Claude Code on VPS Only: levelsio's Year-Long Production Workflow (June 2026)

Update — August 9, 2026: Same Hetzner + Tailscale + Cloudflare Tunnel ingredients, inverted — a Berlin engineer replaced the VPS with a rooted CMF Phone 1 running Termux.

On June 28, 2026, Pieter Levels posted a thread that became one of the year's most-cited agent-dev manifestos — ~1.2M views by mid-July — describing almost a year of coding solely on a VPS with Claude Code. No laptop open. No battery drain. Switch phone or any device. Agents keep going overnight with /goal. Hacky projects start on the server and go live in seconds because the code already sits on production infrastructure.

Two weeks later, Levels extended the same stack for Nomads iOS — Termius from iPhone, Xcode on rented macOS — in our companion VPS + SSH iOS guide. This post is the June 28 manifesto: deploy evolution, live production edits, solo vs team staging, and the Hetzner + Tailscale + Cloudflare checklist Levels shared in replies.

TL;DR — levelsio VPS Production Manifesto

What changed — three stages of deploy evolution

Levels did not jump straight to live production edits. His thread sketches a deliberate regression toward speed — each stage traded safety ceremony for seconds saved.

explainx.ai's read: Stage 3 only works when you treat the VPS as a disposable agent harness host, not sacred infrastructure. Levels' two ten-second outages are the honest error budget — not zero, but acceptable for a solo builder shipping fast. Teams with SLAs should stop at stage 1.5: agent edits staging, human or CI promotes to prod.

The opposite philosophy — local-first, repo upload, no cloud credentials on a shared box — is what Grok Build open source advocates. Both camps agree the laptop is leaving the critical path; they disagree on whether production should be one SSH session away from rm -rf.

Why VPS-only feels like "living in the future"

Levels listed benefits that map cleanly to how long-running agents behave in 2026:

1. No laptop open / no battery drain — the harness runs on wall power in a datacenter
2. Device agnostic — phone, tablet, hotel PC; SSH is the UI (July iOS thread made Termius explicit)
3. Overnight continuity — /goal and tmux survive your sleep; wake up to a diff, not a dead process
4. Instant hacky deploys — spin a project on the same machine that serves traffic; no "works on my machine" transfer
5. Blast-radius separation — after Codex $HOME deletion reports, a wiped /home/dev on a €5 VPS beats a wiped laptop home directory

That last point is why solo builders tolerate stage 3 live prod edits. The VPS is already the sandbox — until it isn't, because production databases and customer sessions live on the same disk.

Solo on prod vs teams on staging

Levels drew an explicit line in the thread:

Staging pattern that preserves speed:

[Claude Code on staging VPS] → git push / rsync → [prod VPS auto-pull or blue-green]

Agents get full shell freedom on staging; production receives only reviewed artifacts — aligned with thin prompts, thick artifacts: chat stays thin, git carries truth.

For harness mechanics — permission modes, tool loops, checkpoint restore — see What is an agent harness?.

Hetzner + Tailscale + Cloudflare — setup checklist

Levels' setup reply on the June 28 thread named a concrete stack. This is explainx.ai's expanded checklist — official docs linked where noted; no competitor product URLs.

1. Provision Hetzner VPS

• CX or CPX tier: 2+ vCPU, 4+ GB RAM for comfortable Claude Code + build tools
• Ubuntu 24.04 LTS, single non-root dev user
• Enable Hetzner Cloud Firewall: block all inbound by default

2. SSH hardening

# As root on fresh VPS — create dev user
adduser dev
usermod -aG sudo dev
mkdir -p /home/dev/.ssh && chmod 700 /home/dev/.ssh
# Paste Ed25519 public key → authorized_keys (chmod 600)

# Disable password auth in /etc/ssh/sshd_config
# PasswordAuthentication no
# PermitRootLogin no
sudo systemctl reload sshd

3. Install Claude Code + tmux

# As dev — persist sessions for phone reconnect
sudo apt update && sudo apt install -y git tmux build-essential
tmux new -s code
# Install Claude Code per official Linux docs
# ANTHROPIC_API_KEY in ~/.config/agent/env (chmod 600) — never in prompts

4. Tailscale (private mesh)

• Install Tailscale on the VPS and every client device
• SSH via Tailscale IP or MagicDNS hostname — no public SSH port required
• Prefer Tailscale SSH or key-based auth; disable exposure of port 22 to the public internet

5. Cloudflare Tunnel (optional public HTTPS)

• Run cloudflared on the VPS for web preview or admin UIs
• Tunnel terminates TLS at Cloudflare; origin stays off the public internet
• Pair with Cloudflare Access if you expose anything beyond localhost

6. Firewall summary

7. Agent session hygiene

tmux attach -t code || tmux new -s code
cd ~/your-product
claude   # harness runs here; /goal for overnight tasks
# Ctrl-b d to detach — iPhone Termius reconnects to same pane

For OpenCode on a VPS with local models instead of cloud API keys, see run open-source models in OpenCode. For visual session tabs instead of tmux, compare OpenCode Desktop worktrees.

Theo's six-month thesis — agents leave the laptop

Levels agreed with Theo (~Theo — builder and commentator in the agent-dev space): roughly six months until most developers move coding agents off laptops onto servers or cloud sandboxes.

Evidence in the thread and adjacent coverage:

• Karpathy reportedly runs Claude via cloud/Slack — agents operate server-first, not IDE-first (Anthropic pre-training hire)
• Community sandboxes — kayhe's sandbox setups, Phil Freo's Docker dev environments, peetzweg's 160k GitHub contributions from agent-assisted workflows
• Peter Yang asked whether dedicated Claude Code / Codex desktop apps matter if the harness lives on a remote box — the answer trending toward no, SSH + tmux + /goal is enough UI

explainx.ai tracks the same shift in phone agent guides: the supervisor is mobile; the worker is always-on infrastructure.

Safety, backups, and when NOT to live-edit prod

3-2-1 backups (Levels' recommendation)

Automate daily DB dumps and pre-agent snapshots before destructive refactors. Treat snapshots as cheap rollback — not a substitute for git.

Hook-level guards

Even on a VPS, agents run shell. Layer defenses from Destructive Command Guard:

• Block rm -rf /, git reset --hard, mass chmod
• Run harness as unprivileged user — not root
• Never enable full access on hardware you cannot wipe — see Codex $HOME chain

When NOT to live-edit production

Do not let Claude Code touch prod directly when:

• Customers pay for uptime and a 10-second outage has real cost
• Multiple engineers share one VPS without branch isolation or worktrees
• PCI, HIPAA, SOC2, or similar requires segregation of duties
• No staging mirror exists — you cannot diff before traffic hits changes
• Backups are untested — you have never restored from them
• Secrets live in the repo the agent reads — .env, signing keys, prod DB URLs
• Database migrations need rollback plans agents cannot infer

In those cases: staging VPS + git promotion + CI smoke tests. Solo indie? Levels' error budget is instructive — two ten-second incidents in twelve months — not a guarantee for your stack.

Community replies — what builders are actually running

The June 28 thread spawned practical forks without changing the core thesis:

Local-first counterweight: Grok Build keeps inference and repo on hardware you own — opposite deployment, same agent loop.

How June 28 and July 8 fit together

Same VPS brain; July adds second SSH hop to macOS for Xcode. Start here for why; read July for how on iPhone.

Summary

June 28, 2026: Pieter Levels documented ~12 months of VPS-only Claude Code — overnight /goal runs, three-stage deploy evolution ending in live production edits, and two ~10-second outages in a year. Solo on prod; teams on staging. Stack: Hetzner, SSH, Tailscale, Cloudflare Tunnels, inbound firewall deny.

explainx.ai's read: this is the server-first agent harness thesis Theo and Karpathy echo — your laptop becomes a terminal remote, not the source of truth. Pair speed with 3-2-1 backups, DCG hooks, and honest when-not-to rules. For Nomads iOS and Termius, continue to the July 8 companion post.

Related on explainx.ai

• CMF Phone 1 Termux home server — VPS replacement (Aug 2026) — skip Hetzner; rooted Android as the box
• Claude Code VPS + SSH — iOS and MacinCloud (Jul 8) — sibling extension: Termius, Xcode hop
• Codex $HOME deletion — full access sandboxing — why disposable VPS beats laptop prod
• /goal command — long-running agents — overnight runs Levels cited
• Grok Build open source — local-first vs cloud — opposite deployment philosophy
• Andrej Karpathy joins Anthropic pre-training — server/cloud agent orchestration context
• Thin prompts, thick artifacts — Thariq — git as promotion artifact
• What is an agent harness? — permission modes and tool loops
• OpenCode Desktop tabs — worktrees — visual parallel to tmux
• Run open-source models in OpenCode on VPS — no cloud API keys on box
• Destructive Command Guard — hook-level shell protection
• Run AI coding agents from your phone — mobile supervisor patterns

Official infra docs: Hetzner Cloud · Tailscale install (Ubuntu) · Cloudflare Tunnel

Context: Pieter Levels X thread, June 28, 2026 (~1.2M views, edited 2:52 PM) · Companion July 8, 2026 Nomads iOS thread (~322.5K views)

Thread view counts, agent CLI behavior, and Hetzner/Tailscale/Cloudflare docs reflect community reports through July 16, 2026. Live production editing carries real outage risk — verify backups and staging before copying a solo-indie error budget. Published June 28, 2026 (story date); last updated July 16, 2026 with July 8 iOS cross-links and Codex safety context.