OpenClaw and Hermes with a self-hosted LLM: local, on a VPS, or via an API?
Short answer: the agent and the model are two different workloads and they want two different machines. A cheap VPS runs the agent beautifully and cannot run a useful model. A Mac or a GPU box at home runs the model beautifully and is a poor place for an always-on agent. So most people who want "no API bill" should end up with the hybrid: agent on a 6 to 7 euro VPS, model at home, a tunnel between them, and a cloud API as fallback for the rare hard task. If you only want one box, and the box is a VPS, use an API. If the box is at home, put both there and accept the uptime trade.
Two workloads, not one
The confusion in every "openclaw local llm" thread comes from treating the agent and the model as one thing. They are not.
The agent (OpenClaw or Hermes) is a light, always-on process that holds your channels, your schedule, your memory and your tools. It wants uptime, a public IP and a few gigabytes of RAM. A VPS is the natural home.
The model is a heavy, bursty workload that wants a GPU or an Apple Silicon chip with a lot of unified memory, and does nothing between requests. A machine at home is the natural home, because you already own it and the electricity is cheaper than renting a GPU.
Once you see it that way the hosting question splits in two and both halves get easy.
The four setups, honestly
| Setup | Agent | Model | Good for | The catch |
|---|---|---|---|---|
| Everything at home | Mac or mini-PC | Same machine, Ollama or LM Studio | Privacy absolutists, local-model tinkerers, Hermes users | Uptime and reachability. Your agent is down when your house is. |
| Everything on one VPS | VPS | Same VPS, CPU inference | Almost nobody | Too slow and too dumb for tool use unless you rent a GPU VPS at many times the price. |
| VPS agent, cloud API model | VPS | OpenAI, Anthropic, OpenRouter and so on | Most people starting out; anyone who values simplicity | A token bill, and your prompts leave your control. |
| Hybrid: VPS agent, home model | VPS | Home machine over a tunnel, API as fallback | Anyone who wants both uptime and a small or zero API bill | More moving parts. Worth it once you have run the others. |
Why a 7 euro VPS cannot run the model
People try this every week and the result is always the same. A 2 vCPU, 4 to 8 GB VPS can load a 3B parameter model quantised to 4 bit and produce a few tokens per second. That is fine for "tell me a joke" and useless for an agent that has to read a tool description, decide, call the tool, read the result and decide again. Agent work multiplies the token count and punishes a weak model, because one wrong decision early derails the whole task.
The sizes that hold up for tool use in our experience start around 7B to 8B and get comfortable at 14B and up. Those want a GPU with enough VRAM, or Apple Silicon with 16 GB or more of unified memory. GPU VPSes exist and they work, but the cheapest you will find is in a different price class from the 6 to 7 euro boxes this site is about. If you are going to pay that, compare it with the API bill first; for most personal use, the API is cheaper.
The VPS is for the agent. Keep it that way.
Ollama, llama.cpp or LM Studio
All three expose an OpenAI-compatible API, which is what OpenClaw and Hermes expect. Pick by where the model runs, not by features:
- Ollama for a headless machine nobody looks at. One command to run a model, easy to keep as a service, sensible defaults. Our default on the home server.
- llama.cpp server when you want control: specific quantisations, specific GPU offload, the newest model the day it lands. More knobs, more to read.
- LM Studio for a desktop where you are still experimenting and want to see what is happening, swap models, watch memory. Good for finding the model you like, then move it to Ollama.
"openclaw lm studio vs ollama" is a real search and the answer is: LM Studio to choose, Ollama to run.
Hermes is built for this; OpenClaw tolerates it
Hermes Agent was designed around bring-your-own-model. Pointing it at a local endpoint is a config line, it behaves well with smaller models because its tasks are narrower, and its self-written skills reduce how much the model has to reason from scratch each time. If "no API bill" is the whole point, Hermes is the agent.
OpenClaw works with a local model, but it was built in a world of large cloud models and it shows: the channel gateway, the skills marketplace and the browser automation all lean on a model that follows instructions well. With a 7B model it is usable for chat and routines; browser automation gets flaky. Our setup gives OpenClaw the home model for everyday things and a cloud model as fallback for anything that involves a browser.
The hybrid, step by step
- Agent on a VPS. Hetzner for developers, Contabo for 8 GB on a budget. The chooser has the rest.
- Model at home. Ollama on a Mac or a mini-PC with a GPU. Pick the largest model your memory takes with room to spare.
- A tunnel between them. Tailscale is the five-minute answer: both machines on the same private network, no ports opened. Cloudflare Tunnel works too.
- Point the agent at the home endpoint over the tunnel address, with a cloud API configured as fallback.
- Let the home machine sleep. Wake-on-LAN or a scheduled window for the model; the agent on the VPS stays up and queues or falls back when the model is asleep.
It is more parts than either pure setup. It is also the only setup where the agent is always reachable and the everyday token bill is close to zero.
What it costs, over a year
| Setup | Fixed | Variable | A year, typical personal use |
|---|---|---|---|
| VPS agent + cloud API | about 80 to 170 euro VPS | API tokens; light daily use is typically single-digit euros a month, browser-heavy use more | roughly 150 to 400 euro |
| Everything at home (owned machine) | about 15 to 30 euro electricity | none | under 50 euro |
| Hybrid | VPS plus about 15 to 30 euro electricity | API only as fallback | roughly 100 to 220 euro |
| GPU VPS | hundreds per month | none | thousands; only if the box does other work too |
Ranges, not quotes. VPS figures from the comparison; electricity at a European average of 0.18 euro per kWh for a machine idling most of the day; API figures from our own invoices for personal use, August 2026. Yours will differ; the shape of the comparison will not.
Not sure you want a local model at all? VPS or a Mac mini at home covers the same choice from the hardware side, and What they do shows which jobs need a strong model and which do not.
Sources
OpenClaw and Hermes Agent documentation on model endpoints; Ollama, llama.cpp and LM Studio documentation; our own setups and invoices, January to August 2026. Model size guidance reflects our experience with tool use, not a benchmark.