Best VPS for Browser-Use (2026): Real Specs for AI Browser Automation
Browser-Use is the browser automation agent that turned a research project into 92K plus GitHub stars by mid 2026. The pitch is simple: an agent with vision and DOM access that can drive any website. The hosting picture is heavier than the demo videos suggest, because every agent run spins up a real Chromium instance.
I ran Browser-Use across three VPS tiers for two weeks each, with single tab and parallel automation modes, to figure out where the breakpoints actually are.
What Browser-Use Actually Needs
Three resource curves:
- The agent. Python plus the orchestration loop. 200 to 400 MB resident.
- Chromium. This is the big one. Headless Chrome in automation mode runs 800 MB to 2 GB per session depending on the pages it visits.
- Concurrent sessions. Each parallel task is another Chromium instance. RAM grows linearly.
The thing nobody mentions: heavy pages with many tabs and the vision mode taking screenshots will push a single session past 3 GB. Plan accordingly.
VPS Comparison for Browser-Use
| Provider | Plan | vCPU | RAM | Disk | Monthly | Best fit |
|---|---|---|---|---|---|---|
| Hetzner Cloud | CCX13 | 2 | 8 GB | 80 GB NVMe | 14.86 EUR | Single session, daily use |
| Contabo VPS | VPS S | 4 | 8 GB | 100 GB NVMe | 4.50 EUR | Budget, single session |
| Hetzner Cloud | CCX23 | 4 | 16 GB | 160 GB NVMe | 29.74 EUR | Two to four parallel sessions |
| Vultr | HF General 32 GB | 8 | 32 GB | 640 GB NVMe | 96 USD | Heavy parallel automation |
Hetzner Cloud CCX13: The default
This is what I run for single agent automation. NVMe handles the Chromium cache writes well, dedicated vCPU keeps the agent loop responsive, Falkenstein latency to most sites is good. 8 GB RAM is enough for one Chromium plus the agent comfortably.
Pros for this workload:
- Sub 80 ms latency to model providers, which matters for vision queries
- 20 TB outbound is generous for normal automation volume
- Snapshots make rollback after a config change painless
Real downside: jumps to CCX23 once you want parallel sessions. The 8 GB ceiling is real.
Get Hetzner: Hetzner Cloud.
Contabo VPS S: Budget single session
4 vCPU and 8 GB at 4.50 EUR. For a single Browser-Use session running on a schedule, this works fine. The trade offs:
- Slow provisioning
- Outbound latency higher than Hetzner, which makes vision mode feel slower
- Bandwidth caps on cheaper tiers can be tight under heavy automation
For low volume scheduled jobs, the price wins.
Get Contabo: Contabo VPS.
Hetzner Cloud CCX23: For parallel work
If you want two to four parallel Browser-Use sessions, CCX23 with 16 GB RAM and 4 dedicated vCPU is the right tier. Each session gets enough headroom and the dedicated cores keep the agent loops responsive when multiple are active.
Get Hetzner: Hetzner Cloud.
Vultr HF General 32 GB: Heavy automation
For real production automation with eight or more parallel sessions, you need 32 GB. Vultr’s High Frequency 32 GB plan at 96 USD a month is the cleanest cloud option in this tier. Multiple regions help if your target sites are geographically diverse.
Get Vultr: Vultr.
Things Worth Knowing
Three real issues:
- Anti bot detection. Cloudflare and DataDome will challenge or block automation from cloud IPs. Plan for residential proxies if your target sites are protected.
- Cookie persistence. Browser-Use is stateless by default. Save and reload sessions explicitly if you need to stay logged in.
- Page load timeouts. Default Playwright timeouts are too short for slow sites. Bump them or you will see failures that look like agent bugs.
What I would actually pick
If you are starting today:
- EU, daily single session: Hetzner CCX13
- EU, budget scheduled jobs: Contabo VPS S
- EU, parallel automation: Hetzner CCX23
- Heavy production automation: Vultr HF General 32 GB
For the broader self hosting picture, see the SelfHostVPS comparison. Browser-Use iterates frequently and I refresh this page when the agent loop or browser layer changes meaningfully.
Frequently asked questions
What is the minimum VPS spec for Browser-Use?
4 vCPU and 8 GB RAM is the floor. Chromium itself is hungry under automation, especially with vision mode enabled. The agent loop plus a single Chromium process sits around 3 to 4 GB resident. Parallel sessions multiply that fast.
Does Browser-Use need a GPU on the VPS?
Not for the agent. Vision mode uses the multimodal capability of the model, not local GPU compute. A CPU only VPS is the standard setup. GPU only enters the picture if you also self-host a vision capable model.
Can Browser-Use run on a 5 dollar Contabo box?
Only for very simple single tab tasks with a small viewport. Real browser automation on the cheapest tier is painful because Chromium swaps and the agent loop times out waiting for page loads. Spend the extra 4 EUR.
Why does Browser-Use need so much bandwidth?
Every page the agent visits is downloaded in full, including images, fonts, and scripts. Heavy sites can be 5 to 15 MB per page. A busy automation can pull 1 to 2 GB an hour. Hetzner's 20 TB outbound is generous, Contabo's bandwidth on cheaper tiers can be tight.