Long agent runs have a boring failure mode: you close the laptop and the session dies. Claude Code, the Cursor CLI, and the Codex CLI all run in a terminal, so you can start them on a small Linux VPS over SSH and reattach later.

Why a server at all?

• The agent keeps going while your laptop sleeps or you are offline. Reattach over SSH.
• Isolation: shell access on the VPS instead of on a personal machine full of browser logins.
• Phone control: Claude Code supports Remote Control (paid plans; see below).

A VPS does not change model costs. You still need a subscription or an API key. Use a cost calculator if you want monthly tool spend side by side.

How much server do you need?

Claude Code’s docs want Ubuntu 20.04+ / Debian 10+ / Alpine 3.19+, at least 4 GB RAM, x64 or ARM64, and a network. The models run in the vendor cloud; RAM is mostly for your project (build, tests, DB, dev server).

Example entry-level Contabo VPS (site prices, retrieved 2026-10-27; promo for first 24 months):

Setup fees and post-promo prices: check the order flow. Source: contabo.com/de/vps (no tracking link in this cross-post).

Step 1: Prepare the box

sudo apt update && sudo apt install -y git tmux
tmux new -s agent      # start
# … later …
tmux attach -t agent   # reattach

tmux is the important bit: the session survives disconnects. Anthropic also recommends tmux/screen when you start Claude on a remote machine.

Step 2: Claude Code — install & login

curl -fsSL https://claude.ai/install.sh | bash
claude --version
claude doctor

• Interactive SSH: claude → press c to copy the login URL, open it locally, paste the code back.
• Scripts: claude setup-token → year-long OAuth token as CLAUDE_CODE_OAUTH_TOKEN (or use ANTHROPIC_API_KEY).
• Headless: claude -p "…".
• Remote Control (phone): claude remote-control / /remote-control — Pro/Max/Team/Enterprise only; needs a normal claude auth login, not API key / setup-token. Outbound HTTPS only.

Free plan does not include Claude Code.

Step 3: Cursor CLI

curl https://cursor.com/install -fsS | bash
NO_OPEN_BROWSER=1 agent login   # prints the login URL
agent status
# headless / plan modes: agent -p … · --mode=plan|ask
# background → Cloud Agents: agent &  then agent ls / agent resume

CURSOR_API_KEY works for non-interactive use. Docs: cursor.com/docs/cli.

Step 4: Codex CLI

curl -fsSL https://chatgpt.com/codex/install.sh | sh
codex login --device-auth   # device-code login (beta; enable in ChatGPT security settings)
# SSH tip: tunnel local port 1455 if the flow needs it
codex exec "…"
codex resume
codex login status

Treat ~/.codex/auth.json like a password. OpenAI warns not to expose Codex execution in untrusted/public environments. API path: --with-api-key (API pricing).

Security checklist

• Separate non-root user; SSH keys only; no password auth if you can.
• Secrets in env / secret store — not in the repo the agent can cat.
• Firewall: only SSH (and what you deliberately open).
• Don’t point an agent with broad shell access at production credentials.
• Prefer tmux so a dropped SSH session does not orphan a half-finished run without a way back.

What does it cost overall?

VPS (a few €/month) plus the tool plan or API spend. The VPS is usually the smaller line item; token/subscription cost dominates. Measure before stacking Ultra + Max “just in case”.

Verdict

A small Linux VPS, a dedicated user, tmux, and the right headless login path (NO_OPEN_BROWSER=1, device-auth, setup-token / OAuth) are enough to keep coding agents alive when the laptop is closed. Start with one tool you already pay for, confirm login over SSH once, then automate.

Sources

• Claude Code · Setup, authentication / remote-control docs (retrieved 2026-10-27)
• Cursor CLI · Overview & authentication
• Codex CLI · install & auth
• Contabo VPS (prices change — verify at checkout)

I write Syntaxlabor, a free German-language digest on AI coding tools (Cursor, Claude Code, agents) for indie devs, 3× a week. Archive without signup: syntaxlabor.de/digest.

Both tools promise the same thing: type less, ship more. The difference is the interface and the control model. Cursor is an AI-native IDE; Claude Code is a coding agent that grew up in the terminal.

Since Cursor's new pricing there is a second, often overlooked difference: where your Claude usage gets billed. A lot of older comparisons still assume "Cursor Pro $20 vs Claude Pro $20". That is no longer what the Cursor pricing page says.

TL;DR

What is what?

Cursor

An AI-native IDE with Tab completion, Agent, Cloud Agents and its own CLI. Plans according to cursor.com/pricing:

• Free $0
• Plus $8: "Only Grok models are available on this plan"
• Premium $30: 4× the usage of Plus, Cloud Agents & Automations, code hosting, AI code reviews via Bugbot, third-party models at API list price
• Super $100 (15×) and Ultra $200 (50×)

The usage pool refills weekly. Cursor itself recommends Premium for new AI users and Plus for people with light AI usage.

Claude Code

Anthropic's coding agent. It runs in the terminal, as a native extension for VS Code (and, per the product page, inside Cursor) and JetBrains, on the web, in the desktop app, on iOS/Android and in Slack. According to claude.com/pricing, Claude Code is included in all paid plans (not the free plan) and shares the usage pool with chat. Alternatively you can use a Console account with pay-as-you-go API tokens.

Head-to-head

Sources: Cursor Pricing, Cursor Worktrees, Claude Pricing, Claude Code, Claude Code common workflows. Prices change, so always check the live pages before you buy.

When Cursor is the better choice

• You want to work in the diff right away: Tab-accept, inline edits, review inside the editor.
• You think in files and tabs, not shell sessions.
• You're fine with Grok, in which case Plus ($8) is the cheapest entry in this comparison.
• You want (from Premium) Cloud Agents, Bugbot reviews and models from several vendors in one tool, and you accept per-token cost for third-party models.

Typical vibe flow: prompt → Cursor edits → you review the diff → commit.

When Claude Code is the better choice

• You let agents run longer, from issue to pull request, including tests.
• You live in the terminal (git, tests, deploy scripts), or you want to wire Claude into scripts and CI via claude -p.
• You use Claude models anyway: on Pro the usage comes out of your plan's pool instead of per-token in Cursor.
• You want parallel sessions without collisions: claude --worktree feature-x gives each session its own checkout.

Typical flow: describe the task → check the plan in plan mode → the agent implements and tests → you review commits/PRs. For side questions there's /btw (now also in the Cursor CLI, by the way).

Costs, honestly calculated

• Cursor Plus only ($8): cheap, but Grok-only. If your workflow needs Claude, Plus won't do.
• Cursor Premium only ($30): Claude & GPT in the editor, but on top, at API list price. On big refactors the token spend can overtake the plan price.
• Claude Pro only ($20, or $17 yearly): Claude Code included, usage from the plan pool. Strong for agent work, but not a full AI IDE.
• Hybrid: Cursor Plus + Claude Pro = $28/month; Cursor Premium + Claude Pro = $50/month (+ third-party tokens in Cursor if you use them). Only worth it if your output measurably goes up.
• The hybrid trick: according to Anthropic, Claude Code also runs as an extension inside Cursor. Use Cursor Plus ($8, Grok) as your editor and do the Claude work through Claude Code on Pro, and you pay no Cursor token money for Claude.

Practical recommendation: test for two weeks

1. Week 1: Cursor Free or Plus. Does IDE-AI actually speed you up?
2. Week 2: one month of Claude Pro ($20). Do one bigger refactor with Claude Code only.
3. Decide: mostly UI and small diffs → keep Cursor; mostly agent/repo tasks → prioritize Claude Code; both daily → hybrid with a hard monthly budget.

Whichever tool you pick: learn isolation with worktrees. Both tools now use separate checkouts instead of one agent working on main.

Common mistakes

• Upgrading a plan without checking the token spend of the last few weeks.
• Using Claude inside Cursor Premium without a spending cap, then seeing the API bill.
• Letting agents work directly on prod branches without review.
• Forgetting Privacy Mode or the training opt-out on client code.
• Two power-tier subscriptions (Ultra + Max) "out of FOMO". Measure first.

Verdict

There's no universal winner. Cursor wins as the everyday IDE, Claude Code as the agentic teammate in your repo. The most important point in 2026, though, is money: Claude models are not covered by a Cursor subscription, but they are in Claude Pro. For many indie devs the best answer is therefore editor + terminal agent with a clear monthly budget.

Sources

• Cursor · Pricing, retrieved 2026-10-19
• Cursor Docs · Worktrees
• Cursor CLI Changelog
• Claude · Pricing, retrieved 2026-10-19
• Claude Code · Product page
• Claude Code Docs · Common workflows
• Claude Code Changelog

I write Syntaxlabor, a free German-language digest on AI coding tools (Cursor, Claude Code, agents) for indie devs, 3× a week. If you read German, the archive is open without signup: syntaxlabor.de/digest. No affiliate links in this post.