AgentBox Skills

AgentBox installs Claude Code skills so the agent can drive boxes for you — fork a session, operate the CLI, and configure a box from the inside

AgentBox ships a handful of Claude Code skills (mirrored as Codex prompts and OpenCode commands) so the agent itself can drive AgentBox. One forks your current session into a new box, one teaches the agent to operate the agentbox CLI on your host, and one runs inside a box to configure it. The npm package installs them for you.

NOTE

These are the AgentBox-provided skills, not the same thing as your own ~/.claude skills — though those flow into every box too (see Keeping boxes in sync). For the agent session itself, see Run an agent.

Install the skills

The skills come with the npm package. The first-run agentbox install wizard installs them as its last step; re-run it (or the --skills-only shortcut) any time to refresh them.

npm -g install @madarco/agentbox
agentbox install                  # full wizard; installs the host skills as its last step
agentbox install --skills-only    # just (re)install the host skills

They land in ~/.claude/skills/agentbox/ and ~/.claude/skills/agentbox-info/, with mirrors at ~/.codex/prompts/agentbox.md and ~/.config/opencode/commands/agentbox.md. The installer is idempotent and only touches files it wrote — pass --force to overwrite a copy you've hand-edited. See CLI commands for the rest of the install flags.

/agentbox — fork a session

/agentbox [provider] is the user-invokable slash command. Run it from Claude Code, Codex, or OpenCode and it teleports your current session — the full conversation, plus the active plan if you're in plan mode — into a fresh box and resumes it there. The provider argument is optional and defaults to local Docker.

/agentbox            # fork this session into a local Docker box
/agentbox hetzner    # fork into a Hetzner cloud box

Under the hood it calls agentbox fork. It's host-only in v1 (forking from inside a box isn't supported yet). The full flag reference — --agent, --name, --plan, and the rest — lives in CLI commands.

agentbox-info — operate AgentBox from chat

agentbox-info is a host-side reference skill. Claude loads it automatically — there's nothing to type — when you ask it to spin up boxes, queue background runs with -i, or push commits safely through the host relay. It's the knowledge that lets an agent already running on your laptop operate the agentbox CLI on your behalf.

TIP

You never invoke agentbox-info directly; it's auto-loaded by relevance. Just ask your host agent to "start a box" or "queue three runs" and it knows the commands. See Orchestration & queue boxes.

agentbox-setup — configure a box from the inside

agentbox-setup runs inside a box. When the agent in a box has no agentbox.yaml — or you ask it to (re)configure one — this skill walks it through discovering the project's services and tasks and writing /workspace/agentbox.yaml, so the in-box supervisor can boot the workspace deterministically.

It ships in the box image rather than in your host ~/.claude, so it's only active where it's useful. See Services & tasks for the file it produces and how the supervisor runs it.

Keeping boxes in sync

Your host ~/.claude is the source of truth. Every agentbox create / agentbox claude syncs it into the next box, so your skills, plugins, MCP servers, and login all flow in automatically. A running box doesn't resync on a plain attach — use agentbox claude start after you install a new skill or sign in on the host, and it picks up the change.

HEADS UP

attach never resyncs your host ~/.claude. After installing a skill or signing in on the host, use agentbox claude start (or a fresh agentbox claude) so the box picks it up.

Next steps

  • Run an agent — launch and fork agent sessions, with the full agentbox fork reference.
  • Services & tasks — the agentbox.yaml that agentbox-setup writes.
  • CLI commandsagentbox install and every other command.

On this page