Skip to main content
The Orka3 CLI agent skill gives AI coding agents deep knowledge of the Orka CLI and, when you’re authenticated and connected to the VPN, the ability to run commands directly against your cluster. You can ask questions in plain English, have your agent generate and execute orka3 commands on your behalf, or use it to build and test automation workflows, all from your terminal. The skill isn’t Claude-specific: it’s structured Markdown context any agent that supports custom instructions or project context can load. Claude Code and Gemini CLI can run orka3 commands directly on your machine; Cursor, Windsurf, GitHub Copilot, Claude Desktop, and ChatGPT or similar tools use the same content as guidance, generating the correct command for you to run yourself. This page focuses on the Claude Code walkthrough since it’s the most complete integration today, see Getting started below for the other agents.
Direct cluster access requires an active VPN connection and an authenticated orka3 session. Without those, your agent answers from documentation only and suggests commands for you to run yourself.

​
What you can do

Ask questions and get accurate answers
  • “What images are available in my namespace?”
  • “What’s the difference between vm save and vm commit?”
  • “How do I pre-cache an image across all ARM nodes?”
Have Claude run commands for you With your permission, Claude can execute orka3 commands directly: deploy VMs, check node status, manage images, create service accounts. You confirm before anything runs. Build and test automation Claude can help you write CI/CD pipeline configuration, generate service account setup scripts, and test workflows end-to-end against a live cluster. Troubleshoot live issues
  • “My VM deployed but SSH isn’t working. Check what’s running.”
  • “Image pull is stuck at Pending. What’s the status?”
  • “List all VMs in the staging namespace and show me which ones are idle”

​
How it works

The skill loads as a system prompt when you open Claude Code. It gives Claude structured knowledge about:
  • Every orka3 command and its flags
  • CI/CD authentication patterns (service accounts, token scopes)
  • Intel vs. Apple silicon feature differences
  • Async operation patterns (save, commit, push, imagecache)
  • Common error messages and how to resolve them
When the orka3 CLI is available in your environment, Claude can also run commands directly and interpret the output in context.

​
Getting started

Installation instructions and the full skill source are in the GitHub repository: github.com/macstadium/orka3-cli-agent-skill For Claude Code, you’ll need Claude Code installed and a valid Anthropic API key. The skill installs in a single command.

​
Agent compatibility

“Execution” means the agent can call orka3 directly on your machine. “Guidance only” means the agent gives you the correct command; you run it yourself. Either way, the skill content is the same, see the repository’s README for per-agent installation steps. Customers have used the skill this way with GitHub Copilot specifically, loading the same context as a repo instructions file and getting accurate, correct commands back without having to feed Copilot Orka-specific context by hand.
AI agents can make mistakes, especially with destructive operations like deleting VMs or modifying namespaces. Always review what the agent plans to do before confirming, and keep a human in the loop for anything touching production infrastructure.

​
Tips for getting the most out of it

  • Give Claude context about your environment. Mention whether you’re on Intel or Apple silicon nodes, and which Orka version you’re running. Claude will tailor commands accordingly.
  • For CI/CD questions, describe your pipeline. Tell Claude which CI platform you’re using and whether you’re running inside a container. The auth approach is different.
  • Ask follow-up questions. If a suggested command doesn’t match your setup, explain why and Claude will adjust.
  • For long-running operations, ask how to check status. Image saves, commits, and pushes are async. Claude knows the right status-check commands.