nanabox

First-run guided onboarding for a freshly provisioned nanabox, run by the box's root/admin agent. Interactively walks the user through every runtime capability — captures what the box is for (and writes it into the system prompt), optional GitHub access, creating their first non-root agent, inter-agent messaging, shared secrets, the headless browser, publishing web apps, and custom skills. Use when a box is new, on the root agent's first session, or when the user says "onboard me", "get me started", "give me a tour", or "what can this box do?".

Drive the long-lived Chromium running on this nanabox over the Chrome DevTools Protocol. Use when the user asks to scrape a page, automate a browser flow, fill a form, log into a site, take a screenshot, or otherwise control a real browser. The user can watch and take over the same session from their own browser tab — useful for one-off logins, captchas, or MFA.

Publish a web app (a page, an API, a webhook receiver) from this nanabox. Use when the user asks to host a page, expose an API, share a script as a URL, set up a webhook, or otherwise serve something over HTTP from this box. Your app binds a unix socket; Caddy reverse-proxies your handle's path to it, and an ephemeral cloudflared quick tunnel makes it reachable on the public internet.

List the other agents on this box and send them a message (one-to-one or broadcast). Use when the user asks "who else is running here?", "tell the other agents X", "broadcast to all agents", or when you want to coordinate work with a peer agent — e.g. handing a sub-task to a peer while you stay focused on your repo. Messages are fire-and-forget; the receiver reacts on its next turn.

Understand and operate the agents on this nanabox — list them, peek at a session for debugging, restart a wedged one, and push a one-off prompt in. Use when the user asks about "the agents" or a specific agent's session. Creating/removing agents and changing box config are the box operator's job, not yours.

End-of-day consolidation routine. A box-wide systemd timer fires `/nightly-cycle` overnight (around 03:00–05:00 UTC; the firing is staggered across the fleet and across agents on a box to avoid an API-rate-limit burst) for each agent on this box. The agent reviews the day's work, distills durable insights into memories and skills, writes a brief day summary at its per-handle path `~/.agents/<handle>/nightly-summary.md`, then `/clear`s the session. When fired by the timer (vs run manually), the box ALSO restarts the agent's systemd unit after the summary lands so the next session re-reads the runtime CLAUDE.md + skills — which `/clear` alone doesn't reload.