Antrum

Your GPUs. Your models. One door in.

An OpenAI-compatible endpoint for the hardware you own, managed from one dashboard and reachable from any app. The traffic is encrypted to your box, and we hold no key to read it.

Join the waitlist See how it works

A recreation of the dashboard's Overview for a demo fleet: four nodes, their VRAM and a week of requests.

From a box to an endpoint

  1. First, enroll a machine with one command.

    The dashboard writes this line for you with a single-use code. The installer checks the GPU, driver, VRAM and disk, opens no inbound ports, and the node shows up in your fleet.

    curl -fsSL https://api.antrumhub.com/install.sh \
  | sudo sh -s -- --code <one-time code> \
  --ca-sha256 <pin>
ok clockok hubok edgeok amdgpu driverok amd gpuok gpu accessok diskenrolled as 7f3a9c2einstalled under /
    A recreation of an install on an AMD box: the line the dashboard writes, then the installer's checks.

  2. Then load a model from the catalog, or bring your own.

    Pick one that fits, name any Ollama model, or give vLLM a Hugging Face id. The fit is checked against the card's free VRAM before anything downloads.

    A recreation of the dashboard: an engine and a model installing on a node with 9.0 GB free, and a model flagged as too large before download.

  3. Point any OpenAI-compatible app at your endpoint.

    Every node gets its own URL and its own certificate. A key from the dashboard is the bearer; revoke it and it stops working.

    Endpoint

    https://<node-id>.nodes.antrumhub.com/v1

    BASE=https://<node-id>.nodes.antrumhub.com/v1
curl $BASE/chat/completions \
  -H "Authorization: Bearer <your key>" \
  -d '{"model":"llama3.2:3b",
       "messages":[{"role":"user","content":"hi"}]}'

    Hi there. How can I help you today?

    115.8 tokens/s through the relay, measured on an RTX 2080 Ti

    A recreation of the dashboard: each node's page shows its URL and a request you can paste, and the reply streams back.

What we can see

Your prompts never touch our servers. TLS ends on your hardware: each node holds its own certificate and private key, issued to its own hostname, generated on the node. Our relay forwards encrypted bytes by reading only the connection's destination name — it has no key, so it cannot decrypt, log, or train on anything you send. Management traffic (health, model lists, key hashes) is the only thing we ever see.

The agent runs on your box and its traffic is yours to inspect. Your browser's padlock on a node's endpoint shows that node's own certificate.

How a request reaches your node Your app sends encrypted traffic to the Antrum relay, which passes it through unread to your node. The lock is on your node: it holds the certificate and key. Your node sends only management data (health, model lists, key hashes) to the Antrum hub. Your app or browser encrypted, passed through Antrum relay no key, reads the name only still encrypted Your node TLS ends here management only: health, model lists, key hashes Antrum hub
The thick path carries your requests; the lock sits on your node, not on our relay.

What ships today

On each box

Self-updating agent
probation and rollback
A new agent version runs on probation and goes back to the previous binary on its own if it is not healthy within 90 s.
Two engines, two vendors
AMD and NVIDIA
Ollama and vLLM on AMD cards under Linux, Ollama on NVIDIA cards under WSL2. The agent picks the build for your card.
MCP endpoint
one per node
Every node serves MCP at its own URL, behind the same keys as its OpenAI-compatible API.

In the dashboard

Test bench
live tokens/s
Send a prompt to any node from the browser and watch tokens per second while it streams.
Keys
scopes, quotas, usage
Limit a key to chosen nodes and models, cap its requests and tokens per day, and see what it used.
Telemetry
average and peak
GPU utilization and VRAM charted with average and peak, plus temperature and power for every card.
A recreation of the dashboard: a key's scope, its nodes, its models and what it used today.
A recreation of the dashboard: GPU utilization on a node's page, average and peak.

What the box replaces

$0.14
per million tokens, in and out, for a hosted model of 9B parameters or less: the cheapest listed row.
$0.20
per million prompt tokens, and $0.60 per million completion tokens, for anything larger: the cheapest listed model above 10B.
115.8
tokens per second through the relay, measured on a consumer RTX 2080 Ti.
10.0M
tokens a day if that card generated without a pause. At $0.14 per million, about $42 a month of hosted serving.

Your own month

$8.40 a month at hosted rates

Model size

An estimate that errs low. Your power bill and hardware are not in it.

Rates: hosted open-model serving, the lowest listed row for each size, from a public price list checked 2026-11-06.

Works with what you use

Join the waitlist

Antrum is in private beta. Tell us what you'd run and we'll invite people in batches. A few questions, about a minute.

We'll send a confirmation email first; you're on the list once you click it. We store your email and these answers with our email provider, never with your fleet data.