DEV Community

Josep
Josep

Posted on Originally published at Medium on

The AI Dev Environment Nobody Teaches You

Architecture of a remote agentic development environment

The idea in one sentence

Don’t run Claude Code, Codex or OpenCode on your laptop. Run them on a VPS that is on 24/7 , which clones your repos, installs whatever is needed, runs, tests and reports back. You talk to it from your laptop or your phone. Your personal machine stays personal.

The 6 layers of the setup

The setup, layer by layer

1. Machine

  • A Linux VPS with no desktop environment : lighter, and a backup of the OS is a backup of your entire dev environment. Windows would also work, but I don’t recommend it.
  • Rough sizing: 4 vCPU / 8 GB RAM is already enough to run several agent instances in parallel.
  • Provider: I recommend Hetzner , specifically the Server Auction section (dedicated servers other customers cancelled; somewhat older hardware but very cheap; filter by budget, e.g. up to $100/month). A fixed monthly cost instead of the per-hour billing of Codespaces and the like.
  • Keep the configuration reproducible so you can tear the server down and bring up an identical one tomorrow (upgrade, migration, etc.).

2. Security

  • Never work as root.
  • Firewall on, limiting which devices can connect.
  • Tailscale as VPN / private network: installed on the server, the laptop and the phone, so only those devices can reach the server. No exposing password-based SSH to the whole internet.
  • The reason is obvious: your clients’ code and credentials live in there.

3. Agents

  • Install the CLIs you already pay a subscription for: Claude Code, Codex, OpenCode, Kimi Code… All on subscription, not API : much cheaper and a fixed cost.
  • The server needs access to your GitHub to clone repos. The agents themselves install Docker, Node, Go, Python… whatever each project needs.

4. Messaging layer

  • An intermediary like Hermes (or OpenClaw): installed on the server and connected to your ChatGPT subscription, it takes orders via Telegram, WhatsApp, Discord or Slack. Hermes can’t use your Claude subscription directly, but it can delegate to Claude Code if it’s installed on the machine.
  • His personal setup: several Telegram bots pointing at the same instance. One “personal assistant” bot (email, articles, tasks) and two or more “developer” bots that know where the projects live and always delegate the actual code writing to Claude or Codex. Having two dev bots lets him launch tasks on different projects in parallel from his phone.

5. Orchestration and terminal (advanced)

  • Herdr : a “tmux for agents”. It opens Claude Code, Codex, etc. in one terminal and, crucially, they talk to each other : Claude can spawn a Codex instance, resume it or read its output. It has a Telegram plugin.
  • Direct SSH access from mobile apps: Termius , Termux (Android), a-Shell (iOS), combined with Tailscale.
  • VS Code over Remote SSH when you need to see the code of a large project; execution still happens on the server.

6. Planning

  • With several projects and sub-agents you need a tracker ( Linear, Notion , Jira…) as the single source of truth: pending, in progress, reviewed by the client, reported bugs.
  • The agent can read and write it via MCP ; no need to fill it in by hand. This is what I like to call the “software factory” approach, as opposed to prompt-by-prompt vibe coding.

Pros and cons of moving your agents to a VPS

The upsides

  • 24/7 : a client reports a bug at night → a message from your phone → the agent fixes it.
  • Several projects in parallel on a single machine.
  • Total portability : travel with any underpowered laptop.
  • Agnostic : a new agent or model comes out, you install it on the server and you’re done.

The downsides

  • Complex projects still call for an editor to review things (solution: VS Code Remote SSH).
  • Port conflicts between projects on the same stack (everything wants 3000). He mentions Portless (named URLs instead of ports) as a mitigation.
  • E2E testing (Playwright, browsers) eats a lot of RAM and can slow down every agent at once.
  • Local models (he tried a DGX Spark, around $4,000): fine for landing pages and typical web apps, but they struggle with complex software; today they don’t replace cloud models.
  • It isn’t finished: he says himself he’s been tuning it for months and aims to have it polished by year end.

Who is it for?

  • Yes : if you juggle several clients or projects and ship changes daily (he works on 2–3 projects a day).
  • No : one project and one client; it’s overkill.
  • Ready-made alternatives: Cursor Cloud Agents and Claude Code on the web do the same (clone, change, open a PR), but they bill close to API prices and you don’t control the machine.

My takeaway

The interesting part isn’t “delegate to the cloud and the app builds itself”. It’s having an isolated, always-on environment you can talk to from anywhere, one that doesn’t depend on whichever agent is in fashion this month. If you work on a single project, don’t overcomplicate it. If you have several and already live between terminals, this is probably the next step.

Tools mentioned: Tailscale, Hetzner Server Auction, Portless, Herdr.

Top comments (0)