Automate Operations

Use an AI agent for operational tasks: managing cron schedules, updating clusters, importing workflows.

Operational work on Pipekit doesn't always need a human reading dashboards. The MCP server exposes enough write-side tools to let an agent run routine ops.

This is the most powerful (and the most destructive) pattern. Read Security & Transports before granting an agent these tools, and lean on the destructiveHint: true annotation to require approval prompts.

Operations the agent can run

The exact flow depends on the task. A few common shapes:

Manage CronWorkflows

The agent uses create_cron, update_cron, suspend_cron, resume_cron, delete_cron, and trigger_cron to manage scheduled workflows.

Typical prompt:

Pause the nightly-export cron on all clusters until I tell you to resume it.

The agent calls list_runs or queries Pipes to find the cron, then suspend_cron per cluster.

For the underlying model, the MCP server ships pipekit://guides/cron-workflows as a resource. The agent reads it as needed.

Update Cluster metadata

update_cluster changes a Cluster's name, description, or active status. Typical prompt:

Mark cluster eu-west-2 as inactive. We're decommissioning it next week.

The agent calls list_clusters to confirm the name, then update_cluster with --status=inactive.

Import from an Argo Workflow Archive

import_workflows walks an Argo Workflow Archive and surfaces the archived workflows as Pipes in Pipekit.

Typical prompt:

Import the last 30 days of workflows from the archive at argo-server:2746 into cluster data-prod.

The agent calls import_workflows with the appropriate --argoServerURI and --cluster-name parameters. For the underlying mechanics, the agent can fetch pipekit://guides/importing-workflows.

Safe use of destructive MCP tools

• Use the HTTP transport with a dedicated token if you want fine-grained control over what the agent can reach. See Security & Transports.

• Run a dedicated agent against a non-prod Organization first. Operational tools change real state. Practice in dev before pointing the agent at production.

• Honor MCP annotations. destructiveHint: true exists because clients can prompt before destructive calls. Make sure your client surfaces those prompts.

Permission notes

This pattern uses several destructiveHint: true tools. In MCP-annotation-aware clients, the agent has to ask for approval before calling them. For the full list of which tools are destructive, see Tool Inventory.