AI Agent Server Requirements

What a self-hosted agent actually needs — and why a GPU is usually not part of it.

A self-hosted AI agent needs far less hardware than most people expect: 2 vCPU, 4 GB of RAM and about 20 GB of disk is enough for everyday use, and no GPU is required. The agent orchestrates — it plans, calls tools and manages memory — while the heavy model inference runs on whichever AI provider you bring a key for. A GPU only enters the picture if you also want to run the language model itself on your own hardware, which is a separate machine and a separate decision.

Do you need a GPU to self-host an AI agent?

Almost certainly not, and this is the single biggest misconception about self-hosting agents. People conflate two very different jobs: running the model, and running the agent.

  • Running the model — turning tokens into tokens. This is the part that wants a GPU, lots of VRAM, and a large machine.
  • Running the agent — deciding what to do next, calling tools and APIs, storing memory, retrying on failure, and driving the loop. This is ordinary application work: some CPU, modest RAM, a little disk.

A bring-your-own-key agent like Hermes Agent only does the second job. When it needs the model to think, it makes an HTTPS request to OpenRouter, OpenAI, or Anthropic using your own account, and waits for the answer. Your server spends most of its life idle, waiting on the network. That is why a small instance is enough, and why a GPU would sit unused.

The exception is if you want to run the language model on your own hardware too, pointing the agent at a local OpenAI-compatible endpoint. That is a legitimate setup, but it is a different machine with different requirements — VRAM becomes the binding constraint, and the numbers below no longer apply to it.

Minimum and recommended requirements

ResourceMinimumRecommendedWhy
CPU2 vCPU4 vCPUThe agent loop, tool calls and TLS are CPU-light; more cores help when several tasks run at once.
RAM4 GB8 GBHolds the runtime, the agent’s working memory and any tool processes it spawns. Long-running tasks with large contexts are the main consumer.
Disk20 GB80 GBAgent software, logs, and accumulated memory and artefacts. Agents that fetch or generate files grow faster than you expect.
GPUNoneNoneInference happens at your AI provider. A GPU is only needed if you self-host the model as well.
NetworkAlways-onAlways-onAn agent that monitors, schedules or responds to webhooks must stay reachable — a laptop or a serverless function will not do.
Uptime24/724/7Scheduled and event-driven tasks fire whether or not you are at your desk.

These are the requirements for the agent runtime itself. Your AI provider bills separately for model usage, and that cost scales with how much the agent thinks, not with the size of the server it runs on.

How to size for your workload

Sizing an agent is not about how many tasks you have — it is about how many run at the same time, and how much each one holds in memory while it works.

  • Personal assistant, a handful of scheduled tasks, occasional ad-hoc work: 2 vCPU / 4 GB is comfortable. The instance is idle most of the time.
  • Several concurrent agents, long multi-step research runs, or tasks that pull down and process files: 4 vCPU / 8 GB. Concurrency and working set are what push you here, not task count.
  • Agents that shell out to heavy tooling — builds, browser automation, media processing: size for the tool, not the agent. The tool will dominate both CPU and RAM.

Start small. Because the agent is mostly waiting on the model provider, an under-loaded instance is the normal state, and scaling up later is a plan change rather than a migration.

What that costs on Flux

PlanResourcesPrice
Hermes Agent2 vCPU · 4 GB RAM · 20 GB disk$4.02/month
Hermes Agent Pro4 vCPU · 8 GB RAM · 80 GB disk$7.49/month

Both plans are pay-as-you-go with the first month free, and both include the password-protected dashboard, built-in Tailscale for private-network access, TLS and DDoS protection. Your AI provider usage is billed by them, not by us, because Hermes is bring-your-own-key. See the step-by-step guide to self-hosting an AI agent for the deployment walkthrough.

Do these numbers change on a DIY VPS?

The agent needs the same CPU, RAM and disk wherever it runs — but on a VPS you also carry the operating system, a container runtime, a reverse proxy and whatever monitoring you add, so budget headroom on top of the figures above rather than matching them exactly.

The bigger difference is not capacity, it is everything around it: certificate renewal, firewall rules, patching, backups, and being the person who notices when the box stops responding at 3am. Those are the real requirements of self-hosting, and they do not show up in a spec table.

Frequently asked questions

No, not for the agent itself. With a bring-your-own-key agent like Hermes, model inference runs on your AI provider (OpenRouter, OpenAI, Anthropic) and the agent only orchestrates — planning, calling tools and managing memory. That is CPU-light work. You would only need a GPU if you also chose to run the language model on your own hardware, which is a separate machine.
4 GB is enough for everyday use — a personal assistant, scheduled tasks and occasional ad-hoc work. Step up to 8 GB if you run several agents concurrently, do long multi-step research runs, or have tasks that download and process files. RAM is driven by concurrency and working set, not by how many tasks you have defined.
Very little on average. The agent spends most of its time waiting on network responses from the model provider, so 2 vCPU covers normal use. More cores mainly help when several tasks run at the same time, or when the agent shells out to heavy tooling like builds or browser automation.
The resource requirements are low enough that the hardware is not the obstacle — the obstacle is uptime. An agent that monitors systems, runs on a schedule, or responds to webhooks has to be reachable around the clock, which rules out a laptop that sleeps and makes a home device dependent on your own power and connection.
Around 20 GB to start. That covers the agent software, logs, and accumulated memory. Agents that fetch documents, generate files or keep long histories grow faster than expected, so 80 GB is the safer choice if yours produces artefacts rather than just answers.
No. Quality and speed come from the model you point it at, and that runs on your provider. A larger instance helps the agent handle more work at once and hold bigger working sets — it does not improve the reasoning.

Keep reading