boxd-cli: it teaches the agent on your laptop how to drive the boxd CLI. The agent loads it automatically whenever you ask it to do something with boxd, so there is no slash command to remember.
What it covers
With the skill loaded, the agent knows how to:- create, fork, and destroy machines, and pick the right lifecycle verb (stop, pause, hibernate)
- run commands in machines with
machine exec, stream long output, and move files withmachine cp - publish ports as HTTPS subdomains or raw TCP/UDP forwards
- manage snapshots, checkpoints, env vars and secrets, orgs, and API keys
- delegate work to agents running inside a machine and follow up on their sessions
exec, the network-idle suspend trap for CPU-only jobs, and the per-command JSON schemas for parsing --json output.
Install
The skill ships with theboxd CLI. One installer drops the binary and the skill:
~/.claude/skills/boxd-cli/SKILL.md- symlinks to the same skill in
${CODEX_HOME:-~/.codex}/skillsand~/.config/opencode/skills, when Codex or OpenCode is installed, so~/.claude/skills/stays the single source of truth
Manual install (if you prefer)
If you don’t want to run the CLI installer, fetch just the skill:/skills refreshes it without a restart.
Inside a VM
Every boxd VM image carries the in-VM counterpart, a skill namedboxd, already installed with nothing to do. It teaches the agent working inside a machine to drive the in-VM CLI: fork the current VM, manage its proxies, and run commands in sibling machines. Inside a VM the CLI signs in as you automatically, so the skill needs no key and no configuration. It works for all three coding agents pre-installed on every VM: it lives at ~/.claude/skills/ for Claude Code and is symlinked into Codex’s and OpenCode’s skill directories, the same fan-out the laptop installer does.