Best VPS for OpenHands (2026): Real Specs for the Coding Agent

OpenHands is the project formerly known as OpenDevin, now at 74K plus GitHub stars and shipping monthly. The hosting picture changed a lot in early 2026 when the team pulled the runtime out of the main process and into a Docker sandbox. Good for safety. Painful if you read the old docs and provision a 2 GB VPS.

I have been running OpenHands as my background coding agent for nine weeks. Here is what the spec sheet does not tell you.

What OpenHands Actually Needs

Three resource sinks, in order of how often they bite people:

1. The sandbox. Docker-in-Docker, with a runtime image that ranges from 800 MB to 2.5 GB depending on the language stack. Every active task spawns one.
2. Build and install steps. When the agent decides to run npm install or pip install on a real repo, RAM usage spikes hard.
3. The model. If you also self-host the LLM on the same box, you need a GPU and a different tier of hardware entirely.

Most users want path 1 plus 2 on a CPU box, with the model living on Anthropic or OpenAI. That is the setup the rest of this page covers.

VPS Comparison for OpenHands

Hetzner Cloud CCX23: My current pick

The CCX line is the sweet spot for OpenHands. CCX13 with 8 GB works for solo light use, but the moment you ask the agent to do two things in parallel you want CCX23. NVMe matters because the sandbox writes a lot during builds, and Hetzner’s I/O is consistent in a way Contabo’s older nodes were not.

What works for OpenHands here:

• Cloud snapshots make rollback after a bad task painless
• Falkenstein and Helsinki regions both run under 80 ms to the Anthropic API
• Hourly billing means you can spin up a beefier box for a specific job

Downside: outbound bandwidth is 20 TB included, which sounds like a lot until the agent decides to clone five gigabyte repos in a loop. Watch the meter.

Provision Hetzner: Hetzner Cloud.

Contabo VPS M: The 8 EUR sleeper

Hard to argue with 6 vCPU and 16 GB for 8.50 EUR. On the newer NVMe plans, the I/O profile is fine for OpenHands. The two real downsides: provisioning takes hours rather than seconds, and outbound latency to the major model APIs is consistently 100 to 200 ms worse than Hetzner. If you keep the box running long term it is fine. If you spin up and tear down a lot, the wait time is brutal.

Pros:

• Cheapest box that actually handles the sandbox without swapping
• 400 GB disk is generous for keeping multiple repo clones around

Negative I will not gloss over: their support reply times on the cheaper tiers can stretch into days.

Get Contabo: Contabo VPS.

DigitalOcean Premium AMD: For US teams

If your model endpoint is in US-east and your team is too, DigitalOcean’s NYC3 nodes cut a round trip out of every model call. The 8 GB tier handles a single OpenHands worker, but I would not run more than one heavy sandbox on it. The newer Premium AMD nodes are noticeably snappier than the older shared CPU plans, worth the price bump.

Honest downside: 28 USD for 8 GB is steep against Hetzner’s EU pricing. Pay it if latency matters more than EUR per gigabyte.

Set up DigitalOcean: DigitalOcean.

Vultr HF General 16 GB: For multi-sandbox workloads

If you are using OpenHands as a team tool and need three or four concurrent sandboxes, Vultr’s High Frequency line is the easiest place to land. The 16 GB plan with 4 vCPU is enough headroom that no single task starves the others, and Vultr’s worldwide region map matters when your team is split across continents.

Get Vultr: Vultr.

Common Gotchas When Self-Hosting OpenHands

Three things that tripped me up:

1. Docker socket exposure. The sandbox needs the host docker socket. That is fine on a single-user VPS, dangerous on a shared one. Do not run OpenHands as a service for untrusted users without a separate runtime user.
2. Cleanup. Old sandbox containers accumulate fast. Set up a daily docker container prune or you will run out of disk inside a week.
3. Memory limits. Cap each sandbox at 4 GB explicitly. Without limits, a runaway npm install will eat the whole VPS and crash the orchestration layer with it.

What I would actually pick

Starting today and you want OpenHands as your background coder:

• EU, daily use: Hetzner CCX23
• EU, lowest spend: Contabo VPS M, accept the slower provisioning
• US team: DigitalOcean Premium AMD 8 GB or larger
• Team setup, three plus sandboxes: Vultr HF General 16 GB

For the broader self-hosting picture, see the SelfHostVPS comparison. OpenHands moves fast, so I refresh this page after major sandbox or runtime changes hit main.