VPS for Claude Code and Codex: which size and which provider
What Claude Code and Codex need from a server, measured, and what 4 GB and 8 GB cost at OVHcloud, Hetzner, Vultr, DigitalOcean and others in October 2026.
Both makers put the floor at 4 GB of memory, and OpenAI recommends 8 GB for Codex. For both agents on one server, with the builds and tests they run, get 4 vCPU and 8 GB; CPU speed and disk matter less, because the model runs on Anthropic’s and OpenAI’s servers. As listed on October 14, 2026, before tax and with no long commitment, that size costs €8.49 a month at OVHcloud, $40 at Vultr, and $48 at DigitalOcean and Akamai. The roughly $5 Hetzner server older advice points to is listed as “not available”, and Hetzner’s orderable 4 vCPU, 8 GB plan is €35.49.
What the two agents actually use
Anthropic’s setup page lists “4 GB+ RAM, x64 or ARM64 processor” for Claude Code. OpenAI’s install notes give Codex “4-GB minimum (8-GB recommended)”. Neither gives a disk figure, or one for running both. Installed side by side on an Ubuntu 24.04 server in October 2026:
| Claude Code 2.1.299 | Codex 0.162.1 | |
|---|---|---|
On disk | 254 MB per version; the installer keeps the two newest | 432 MB from npm, plus 432 MB for the background service it downloads on first start |
Memory | 394 to 573 MB for each working session | About 420 MB idle, across four processes |
After you quit | Nothing left running | Its background service stays, about 216 MB; codex app-server daemon stop ends the service, though its small helper process kept running after it |
Linux builds | x64 and arm64 | x86_64 and arm64 |
So 4 GB runs either agent, or both while they are only talking to you. What fills the rest is the work they start: a dependency install, a test suite, a Docker build, a browser. Claude Code warns when a session’s heap passes 2.5 GB, and installing it “needs roughly 512 MB of free memory”; on a small server the install is the usual casualty, killed by the out-of-memory killer. Anthropic’s install troubleshooting gives adding swap as the first remedy, and a 2 GB swap file is cheap insurance on any size. Both agents ship arm64 builds, so an Arm server works too.
What a server costs, October 2026
Monthly prices as each provider listed them on October 14, 2026, before tax, for the cheapest plan of each size:
| Provider | 2 vCPU, 4 GB | 4 vCPU, 8 GB | Backups | Watch for |
|---|---|---|---|---|
OVHcloud | €4.49 / $5.35 | €8.49 / $10.00 | Daily, last 24 hours, included | The “from” price on its page (€3.81, €7.21) is for 12 months paid upfront |
Hetzner (EU) | CX23 €5.49 / $6.49, marked not available; CPX22 €19.49 / $22.99 | CX33 €8.49 / $9.99, marked not available; CPX32 €35.49 / $41.99 | 20% of the server price | Plus €0.50 / $0.60 for an IPv4 address |
Hetzner (US) | CPX21, 3 vCPU: $37.49 | CPX31: $73.49 | 20% | 2 to 3 TB of traffic on these plans, against 20 TB in the EU |
Vultr | $20 | $40 | Plus 20% | |
DigitalOcean | $24 | $48 | Plus 20% weekly, 30% daily | Billed per second, capped at the monthly price |
Akamai (Linode) | $24 | $48 | Plus $5 / $10 | Dearer in Jakarta and São Paulo |
AWS Lightsail | $24 | $84 (its $44 plan has 8 GB but 2 vCPU) | Snapshots $0.05 per GB | Burstable CPU |
Hostinger | KVM 1 is 1 vCPU: $6.49 on 24 months upfront | KVM 2 is 2 vCPU, 8 GB: $8.99 on 24 months upfront | Weekly free; daily $3 / $6 | Renews at $11.99 / $14.99; month to month, KVM 2 is $13.99, then $24.49 |
Two readings of that table matter more than the cheapest cell. A long prepaid term is a bet that the setup will still suit you in two years, and Hostinger’s terms reserve the right to change “Services’ renewal fees displayed at this Site at the time of your first purchase”. And the core count differs between plans that look alike: Hostinger’s KVM 1 has one vCPU, and its KVM 2 and Lightsail’s $44 plan have two each beside their 8 GB.
Hetzner in 2026
Hetzner was the usual answer for a cheap agent server, and its 2026 prices are why that advice has aged. Its CX23, 2 vCPU and 4 GB, cost €2.99 ($3.49) a month until April 1, 2026, when a rise for new and existing servers took it to €3.99 ($4.99), citing RAM and SSD costs. On June 15 a second rise for new orders took it to €5.49 ($6.49). Today the whole cost-optimized line, its x86 CX and Arm CAX plans, is marked “not available”, and Hetzner says of that line that “the number of available servers is limited”.
What can be ordered is the regular-performance line, which the June rise hit hardest: the CPX22 went from €7.99 to €19.49, and the 4 vCPU, 8 GB CPX32 from €13.99 to €35.49. At the start of the year they were €5.99 and €10.49. A server bought before June 15 keeps its price until you rescale it or restore it after deletion. Hetzner’s US locations offer only the older CPX plans, from $37.49 for 3 vCPU and 4 GB.
Do the templates save anything?
A few providers sell an image with an agent on it. Hostinger’s “Ubuntu 24.04 with Claude Code” template comes with Claude Code installed; it has no Codex template, and its help article does not say how Claude Code was installed or whether it updates itself. DigitalOcean’s Codex 1-Click image is set up to send Codex to DigitalOcean’s own paid model service, with your ChatGPT sign-in optional; it runs Codex as root, suggests a 1 GB minimum where OpenAI says 4 GB, and its listing showed version 0.154.0 when 0.162.1 was current. Vultr, Lightsail, Hetzner and OVHcloud have no image for either agent.
What an image saves is one command. Anthropic’s installer put Claude Code on the server above in 11 seconds and npm installed Codex in 6. The parts that take the time are the same on any image: signing in on a machine without a browser (Claude Code has you paste a code back; Codex uses device-code sign-in, which you switch on first in ChatGPT’s settings), keeping the agents updated, and securing the server and the keys the agents hold.
Three things that bite on a rented server
- Outgoing mail is blocked. Hetzner blocks ports 25 and 465 on its cloud servers, DigitalOcean blocks 25, 465 and 587, and Vultr blocks 25. An agent that should email you needs a mail service on 587 or an HTTPS API; the notifications guide covers the options.
- The provider’s rules apply to what the agent does. Hetzner’s cloud terms prohibit “The scanning of foreign networks or foreign IP addresses”, and DigitalOcean’s acceptable use policy has a section on AI systems. An agent asked to probe hosts you do not control is breaking your provider’s terms, on your account.
- Free tiers can reclaim an agent that waits. Oracle Cloud’s Always Free Arm allowance (2 OCPUs and 12 GB) runs both agents, but Oracle may reclaim an instance whose CPU, network and memory use stay under 20% over seven days. That describes an agent waiting for its owner.
A cloud computer with the agents already installed
The other way to skip the setup is a machine built for coding agents. The guide to servers with Claude Code and Codex installed compares the field. Everpod, which publishes this guide, sells one: its developer pod is a cloud computer of your own with your pick of Claude Code, Codex, OpenCode, Pi, Hermes and OpenClaw installed, from $24 a month. A developer pod comes in three sizes: S with 4 vCPU, 8 GB of memory and 80 GB of disk, room for about 8 coding agents working at once; M with 8 vCPU, 16 GB of memory and 160 GB of disk, about 16; L with 16 vCPU, 32 GB of memory and 320 GB of disk, about 32. A developer pod is one machine that stays yours: what you installed, the servers you left running, Docker and your work in progress are where you left them tomorrow. It accepts no incoming connections from the open internet: no open ports and no public SSH. A server your agent starts is reachable from your devices and from nowhere else. By default the machine does not start connections to your other devices. The whole disk is backed up daily, and the backups rotate over roughly 7 days. You sign in to each agent with your own subscription or API key. No model usage is included, and your sign-ins never pass through us. You still log in over SSH and run the agents yourself; what it saves is choosing and provisioning the server, the installs, and the server’s side of the security.