When the AI model runs in the cloud, the agent's computer needs little: 4 GB of memory is the minimum Claude Code and Codex publish, and OpenClaw's gateway runs in 1 GB. A local model needs far more, as the tables below show.
The options at a glance
Any of these can run an agent whose model is in the cloud. They differ in what they cost to leave on, and in what else they can do.
| Computer | Power when idle | Running cost | Good to know |
|---|---|---|---|
| An old laptop | 6.4 W | $0.86 a month in electricity | Costs nothing extra if you already own it, but you must set it never to sleep, including with the lid closed, and it stops when it loses power or Wi-Fi. |
| An office-surplus desktop (e.g. Dell OptiPlex) | 13 W | $1.74 a month in electricity | Uses about three times the idle power of a Mac mini or N100 mini PC, though that is still under $2.50 a month at the US average electricity price. |
| A mini PC (e.g. Intel N100) | 4.8 W | $0.64 a month in electricity | Small and cheap to leave on, but with no dedicated graphics card a local model would run on the CPU, which Hermes' docs say works but responds more slowly. |
| A Mac mini (Apple silicon) | 4 W | $0.53 a month in electricity | Very low idle power, and it runs macOS, which OpenClaw's docs say you need for iMessage. |
| A Raspberry Pi 5 | 2.7 W | $0.36 a month in electricity | The lowest power draw here, but it is ARM-based, so some optional tools may lack builds, and it is too slow for local models. |
| A rented server (VPS) | – | $15–24 a month | Always on and reachable without touching your home network, but you pay every month and your agent and its files live on someone else's computer. |
| The computer I use every day | – | None extra | No extra cost, but the agent only works while that computer is on and awake, and it shares your files, accounts and screen. |
What "self-hosted" means, and what it doesn't
Self-hosted means the agent program runs on a computer you control: its memory, its files and the accounts it's signed in to live there. It doesn't usually mean the AI is on that computer. Most people connect the agent to a model in the cloud, such as Claude, ChatGPT or Gemini, and that provider still sees the prompts, screenshots and files the agent sends it. Only a local model keeps those on your own hardware, and that takes much more memory.
The computer can be your everyday one, a spare in a drawer, a small box you buy for it, or a server you rent by the month. I built ibara, free software that gives an agent a computer of its own through MCP tools, because I'd rather an agent work on its own computer than in my files, accounts and screen. I run my own agents on five computers: two old mini PCs, two office-surplus Dell OptiPlex desktops and a newer mini PC.
Which agents you can self-host
These four install on a computer you control. Each still needs an AI model: a paid plan or API account in the cloud, or a local model. Hermes vs OpenClaw compares the two most popular in detail.
| Agent | Runs on | Memory |
|---|---|---|
| OpenClaw | macOS, Linux, Windows (native or WSL2) | 1 GB minimum, 2 GB recommended |
| Hermes Agent | macOS (Apple silicon), Windows 10/11, Linux, WSL2 | No published minimum |
| Claude Code | macOS 13.0+, Windows 10 1809+ or Windows Server 2019+, Ubuntu 20.04+, Debian 10+, Alpine Linux 3.19+ | 4 GB minimum |
| Codex | macOS, Linux, Windows | 4 GB minimum, 8 GB recommended |
Running the AI model on your own computer
A local model needs far more memory than the agent does. These are the minimums the Ollama and Hermes Agent docs give for common model sizes, in billions of parameters (B):
| Model size | Memory needed | Download |
|---|---|---|
| 3–4B | 8 GB | about 2.5 GB |
| 7–8B | 8 GB | about 5.2 GB |
| 14B | 16 GB | about 9.3 GB |
| 30B+ | 32 GB | about 19 GB |
On an Apple silicon Mac, the processor and graphics share one pool of memory, so the Mac's total memory limits the model size. On a PC, a model runs fast only if it fits in the graphics card's own memory; without one, it runs on the processor, more slowly.
Questions
What is a self-hosted AI agent?
An AI agent that runs on a computer you control, such as your own PC, a Mac mini or a rented server, rather than in the agent maker's cloud. OpenClaw, Hermes Agent, Claude Code and Codex can all run this way. The AI model can still be in the cloud.
How much RAM does a self-hosted AI agent need?
With the model in the cloud, 4 GB is the minimum Claude Code and Codex publish, and Codex recommends 8 GB. OpenClaw's gateway alone runs in 1 GB. A local model needs more: about 8 GB for a 7–8B model and 32 GB for a 30B one.
Do I need a Mac mini to run OpenClaw?
No. OpenClaw runs on Linux, Windows and macOS. A Mac mini is worth it for iMessage, which needs macOS, or for running local models in its shared memory. Mac mini for OpenClaw goes through the choice.
Is a rented server (VPS) or a computer at home better?
A server is always on and doesn't touch your home network, but costs about $15–24 a month for 2 processors and 4–8 GB of memory, and your agent and its files live on someone else's computer. A spare computer at home costs under $2 a month in electricity to leave on.
Is it safe to run an AI agent on my own computer?
It's safer on a computer of its own, with its own accounts and only the logins it needs, than on the computer you use every day. Is OpenClaw safe? covers the risks and the precautions.
Sources
- OpenClaw docs: Install (system requirements)
- OpenClaw docs: Raspberry Pi (minimum and recommended hardware)
- OpenClaw docs: DigitalOcean ($6/month 1 GB Droplet)
- Hermes Agent docs: Platform support
- Hermes Agent docs: Installation
- Hermes Agent README ($5 VPS)
- Hermes Agent docs: Run Hermes locally with Ollama
- Hermes Agent docs: Run local LLMs on Mac
- Hermes Agent docs: Local models
- Claude Code docs: Advanced setup (system requirements)
- OpenAI: Codex CLI install
- openai/codex: Installing & building (system requirements)
- Ollama README v0.12.0 (RAM per model size)
- Ollama library: qwen3 (download sizes)
- Apple Newsroom: Apple unleashes M1 (unified memory)
- Apple: Mac mini tech specs
- Apple Support: Mac mini power consumption and thermal output
- Notebookcheck: Lenovo ThinkPad T480s (i5-8250U) review
- Notebookcheck: Minisforum UN100L (Intel N100) review
- ServeTheHome: Dell OptiPlex 7070 Micro review
- ServeTheHome: Dell OptiPlex 7090 Micro review
- ServeTheHome: Dell OptiPlex 7060 Micro overview
- Tom's Hardware: Raspberry Pi 5 review
- Jeff Geerling: New 2GB Pi 5 has 33% smaller die, 30% idle power savings
- EIA Electric Power Monthly Table 5.6.A (July 2026)
- DigitalOcean: Droplet pricing
- Vultr API: plans (vc2-2c-4gb)
- Hostinger: VPS hosting plans
- Electricity: US average residential price, July 2026 (EIA Electric Power Monthly, released 2026-09-24). Monthly cost = watts × 730 hours ÷ 1000 × price.