Claude Code supports VPS agents for overnight coding sessions

• levelsio's VPS setup thread described a Claude Code workflow that lives on a server instead of a laptop, stays reachable from a phone, and can keep running overnight with /goal.
• In replies to the main thread, levelsio's setup note named the stack plainly: a Hetzner VPS, SSH, Tailscale, Cloudflare Tunnels, and a firewall with inbound traffic blocked.
• Peter Yang's interview post and Peter Yang's codebase clip showed Anthropic making the same broader bet inside the company, where Jess Yan said long-running cloud agents help teams inspect repos, synthesize feedback, and track merged PRs.
• The workflow is already drifting beyond coding, because Moritz Kremb's MCP thread used Claude Code to draft and format a newsletter, while his YouTube workflow post used it to score titles and thumbnail ideas.
• The tradeoff is visible in the same thread, where levelsio's outage reply said live edits on production only broke his site twice in 12 months, and Peter Yang's question captured the bigger tension between cloud agents and local hardware.

You can watch Peter Yang's episode, skim Jess Yan's codebase demo, and inspect the concrete server recipe in levelsio's replies. There is also a surprisingly mundane phone setup in levelsio's Termius reply, plus a cheap but not tiny cost number in his $200 a month estimate.

The cleanest argument for Claude Code on a VPS came from levelsio's main thread, which reads less like a hot take than a field report after a year of doing it.

His claimed gains were concrete:

• no laptop needs to stay awake
• no battery drain
• the same session can continue from a phone or another device
• long jobs can run overnight
• deploy time collapses because the agent is already on the server

A follow-up in levelsio's IDE reply made the workflow even starker: he said Claude Code on the server handled 95 percent of his work and that he had not used an IDE in about 12 months.

The access pattern here is not fancy. In levelsio's setup note, he told people to create a VPS, SSH in, install Claude Code, add Tailscale and Cloudflare Tunnels, and block inbound traffic with a firewall.

The phone angle is part of why this took off. levelsio's Termius reply said he uses Termius on iOS, and levelsio's Tailscale reply said a laptop should sit inside the same Tailscale network.

That makes the machine less like a personal computer and more like a persistent workspace you visit from whatever screen is nearby.

The wildest part of the thread was not that the agent runs remotely. It was that levelsio's main thread said it live edits his production server.

He also filled in the failure rate. In levelsio's outage reply, he said Claude Code only broke the site twice in 12 months, once with a PHP error that took the site down for about 10 seconds before it was fixed.

There was a boundary on that claim. In the same main thread, he said that if he were working inside a bigger company he would point the workflow at staging instead of production.

The external creator workflow lines up with what Anthropic described publicly. In Peter Yang's interview post, Peter Yang said Jess Yan, product lead at Anthropic, walked through building a long-running Claude agent and described internal uses that included understanding the codebase, synthesizing user feedback, and pressure-testing API decisions.

Yan's most specific claim came in Peter Yang's codebase clip, where she said codebase access helped her manage state by tracking which PRs were merged and deployed instead of asking engineers for updates.

That is a different pitch than "AI writes code for you." It is closer to giving a product team a durable operator that can sit near the repo and keep watch.

A smaller but revealing thread showed how people are adapting the terminal itself around Claude Code. LLMJunky's cmux demo showed a pane-splitting setup, and LLMJunky's follow-up said to use cmux on macOS or tmux on Linux.

cmux split-pane demo for Claude Code

This fits the cloud setup better than it first appears. Once Claude Code lives on a server, the terminal becomes the product surface, so tricks like split panes, extra SSH tabs, and side-by-side long runs stop looking niche.

The most useful new detail is that this workflow is already bleeding into creator tools. Moritz Kremb's MCP newsletter thread said he now writes a newsletter in 15 minutes instead of four hours by pulling a transcript into Claude Code, drafting from templates stored in his "Claude OS," then telling Claude to format the result directly in Beehiiv through MCP.

His other examples were similarly operational:

• the YouTube workflow post used yt-dlp, Notion, and vidIQ to rank titles and thumbnail ideas
• the business OS demo described a setup that grades sales calls, writes newsletters, builds Meta ads, and schedules content

That makes the VPS story bigger than overnight coding. The same persistent Claude Code session is being used as a back-end workspace for writing, research, channel analysis, and content operations.