# Self-Hosted vs Managed Claude Code: An Honest Comparison

> Self-host Claude Code on a VPS or pay for managed hosting? A straight comparison of setup, security, persistence, and the real cost of your own time.

- Published: 2026-07-17 · Updated: 2026-07-17 · jurniti
- Canonical: https://www.jurniti.com/blog/self-hosted-vs-managed-claude-code

You can run [Claude Code](/claude-code) on a $6 VPS. Plenty of people do, and for some of them it's the right call. The question isn't whether self-hosting works — it's what the sticker price leaves out, and whether the thing you're actually buying (an always-on agent that survives your laptop closing) is cheaper to rent or to build. Here's the comparison we'd want if we were on the other side of it, including the cases where you should not pay us.

## What "self-hosting Claude Code" actually means

Claude Code is a thin terminal client — the model work happens on Anthropic's side — so the requirements look trivial: a Linux box, Node, a shell. The real build list is longer:

- Rent a VPS and create a non-root user.
- Harden SSH: key-only auth, no password login, ideally a non-standard port and fail2ban, because that port now faces the internet forever.
- Install Node and the Claude Code CLI, authenticate with your subscription or an API key.
- Set up tmux (or zellij) so a dropped WiFi connection doesn't kill a three-hour agent run.
- Write a systemd unit if you want anything supervised across reboots.
- If you want `--dangerously-skip-permissions` — the mode that makes a remote agent genuinely useful — wrap it in a container or VM first. Anthropic's guidance is blunt: bypass mode belongs inside isolation.

First time through, that's a weekend. Second time, an afternoon. None of it is hard for someone who runs servers; all of it is friction for someone who just wants the agent working tonight.

## The part nobody prices in

The VPS is $5–10/mo. That's not the cost. The cost is that you've taken a standing job: OS patches, Node upgrades, the CLI's update cadence, the morning you discover the tmux session died at 2am and the overnight run with it. Each item is small. The sum is a background process in *your* head, running forever.

There's also an exposure question that compounds quietly. Your Claude login or `ANTHROPIC_API_KEY` lives on that box. The box has an open SSH port, and you are its security team. And on budget hosting, "your VPS" is often a container on a shared host kernel — which matters more than usual here, because the workload is an agent executing code no human reviewed, steered by input you don't fully control. Researchers showed in January 2026 that hidden text in a `.docx` could prompt-inject Claude Code into exfiltrating files. On a DIY box, the blast radius of that class of attack is everything the box can reach. (The full argument for why the isolation boundary matters is in [microVM vs Docker for agent hosting](/blog/microvm-vs-docker-agent-hosting).)

## What managed hosting actually buys

jurniti's version of "managed" is narrow on purpose: we run the box, you keep the agent.

- **A dedicated Firecracker microVM** — its own kernel, a KVM hardware boundary, one tenant per box, booting in about 124ms. Bypass mode is safe here by construction: root inside the VM is root over one disposable machine, not yours.
- **BYOK as architecture, not policy.** You run `claude /login` inside your own VM or paste an API key there. Traffic leaves straight for Anthropic. We are never in the path, so there's nothing for us to proxy, log, or mark up.
- **A persistent home.** Your `CLAUDE.md`, slash commands, skills, subagents, and MCP config survive restarts and reprovisions.
- **An in-browser terminal.** Nothing to SSH into means no SSH surface to harden — the single biggest DIY exposure simply isn't there.
- **Templates and forking.** Snapshot a dialed-in setup into a [template](/templates) and fork it for the next project; credentials are scrubbed on publish.
- **~3 minutes from payment to working terminal**, no human in the loop.

Flat [$25–99/mo](/pricing), with a 30-day money-back guarantee on your first purchase.

## When self-hosting is the right call

Credibility requires this section, so here it is, unhedged:

- **You already run infrastructure.** If there's a Proxmox box in your closet or you operate servers professionally, one more service is marginal work and DIY is genuinely cheaper.
- **You want to learn the stack.** Setting this up by hand teaches you SSH hardening, process supervision, and sandboxing. That's a good weekend.
- **Policy requires your own account.** Some employers and clients mandate that code and credentials stay in infrastructure they control. Managed hosting can't satisfy that, ours included.
- **You need hardware we don't offer.** GPUs, a specific region, a beefy bare-metal box — go where the hardware is.
- **Budget is the real constraint.** If $25/mo is material and your evenings are free, a $6 VPS plus your labor is a fair trade. Just wrap bypass mode in a container.

If two or more of those apply, self-host. The rest of this page is for the people nodding at none of them.

## Self-hosted vs managed, side by side

| | DIY VPS | jurniti managed |
|---|---|---|
| Time to working agent | A weekend, then an afternoon each time | ~3 minutes, automatic |
| Monthly cost | $5–10 + your maintenance time | $25 flat |
| Blast radius of a bad command | The whole server | One disposable microVM |
| Bypass mode | Risky unless you wrap it yourself | Safe — KVM-isolated by default |
| Session persistence | tmux you configure | Persistent home + durable terminal |
| Attack surface | Open SSH port, you patch it | In-browser terminal, no SSH exposure |
| Claude credentials | On a box you harden | In your VM, never proxied |
| Refund | None | 30-day money-back |

## Pricing

Starter is **$25/mo** ($250/year — two months free) and comfortably runs Claude Code; it's a thin client, so 1 vCPU / 2 GiB is plenty. Pro is **$49/mo** and Max **$99/mo** for heavier workspaces or a custom TLS subdomain. Every plan carries the 30-day money-back guarantee on your first purchase, and your Claude subscription or API spend stays yours — separate, direct to Anthropic, never marked up.

More background: the full [Claude Code hosting guide](/blog/claude-code-hosting), the [isolation deep-dive](/blog/microvm-vs-docker-agent-hosting), or the [pricing page](/pricing).

## Frequently asked questions

### Is it cheaper to self-host Claude Code on a VPS?

On sticker price, yes — a small VPS runs $5–10/mo versus $25/mo for a managed microVM. But the VPS number excludes your time: initial setup, SSH hardening, tmux/systemd supervision, OS patching, and rebuilding when something breaks. If you already run servers for a living, self-hosting is genuinely cheaper. If an evening of your time is worth more than $20, it usually isn't.

### What do I need to self-host Claude Code?

A VPS with Node.js, the Claude Code CLI, a non-root user, hardened SSH (key-only auth), and tmux or a systemd unit so sessions survive disconnects. If you want to run --dangerously-skip-permissions, you also need a container or VM wrapper — Anthropic's own guidance is that bypass mode belongs inside isolation, not on a raw box.

### Is running Claude Code on a cheap VPS safe?

It's a trade. Your Claude credentials live on a box whose SSH port faces the public internet, and you are the security team. Budget VPSes are also often containers sharing a host kernel with other tenants — a thin boundary for an agent that executes code no human reviewed. A dedicated Firecracker microVM gives the agent its own kernel, so a bad command or a prompt-injection attack is contained to one disposable box.

### Can I use my Claude subscription with managed hosting?

Yes. jurniti is bring-your-own-key by architecture: open the in-browser terminal and run `claude /login` with your Claude Pro or Max plan, or export an ANTHROPIC_API_KEY. The credential lives only inside your VM, traffic goes straight to Anthropic, and jurniti never proxies it or marks up your usage.

### When is self-hosting Claude Code the right call?

When you already operate servers and one more box is marginal work; when you want to learn the stack; when policy requires your own cloud account; or when budget is genuinely the constraint and your time is free. Two or more of those and DIY is a defensible choice — the managed pitch is for everyone else.

### How fast is a managed Claude Code box ready?

About 3 minutes from payment to a working terminal. The Firecracker microVM itself boots in roughly 124ms, Claude Code is pre-installed on the PATH, and there's a 30-day money-back guarantee on your first purchase if it's not what you wanted.