Best VPS for ElizaOS (2026): Specs for the Agent OS
ElizaOS, the project that came out of the ai16z agent push and grew into its own foundation, is the most popular TypeScript agent runtime in 2026. It is also the runtime where the choice of VPS matters most, because the plugin model means every connector you load is RAM that does not get released.
I run two ElizaOS agents on a Hetzner box, one bot on a Contabo box, and I tested ten different plugin combinations to write this up.
What ElizaOS Actually Needs
Three resource curves to plan for:
- Base runtime. Node, the agent loop, the action and evaluator engines. About 200 MB resident with a stripped down config.
- Plugins. Each one adds 50 to 300 MB depending on what it does. Knowledge plugins are the biggest, social connectors are mid weight, simple tool plugins are tiny.
- Storage. Embeddings cache, message history, plugin state. Grows steadily as the agent runs.
The thing nobody mentions in the docs: leaving the agent running for weeks accumulates conversation history fast, and if you do not prune you will see the Postgres database climb past 5 GB on a chatty Discord bot.
VPS Comparison for ElizaOS
| Provider | Plan | vCPU | RAM | Disk | Monthly | Best fit |
|---|---|---|---|---|---|---|
| Hetzner Cloud | CCX13 | 2 | 8 GB | 80 GB NVMe | 14.86 EUR | Two to four agents, daily use |
| Contabo VPS | VPS S | 4 | 8 GB | 100 GB NVMe | 4.50 EUR | Budget pick for a long lived bot |
| DigitalOcean | Premium AMD 2 GB | 2 | 4 GB | 80 GB NVMe | 21 USD | Single agent, US latency |
| Hetzner Cloud | CCX23 | 4 | 16 GB | 160 GB NVMe | 29.74 EUR | Five plus agents with heavy plugins |
Hetzner Cloud CCX13: The default
This is what I run my pair on. NVMe handles the embeddings writes well, dedicated vCPU keeps the agent loop snappy when multiple plugins fire on the same message, and Falkenstein latency to all the major model APIs is sub 80 ms. The 8 GB ceiling is enough for two to four agents if you are reasonable with plugins.
Pros that actually matter:
- API plus Terraform makes redeploys after a bad plugin update painless
- Snapshots let you roll back to a known good config
- EU jurisdiction if your agents handle PII
Real downside: the 20 TB outbound cap is fine for chat agents, tight if you start pulling large knowledge corpora.
Get Hetzner: Hetzner Cloud.
Contabo VPS S: Cheapest path to a running bot
4 vCPU and 8 GB for 4.50 EUR a month. On the newer NVMe plans the I/O is fine for ElizaOS even with knowledge plugins. The real trade offs:
- Provisioning takes hours, not seconds
- Outbound to the major model APIs runs 100 to 200 ms slower than Hetzner from EU Contabo data centers
- Customer support reply times on the lowest tier can be slow
For a bot you set up once and leave running, those negatives stop mattering after the first week.
Get Contabo: Contabo VPS.
DigitalOcean Premium AMD 2 GB: For single agents in the US
4 GB RAM is tight for ElizaOS once you load more than two plugins, but a single agent with Discord and a knowledge plugin fits. NYC3 latency to OpenAI and Anthropic is the lowest of the mainstream US options. The droplet snapshot system is genuinely useful when you want to roll back a misbehaving agent.
Honest negative: 21 USD for 4 GB. Hetzner does the same job for two thirds the price unless region matters.
Set up DigitalOcean: DigitalOcean.
Hetzner Cloud CCX23: When you run a fleet
Once you are running five or more agents or your knowledge plugins are pulling in serious data, CCX23 is the next step. 16 GB RAM is enough headroom for a heavy multi agent setup, and the 4 vCPU keeps the action loops from blocking each other.
Get Hetzner: Hetzner Cloud.
Common Gotchas When Self-Hosting ElizaOS
Three things to watch:
- Plugin memory leaks. Some community plugins do not release event listeners properly. Restart the agent on a schedule until the plugin maintainer ships a fix.
- Postgres growth. If you use the Postgres plugin for memory, run vacuum monthly. The default autovacuum is too lazy for chatty agents.
- Discord rate limits. Multiple agents on one VPS sharing an outbound IP can hit Discord’s per IP rate limits faster than you expect. Two agents fine, four plus and you start seeing 429s during traffic spikes.
What I would actually pick
If you are starting today:
- EU, two to four agents: Hetzner CCX13
- EU, single bot, low budget: Contabo VPS S
- US single agent: DigitalOcean Premium AMD 2 GB or 4 GB
- Fleet of agents: Hetzner CCX23
For the full self-hosting picture, see the SelfHostVPS comparison. ElizaOS ships frequently, so I refresh this page when the plugin API or storage layer changes meaningfully.
Frequently asked questions
What is the minimum VPS spec for ElizaOS?
2 vCPU and 4 GB RAM if you run one agent with two or three plugins and point it at a remote model API. Each loaded plugin sits resident in the Node process, so once you add Discord, Telegram, Farcaster, and a couple of knowledge plugins, the heap creeps over 1.5 GB. With four or more agents, plan for 8 GB.
Does ElizaOS need a GPU on the VPS?
Only if you self-host the language model. ElizaOS is a TypeScript runtime and a plugin layer. All the agent reasoning happens behind whatever model provider you configure. A CPU only Hetzner CCX or Contabo NVMe VPS is the normal setup.
Can I run multiple ElizaOS agents on one VPS?
Yes, but plan for about 600 MB to 1 GB RAM per active agent depending on plugin count. Five agents on 8 GB is fine. Ten agents on 8 GB is asking for trouble once a memory plugin starts caching aggressively.
Why does ElizaOS use so much disk?
The knowledge plugins cache embeddings, the message history is stored locally by default, and the Node modules tree is heavy. 80 GB is the realistic minimum even for a single agent. After three months of active Discord and Telegram use I have seen the data directory pass 25 GB on a single agent.