Skip to content

AI assistants

An AI assistant that runs on your workstation can use cv4pve-pepper to open the console of a VM or container for you: you ask for the console of webserver, and the remote-viewer window opens on your desktop. The assistant does not see the screen of the VM; it finds the guest, starts the viewer and tells you what went wrong when the console does not open.

The repository ships a skill, skills/cv4pve-pepper/SKILL.md: a short file in the Agent Skills format that Claude Code, Codex and other assistants load when a request matches its description. It tells the assistant to:

  • connect only with the options file, passed with @, and never print it;
  • use the exact id or name of the guest, and ask when you gave only part of a name;
  • add --start-or-resume, which starts a stopped guest, only after you agree.

Install it with the skills command, which needs Node.js:

npx skills add Corsinvest/cv4pve-pepper

Or copy the file into the skills folder of your assistant: .claude/skills/ for Claude Code,.agents/skills/ for Codex, in the project or under your home folder:

mkdir -p .claude/skills/cv4pve-pepper
curl -fL -o .claude/skills/cv4pve-pepper/SKILL.md https://raw.githubusercontent.com/Corsinvest/cv4pve-pepper/master/skills/cv4pve-pepper/SKILL.md

The skill tells the assistant what to do; the token decides what it can do. The assistant can read the token and call the Proxmox VE API without cv4pve-pepper, and the skill cannot prevent that. Give it a token with VM.Audit and VM.Console only on the guests it may open, and VM.PowerMgmt only if it may start them: see Permissions. Put the connection and the path of the viewer in an options file, so the token stays out of the conversation.