERVER_EGRESS_PORTS, 443 and 80 by default. With network policies off, or with no external addresses set, egress is open. Cilium matches those addresses only outside the cluster; Calico also matches pod addresses, so a Calico site lists its external ranges. An admission policy refuses any per-server NetworkPolicy the operator writes that would widen this; such a server stays not ready. Add only images you trust. Every server runs as the mcp-server ServiceAccount without a Kubernetes API token. The platform mirrors the operator image into its private registry and installs the upstream CRD and controller with the required image-pull configuration.

A tenant's servers start once the tenant runtime reports it is ready for them; creating one earlier is refused with a message to try again. The BSQAI API stores each Managed Server as desired state, and its MCP reconcile worker converges the MCPServer resource every 30 seconds. A Managed Server that runs outside its organization's tenant namespace moves there once the tenant is ready for servers. It restarts once and keeps its configuration and its users' linked credentials. See Managed Connectors move into tenant namespaces.

Remote MCP Servers do not create operator-managed workloads. The BSQAI API stores their HTTPS address and reaches them through the MCP Gateway.

Interacts With

  • BSQAI API, whose MCP reconcile worker creates managed server resources, and which reports their status.
  • MCP Gateway, which authenticates callers and injects linked upstream credentials.
  • AI Web Portal, which manages server definitions, credentials, and tool allowlists.
  • Keycloak, which authenticates users and identifies the active tenant.

On this page