Bottom line: if you want to run an AI agent loop — something like OpenClaw, Hermes, or an n8n workflow that calls an LLM — the agent runtime is light, but you should not plan to run a local model on a budget VPS. A 2 GB box runs the agent software fine; running the model itself needs a GPU box or a cloud API. Know which one you're actually building.
The two different questions:
1) Run the agent (the runtime): cheap — 1–2 GB RAM handles it. The agent calls out to an LLM API.
2) Run the model (the brain): expensive — needs a GPU. Not what a $5 VPS is for.
What an agent runtime actually needs
OpenClaw, Hermes, and similar agent frameworks are mostly an event loop, tool calls, and some memory. They're not GPU-bound. What they need:
- RAM: 1 GB runs a modest agent; 2 GB is comfortable with room for Docker and a small DB.
- CPU: one vCPU is enough for light agent work; more helps when workflows chain many steps.
- Disk: NVMe matters for the agent's memory/state writes and logs.
- Network: solid outbound to the LLM API. Keep the box near where the API is.
Realistic pick for an agent box
| Use case | RAM | vCPU | Approx /mo |
|---|---|---|---|
| Light agent, one at a time | 1–2 GB | 1 | $5–12 |
| Several agents / heavy workflows | 4 GB | 2 | ~$15–25 |
| Local model (LLaMA etc.) | 16 GB+ / GPU | 4+ | $50+ |
Specs are provider-reported; we don't publish fabricated benchmark numbers. See our Hermes VPS requirements & methodology.
New to the whole idea? Read when a VPS beats shared hosting, and see our full Hermes VPS requirements page for exactly what the Hermes agent stack needs.