New

Your agent shouldn't stop because you closed your laptop.

Agent VPS is a machine built for AI agents that run for hours at a time. The model and the API key stay yours. We do one thing: keep the box up, and make it easy to get back when your agent breaks it.

Run Claude Code, Codex, or your own agent loop. Linux if you want Linux, Windows if you bring a licence, and if you need macOS, ship us a Mac mini and we'll rack it.

Tell us what your agent does See specs and pricing

Taipei facility, on our own AS152626 network. Bandwidth isn't metered — your agent can pull as many repos and images as it likes.

The model is yours. We only sell the machine.

Plenty of "AI hosting" is really token resale with a markup. This isn't.

Bring your own LLM endpoint

Agent VPS includes no inference credit of any kind. Your Anthropic, OpenAI, Google, AWS Bedrock, or self-hosted vLLM / Ollama endpoint — the API key lives on your machine and the traffic goes straight out from it.

That's a deliberate split, not a missing feature. It means we never hold your API key, never see your prompts, and never touch your source. Switch model providers without switching hosts, and we have no incentive to sit in the middle taking a cut.

Want to run models locally instead? That's what the accelerator and bring-your-own-GPU options further down are for.

Why not just rent an ordinary VPS

You can, and most people start there. But agents are a different workload from web servers, and they fail in different ways.

It runs long, and it can't drop

A three-hour agent run is normal. Laptop sleeps, SSH drops, the corporate VPN kicks you — and the work is gone. We ship tmux and systemd persistence configured, so a dropped connection just means you stopped watching.

It eats bandwidth

An agent can clone dozens of repos, pull a pile of container layers and walk the whole of npm and PyPI in an afternoon. On metered egress that's an ugly invoice. We bill on port rate, so pulling more costs the same.

It will break the system eventually

Give an agent a shell and sooner or later it deletes the wrong file, installs conflicting dependencies, or wrecks the Python environment. Seven daily snapshots; a rollback takes about a minute. Cheaper than an afternoon of debugging.

It needs a fence around it

A machine with internet access hands your agent unlimited outbound reach. We can put a network-layer egress allowlist in front of it, permitting only the domains and prefixes you name. Security review tends to go better with one.

Four kinds of machine

From a virtual machine to shipping your own hardware to Taipei.

Linux VPS

The usual choice. KVM, root access, install whatever you want. Fits almost any agent that writes code, gathers data, or drives an automation pipeline.

  • OSUbuntu, Debian, Rocky, or an image you specify
  • SizeFrom 4 vCPU up to 32 vCPU / 128GB
  • StorageNVMe, additional volumes on request
  • NetworkStatic IPv4 and native IPv6

Windows VPS (BYOL)

Some toolchains are Windows-only — .NET desktop builds, Office automation, certain industrial and financial software under test. We supply the machine; you bring the Windows licence.

  • LicensingYour Windows Server or VDA licence; we don't resell
  • AccessRDP, restrictable by source address
  • FitsWindows-only builds and automated testing
  • NoteLicence compliance is yours; we'll supply machine details for audits

Mac mini colocation

Apple's licence terms keep macOS on Apple hardware, so colocation is the only legitimate route to a 24/7 macOS environment. Buy the machine, ship it to us, we rack it and give it power and a network port.

  • ModelsM4 / M4 Pro, M5 Pro, or M6
  • FitsXcode builds, iOS CI, macOS-only toolchains
  • Local inferenceUnified memory suits MLX workloads well
  • RemoteKVM over IP or a smart PDU for hard reboots

Accelerators and your own GPU

Two routes if you want local inference on hardware you own: ship us an NVIDIA DGX Spark to colocate, or have us fit your GPU into a dedicated server.

  • DGX SparkGB10, 128GB unified memory, 240W PSU, 10GbE
  • Footprint150 × 150 × 50.5 mm, 1.2 kg — barely any rack space
  • Your GPUSingle or multi-card, subject to power and thermal review
  • Why not at homePower bills, heat, noise, and your upstream link

Who this fits — and who it doesn't

As always, the second list first saves everyone time.

Good fit

  • Letting a coding agent work through a backlog overnight
  • Long-running data collection, enrichment and batch pipelines
  • A machine that never sleeps to catch webhooks or run schedules
  • A persistent build machine for iOS and macOS projects
  • Running local models on hardware you own, in a proper facility
  • Teams tired of hyperscaler egress invoices
  • Overseas teams wanting an APAC box without a Taiwanese entity

We can't help with

  • Hourly GPU compute from a shared pool — we don't operate one
  • Bundled model credits or resold API tokens
  • Large multi-node training clusters
  • Agents used for scraping abuse, account farming or attacks

For the first three we'll say so plainly and point you at someone who does them. The fourth we decline.

Specs and pricing

Standard sizes are listed. Colocation is quoted by model and power draw. Prices exclude tax.

PlanSpecification and fitPrice
Agent VPS StarterLinux4 vCPU / 8GB / 160GB NVMe. A single agent, solo developerNT$1,200/ month
Agent VPS StandardLinux8 vCPU / 16GB / 320GB NVMe. Several agents in parallel, small teamNT$2,400/ month
Agent VPS ProLinux16 vCPU / 32GB / 640GB NVMe. Heavy batch work, build pipelinesNT$4,600/ month
Windows add-onBYOLWindows on any tier above; licence supplied by you+NT$300/ month admin
Mac mini colocationM4 / M5 Pro / M6Rack space, power and network. You buy and ship the machineNT$2,800/ month and up
DGX Spark colocationGB10Rack space, power and a 10GbE portNT$3,500/ month and up
Your own GPUDedicated serverYour card in our chassis, subject to power and thermal reviewQuotedtell us the model first

Bandwidth is not metered; you're limited by the port rate on your plan. Instances resize upward without a rebuild. Colocation requires a separate equipment custody agreement, which comes with the quote.

From enquiry to running

Virtual machines go out the same day. Colocation depends on when your hardware lands.

Describe the workload

What the agent does, how many run at once, whether you need an egress allowlist. If you don't know the sizing, we'll estimate it.

Same-day reply

Provisioning

VMs are live the day you pay — you get the address, credentials and snapshot notes. Colocated hardware is racked, powered and tested on arrival.

VPS same day / colo 1–3 days

An engineer stays on it

Resizing, firewall changes, snapshot rollbacks — just ask. We monitor the hosts, so if yours falls over we usually know first.

24/7 for incidents

Frequently asked questions

What people ask before committing.

How is Agent VPS different from an ordinary VPS?
Same virtualisation underneath; the difference is the default configuration and how bandwidth is billed. Agent VPS ships with persistent execution set up (tmux, systemd), seven daily snapshots, an optional network-layer egress allowlist, and port-rate rather than metered bandwidth. All of that targets how agents actually behave: long runs, heavy external pulls, and occasionally destroying the filesystem.
Do you provide an LLM or API credits?
No, deliberately. Agent VPS includes no inference credit — bring your own endpoint. That means we never hold your API key, never see your prompts, and never mark up model pricing. You can change providers without changing hosts.
Can I run Claude Code, Codex, or my own agent framework?
Yes. It's a standard Linux machine with root. Anything that runs on Linux installs normally. You supply your own API key and traffic goes directly from the machine to your model provider.
Can you see my code and data?
We're an infrastructure provider, so we hold administrative access to the physical hosts — that's true of every hosting company and it's better said plainly. In practice we don't enter customer VMs unless you open a ticket asking us to. You can run full-disk encryption with keys you hold, and we'll sign an NDA. If the data is highly sensitive, colocating your own hardware is the cleaner answer.
What if the agent destroys the system?
Roll back a snapshot. We take one daily and keep seven; a rollback takes about a minute. You can also ask for an on-demand snapshot before a risky run.
Is bandwidth really unmetered?
Yes — you're limited by the port rate on your plan, with no egress invoice. Heavy repo cloning, container pulls and model weight downloads are all in scope. We can price this way because we run the network ourselves (AS152626), which is a different cost structure from reselling a hyperscaler. The limit is that it isn't for sustained saturated transit or abuse.
Can I restrict what the agent is allowed to reach?
Yes. We can apply a network-layer egress allowlist permitting only the domains or prefixes you nominate — your Git server, model API and package registries, for instance — and dropping everything else. For teams who need to clear a security review, this is often the deciding factor.
How does Windows licensing work?
Bring your own licence. We provide the machine and management; you obtain the Windows Server or VDA licence and remain responsible for compliance. We'll supply hardware and tenancy details if you need them for an audit. We don't resell Microsoft licensing.
Why colocate a Mac mini instead of renting one from you?
Apple's licence terms restrict macOS to Apple hardware, and we don't offer time-shared Macs. So you buy the Mac mini — M4, M5 Pro or M6 — and ship it to our Taipei facility. We rack it, give it power and a network port, and provide remote reboot. Ownership stays with you throughout.
What should I know before shipping a DGX Spark?
It's tiny — 150 × 150 × 50.5 mm and 1.2 kg — but it ships with a 240W external supply and the GB10 carries a 140W TDP, so sustained load needs thought about placement and airflow; we'll handle that. Networking is 10GbE plus ConnectX-7, which drops straight into a rack switch. Confirm power specification and delivery timing with us before shipping.
Can I put my own GPU in one of your servers?
Yes, but tell us the model first. What matters is per-card power draw, thermal design and whether it physically fits — with large consumer cards in standard rack chassis, it's usually a mechanical problem rather than a technical one. Send us the model and we'll come back with a workable chassis and a price.
Will the outbound IP get blocked as a bot by Cloudflare or GitHub?
Every datacentre IP carries that risk, and anyone promising otherwise is overselling. What we can offer: addresses from our own space where we control who the users are, not an anonymous recycled pool — and when something is misclassified, we take it up with the IP intelligence vendors and the service directly. If your workload depends heavily on browsing ordinary websites, test during evaluation; we'll hand you an address to try.

Tell us what your agent does

No specification needed. The workload, how long it runs, and any security constraints is enough to start.

Sales and technical

Email
tech@cloudbase.twPreferred — describe the workload and we can scope it
Phone
+886 971 726 606Mon–Fri 09:00–18:00 UTC+8
Telegram
@endichanFastest for urgent matters
Portal
shop.cloudbase.twExisting customers — invoices and tickets

About Cloudbase

Entity
Cloudbase Network Corp. Ltd.Registered in Taiwan
Network
AS152626・AS213430Assigned by APNIC, not resold
Facility
Chief Telecom, TaipeiChief HD / LY / LY2
Verify
PeeringDB net/40307Peering and facility presence are public