What it does Cabinet is a local-first “startup OS” that runs as a Next.js app with a background daemon. You answer five onboarding questions and it spins up a custom AI team — CEO, content marketer, QA agent, whatever fits — each with goals, scheduled cron jobs, and a persistent workspace. All output lives as markdown files on disk, auto-committed to git. No database, no Notion, no vendor lock-in.

The interesting bit The project treats AI memory as infrastructure that needs history and auditability, not as ephemeral chat context. Agents run through structured adapters for Claude Code and Codex CLI, with persisted conversations and transcript-driven live views. The web terminal stays as an interactive surface, but the real work now happens through a scheduler-backed daemon with per-run provider overrides.

Key highlights

• 20 pre-built agent templates across leadership, product, marketing, engineering, and ops
• Cron-based automation: agents scout Reddit, write reports, file tasks while you sleep
• Drop an index.html anywhere in the tree — it renders as an embedded app, version-controlled, no build step
• Full git-backed history with diff viewer and point-in-time restore
• Skills installable from skills.sh or any GitHub repo, attachable per-agent or @-mentioned per-task
• npx create-cabinet@latest install, no global package needed; updates pulled from GitHub releases

Caveats

• Requires Node.js 22+ and either Claude Code CLI or Codex CLI installed globally
• Windows only via WSL; macOS and Linux native
• Node 20 “still works” but throws an EBADENGINE warning from a transitive dependency
• Cloud-hosted version is waitlist-only; self-hosting is the current path

Verdict Worth a spin if you want AI agents that actually accumulate context across sessions and leave inspectable artifacts behind. Skip it if you’re looking for a simple note-taking app or need fully managed SaaS today.