Lay an Ollama Egg on your own host. A persistent model volume, an OpenAI-compatible endpoint and a stable private hostname, for a flat monthly fee instead of a meter that runs every time your product is used.
Runs on your hardware · never published to the internet · unmetered egress
Every hosted model API bills by the token, so the better your product does, the worse the bill gets. An Egg on your own host is a flat monthly fee, whatever you generate. Summarise a million documents on a Sunday and the invoice does not move.
The model runs on hardware you own, in a container with no public front door. Nothing is sampled for training, nothing is retained by a third party, and there is no data-processing agreement to negotiate because there is no processor.
Ollama serves its own API and an OpenAI-compatible one on the same port. Point an existing client at your Egg by changing one base URL, keep your code, and stop paying for the calls it makes.
Generated tokens are egress, and egress is unmetered on every Ruust tier. Stream long completions to as many users as you like: it is the same bill. Fair-use rate limits exist, a bandwidth meter does not.
Run the one-line enrol command on the box you want the model to live on. It can be a server in a rack, a workstation under a desk, or a machine at home.
Pick the Private LLM type and a model-storage size. There is no repo and nothing to build: it runs our Ollama image and comes up ready.
Open the Egg shell and pull a model. It lands on a persistent volume, so it survives restarts, redeploys and resizes. You choose the models and accept their licences.
Your other Eggs reach it by private hostname on the Coop network. Reference it from an app Egg and its base URL is filled in for you.
Keep the client you have
The native Ollama API lives under /api and an OpenAI-compatible API under /v1, both on port 11434. An OpenAI SDK, a LangChain client or a plain fetch all work by pointing the base URL at your Egg.
Nobody drains your box
An open inference endpoint is a magnet for abuse: a single exposed instance is free compute for whoever finds it. An Ollama Egg is never published to the internet. It is reachable from your peered Eggs and your own shell, and nowhere else.
// In your app Egg, OPENAI_BASE_URL references the LLM Egg,
// so Ruust fills in its private address for you.
const res = await fetch(`${process.env.OPENAI_BASE_URL}/chat/completions`, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
model: 'llama3.2',
messages: [{ role: 'user', content: 'Summarise this ticket.' }],
}),
});Not yet. Today an Ollama Egg runs on the CPU, even on a machine with a GPU in it. That makes small quantised models perfectly usable and large ones slow. GPU passthrough is the next piece of work on this Egg type, and we would rather tell you that here than have you find out after you have laid one.
Roughly 8 GB for a 7B model and 10 GB for a 13B, plus room for the operating system. Model files are gigabytes each, so give the Egg enough disk for the ones you plan to keep. Both can be resized later, and storage only ever grows.
Anything Ollama can pull. We ship the runtime, never the models: you pull what you want and accept its licence, which for some open-weight families has real conditions attached. Your models, your call.
Because our managed fleet has no GPUs, and pretending otherwise would sell you a slow model at a premium. You bring the machine, which is also the honest answer to where your prompts live, and we handle the runtime, the volume, the private hostname and the wiring.
An Ollama Egg on your own host, with a model volume that survives a roll, an OpenAI-compatible endpoint and no public front door. Free on your own metal whilst we are in beta: no per-Egg fee, no host cap, no card.
Ollama is an independent open-source project under the MIT licence; Ruust hosts it and is not affiliated with or endorsed by it, nor by any model vendor. Models are pulled by you and remain subject to their own licences. Inference is CPU-only today.