Runebox sync — Claude Code plugin + standalone CLI

Your team's Runebox registry, kept in sync on every engineer's machine: log in once with an API key and your org's approved skills, agents, and commands install and stay current — on Claude Code, Codex CLI, Cursor, or Gemini CLI. One sync core (scripts/sync.py, stdlib-only) ships two ways:

• On Claude Code — the plugin below, with a SessionStart hook.
• Everywhere else — the same script as a standalone runebox-sync CLI via pipx/uvx. No Claude Code required.

Install — Claude Code plugin

claude plugin marketplace add runebox-ai/runebox-plugin
claude plugin install runebox
/runebox:login

/runebox:login asks for your org's API key (starts with rbx_; your admin creates it on the org's Get set up page at runebox.ai), verifies it, auto-detects which harnesses you have installed, and syncs the whole approved catalog in one shot.

Install — standalone CLI (Codex, Cursor, Gemini CLI)

pipx install runebox-sync   # or: uvx runebox-sync login
printf '%s' "rbx_yourkey" | runebox-sync login

Same key as the plugin path — an engineer on Codex/Cursor/Gemini CLI never touches Claude Code at all. Add runebox-sync autosync to your shell profile so it keeps syncing on every new shell (there's no SessionStart hook outside Claude Code):

runebox-sync autosync --apply

Commands

How sync works

• Per org, per enabled harness: one conditional GET against the org catalog. Nothing changed → a cheap 304 and zero output. Something changed → it installs/updates/removes and prints one line. A skill installs identically to every harness (the open SKILL.md format); a command's filename/location differs per harness; agents/output styles/claude_md stay Claude Code-only for now.
• Every file it installs is tracked in a local manifest, per (org, harness). It only ever touches files it installed itself — never your own skills or settings.
• When an admin yanks a bad version, the next sync removes it from your machine. When your key is revoked, sync stops with a single notice.
• Adding a harness the server can target requires a client release (a new allowlisted install root) — a compromised server can't write anywhere new.

Requirements

• python3 on PATH — the sync script is stdlib-only, no dependencies, works standalone or as a Claude Code plugin.

Security notes

• API keys are read from stdin at login (never passed as a CLI argument) and stored in ~/.runebox/credentials.json with 0600 permissions.
• Install paths from the server are validated client-side against a per-harness roots allowlist — a hostile path can't escape those directories, and an unrecognized root name is rejected outright.
• Sync failures never block your session; the hook (and --quiet sync) exits quietly.

Runebox is a private, governed artifact registry for teams using AI coding agents — publish once, scan + review every version, and every engineer's tools stay in sync, whichever harness they run. Free for teams up to 5 seats at runebox.ai.