The best and cheapest VPS for Claude Code in 2026. We priced an always-on server on AWS, Azure, GCP, DigitalOcean, Vultr, Hetzner, Oracle's free tier and your own hardware: $0 to $263 a month.
Coding agents like Claude Code and Codex work best on a machine that stays on. Your laptop isn't that machine. It sleeps when you close the lid, it comes to meetings, and it runs out of battery on the train. So people start looking for a Claude Code VPS, a small server that stays on so the agent can keep working. The first answer is usually AWS.
We priced it out across every cloud we could think of. The same kind of always-on box costs anywhere from $0 to $263 a month, depending on where you rent it. Most of that spread pays for things a personal dev box never uses.
TL;DR: A 4-core box running 24/7 costs about $121 a month on AWS or Azure, $98 on Google Cloud, $40 to $48 on Vultr or DigitalOcean, and $9.99 on Hetzner. An old laptop costs you about $2 a month in power. Add a $20 AI plan and GetPullRequest (GPR) on your phone, and you have a coding agent that works while you're away for about $30 a month.
How much server does Claude Code need? #
When Claude Code or Codex "thinks", that work happens on Anthropic's or OpenAI's servers. Your machine just sends the request and waits. What it actually runs is git, your editor tools, your package manager and your tests. Anthropic lists the minimum for Claude Code as 4 GB of RAM on an x64 or Arm processor.
So the size of the box depends on your project, not on the agent. A Node or Python repo is happy on 2 cores and 4 GB. A Flutter app with an Android build, or three agents running test suites at the same time, wants 8 cores and 16 GB. Most people land at 4 cores and 8 GB, and that's what we compared.
| Your project | Box size that works |
|---|---|
| Scripts, small web apps, one agent | 2 vCPU, 4 GB |
| Typical app with a real test suite | 4 vCPU, 8 GB |
| Mobile builds, Docker, 2 to 3 agents at once | 8 vCPU, 16 GB |
The prices, one by one #
All prices are US dollars for a month of 730 hours, Linux, checked on 16 October 2026. Follow the links, because cloud prices change.
AWS, Azure and Google Cloud: what the hyperscalers charge
| Provider | Machine | Per hour | Per month |
|---|---|---|---|
| AWS EC2 | t3.xlarge, 4 vCPU, 16 GB | $0.166 | ~$121 |
| Azure | B4ms, 4 vCPU, 16 GB | $0.166 | ~$121 |
| Google Cloud | e2-standard-4, 4 vCPU, 16 GB | $0.134 | ~$98 |
Sources: AWS t3.xlarge, Azure B4ms, GCP e2-standard-4 (Vantage, us-east-1 / East US / us-central1).
That's compute only. On AWS you also pay $0.005 an hour for a public IPv4 address, which is $3.65 a month, and about $0.08 per GB-month for the disk, so another $8 for 100 GB. The real AWS bill is closer to $133. Data going out to the internet is billed on top.
A one-year reservation brings the t3.xlarge down to about $0.104 an hour (~$76 a month). That's still more than seven Hetzner boxes.
None of this is a rip-off. You're paying for dozens of regions, enterprise access controls and an uptime promise that production systems need. A box that runs your agent while you sleep needs none of it.
Lightsail, DigitalOcean and Vultr: the simpler clouds
| Provider | Machine | Per month | Traffic included |
|---|---|---|---|
| AWS Lightsail | 4 vCPU, 16 GB, 320 GB SSD | $84 | 6 TB |
| AWS Lightsail | 2 vCPU, 8 GB, 160 GB SSD | $44 | 5 TB |
| DigitalOcean Basic | 4 vCPU, 8 GB, 160 GB SSD | $48 | 5,000 GiB |
| Vultr Regular | 4 vCPU, 8 GB, 160 GB SSD | $40 | 4 TB |
| Vultr High Performance | 4 vCPU, 8 GB, 180 GB NVMe | $48 | 6 TB |
One flat price, with the disk, the IP address and a generous amount of traffic included. Lightsail is the same AWS underneath at about two thirds of the EC2 price. If you want AWS on the invoice, that's the one to pick.
Hetzner vs DigitalOcean vs AWS: the cheapest paid option
Hetzner raised cloud prices on 15 June 2026, some plans by more than double. It's still the cheapest paid option by a wide margin. On the same 4 vCPU, 8 GB box, Hetzner vs DigitalOcean is $9.99 against $48. Hetzner vs AWS is $9.99 against about $133.
| Plan | Machine | Per month (Germany/Finland) |
|---|---|---|
| CX33 | 4 vCPU, 8 GB, 80 GB NVMe | $9.99 (€8.49) |
| CAX21 | 4 Arm vCPU, 8 GB, 80 GB NVMe | $12.49 (€10.49) |
| CX43 | 8 vCPU, 16 GB, 160 GB NVMe | $18.49 (€15.99) |
| CPX32 | 4 vCPU, 8 GB, 160 GB NVMe | $41.99 (€35.49) |
Prices exclude VAT and a €0.50 public IPv4 fee. EU servers include 20 TB of traffic.
Two catches. The cheap CX and CAX plans are "cost-optimized", and on the day we checked, Hetzner's page listed them as currently unavailable. They come back in batches, so keep trying. And the cheapest plans live in Europe. That matters less than you'd think. Your phone talks to the agent through our server, and the agent talks to the model's servers, so you never type into a laggy terminal across the Atlantic.
Is Oracle Cloud's free tier really free?
Oracle's Always Free tier gives you Arm compute worth 2 OCPUs and 12 GB of memory, plus 200 GB of disk, for $0. It's a real machine and it's enough for one agent.
Yes, but read the limits. The catch is written in the same doc. Oracle may take back a free instance that looks idle for 7 days, meaning CPU, network and memory all under 20%. A dev box that waits for your next task looks exactly like that. Use it to try things. Don't build your week around it.
GitHub Codespaces
Codespaces charges $0.36 an hour for 4 cores. Used eight hours a day on workdays, that's about $63 a month. Left running all month so an agent can work overnight, it's $263. Codespaces is built to stop when you stop, which is the opposite of what an agent box is for.
Your own machine
| Machine | Up front | Power per month | Per month over 3 years |
|---|---|---|---|
| Old laptop or desktop you already own | $0 | ~$2 | ~$2 |
| Raspberry Pi 5, 8 GB | $155 | ~$0.70 | ~$5 |
| Mac mini M6 | $899 | ~$1.30 | ~$26 |
Power uses the US average of 18.2 cents per kWh for 2026. Apple says the M6 Mac mini draws 4 watts at idle, and we assumed an average of 10 W with builds. The Pi got 5 W and the old laptop 15 W.
The old laptop in the cupboard is the cheapest dev box there is. It already has your tools, your keys and a warm build cache. RAM prices pushed the Pi up 24% this year, but it's still a fine box for a light repo.
What it costs all in #
The box is only half the bill. The agent needs a model plan too. Claude Pro is $20 a month and includes Claude Code. ChatGPT Plus includes Codex.
| Setup | Box | AI plan | GPR | Total per month |
|---|---|---|---|---|
| AWS EC2, done "properly" | ~$133 | $20 | $0 | ~$153 |
| DigitalOcean | $48 | $20 | $0 | $68 |
| Hetzner CX33 | $9.99 | $20 | $0 | ~$30 |
| Old laptop at home | ~$2 | $20 | $0 | ~$22 |
| Oracle free tier | $0 | $20 | $0 | $20 |
Heavy users move to a bigger AI plan (Max starts at $100) long before they need a bigger box. That's where the money actually goes.
Our take #
We tried hosted sandboxes first. Wiring our backend, Flutter app, database and daemon together inside someone else's sandbox was painful, every test meant a fresh checkout and a full build, and the bill was too high for a small startup whose laptops sat idle most of the day. (The launch post tells the whole story.)
So our view is simple. Spend on the model, not the machine. The model is what writes the code. The machine just needs to be on, have your repo checked out, and run your tests. A $10 box does that as well as a $120 one.
What $30 a month gets you #
With a small always-on box and GPR on your phone:
- Agents keep working when your laptop is closed. Start a task on the train, close the app, check the result at lunch.
- Tickets turn into pull requests. Point an agent at a GitHub issue, a Jira ticket or a Slack thread, and a PR comes back.
- Several tasks at once. Each task gets its own git worktree, so two or three agents can work side by side on 8 GB.
- Overnight runs. Queue the boring backlog before bed and review the PRs from your phone in the morning.
- Your agent, your choice.Claude Code , Codex, Cursor or any of theeight ACP agents .
- Your code stays on a box you control, not in a vendor's sandbox.
How to install Claude Code on a VPS #
Rent the box (or dust off the laptop) and log in over SSH. Install the agent you use, then GPR. For Claude Code, that's the official installer, and you sign in by running claude once:
curl -fsSL https://claude.ai/install.sh | bash
claude
curl -fsSL https://getpullrequest.com/install | bash
gpr setup ~/code/your-repo
Scan the QR code with the GPR app and you're done. The daemon connects out to our server, so you don't need to open any ports. Keep SSH on keys only and leave the firewall closed.
A few habits that save money and pain:
- Start small. Moving up a size on most clouds takes a reboot, not a rebuild.
- Check
gpr statusandgpr doctorafter a reboot, the same as any other machine. - Watch the hourly clouds. A forgotten EC2 box with a big disk keeps billing whether you use it or not.
Questions people ask #
What's the best VPS for Claude Code?
For most people, a 4 vCPU, 8 GB box. Hetzner's CX33 at $9.99 is the cheapest we found. If it's out of stock, Vultr ($40) and DigitalOcean ($48) are simple and reliable.
What's the cheapest VPS for Claude Code?
Oracle's free tier at $0, if it doesn't get taken back for being idle. The cheapest paid one is Hetzner's CX23 (2 vCPU, 4 GB) at $6.49, which just meets Claude Code's 4 GB minimum.
How much RAM does Claude Code need?
Anthropic says 4 GB or more. Plan for 8 GB if your tests or builds are heavy, or if you run more than one agent at once.
Is Hetzner cheaper than AWS?
Yes, by about 13 times for an always-on 4-core box: $9.99 against roughly $133 with disk and IP address on AWS.
Hetzner vs DigitalOcean pricing?
For 4 vCPU and 8 GB, Hetzner is $9.99 and DigitalOcean is $48. Hetzner's cheapest plans are only in Europe, and they sell out.
Can I run Claude Code on a Raspberry Pi?
Yes. Claude Code supports Arm, and the 8 GB Pi 5 meets the memory minimum. It's slow for big builds but fine for a light repo.
Does my Claude Code VPS need open ports?
Not for GPR. The daemon connects out to our server, so only SSH needs to be reachable, and only with keys.
Try it #
GPR is free for projects, sessions, and the GitHub and Slack connections. You pay only for more workspaces or more tasks running at once. Get the app on Google Play, pick the cheapest box on the chart above, and give it tomorrow's longest task.
Keep reading #
- Ship code from your phone with the AI coding agents you already pay for
- Your AI coding agent is stuck in a terminal you have to babysit. So we built GetPullRequest
- Claude Code Remote Control is great. It's also not enough
- Agent Client Protocol (ACP) is the most important AI coding standard nobody's talking about
Sources #
Prices checked 16 October 2026.
- AWS: EC2 t3.xlarge ,Lightsail pricing ,public IPv4 pricing ,EBS pricing
- Azure: B4ms
- Google Cloud: e2-standard-4 ,Compute Engine pricing
- Hetzner: June 2026 price adjustment ,cost-optimized plans ,regular performance plans
- DigitalOcean: Droplet pricing
- Vultr: pricing
- Oracle Cloud: Always Free resources
- GitHub Codespaces: billing
- Mac mini M6: TechCrunch, $899 starting price ,Apple power figures
- Raspberry Pi 5: OMG! Ubuntu, February 2026 price rise
- Claude Code: setup and system requirements
- Electricity: EIA Short-Term Energy Outlook, October 2026
- AI plans: Claude pricing ,ChatGPT pricing