- The author's laptop was the bottleneck in AI coding sessions because closing the lid would kill running sessions like Claude Code.
- Solution: Move all development work (agents, code, databases, editor GUI) to a VPS, accessible from any device via Tailscale VPN.
- Key benefits: close laptop without stopping work, instant handoff between devices, survives dropped connections, no need for caffeine hacks to keep laptop awake.
- Setup uses three core pieces: Tailscale for private network, a VPS (Ubuntu 26.04 LTS, 2 vCPUs, 8 GB RAM), and a browser-based coding GUI (T3 Code, but swappable).
- Tailscale provides a mesh VPN; the VPS exposes only one public port (SSH break-glass) and all other services are bound to localhost and published on the tailnet.
- The server uses btrfs on a separate volume for copy-on-write (cp --reflink) and compression (zstd:3), saving disk space and allowing instant cloning of large directories like node_modules.
- The author stopped using a traditional code editor; instead, they describe tasks to an agent, review diffs, and approve changes via the browser GUI.
- Custom automation includes: systemd timer for daily updates of Claude Code, Codex, and T3 (with rollback), one-command project onboarding (newproject script), and per-worktree preview servers with garbage collection.
- Two agents cannot share a working directory; git worktrees isolate each agent's workspace, with a script (worktree-up) to set up environment and preview server.
- The server also runs scheduled tasks (cron) like nightly dependency bumps and changelog generation, and integrates with Home Assistant via MCP for home control.
- Monitoring via Oh Dear (port scanning) ensures no unexpected public ports are exposed, especially since Docker can bypass firewall rules.
- The entire setup can be reproduced by giving the article as a prompt to a coding agent on a fresh Ubuntu VPS.