Documentation

Documentation

inference.club is a community-run inference network. Members run GPUs at home and expose their local model servers; everyone calls those models through one OpenAI-compatible API. One key works in both directions — to consume inference and to provide it.

Base URL   https://api.inference.club/v1
Auth       Authorization: Bearer <your-api-key>

Start here

What you can build with

The network serves far more than chat. Which modalities are live depends on what models the providers are running right now.

How it fits together

A small always-on cloud box is the control plane — the website, the API, auth, routing, billing, and async orchestration. The heavy GPU compute lives on home hardware that is never exposed to the internet. A private Tailscale tunnel bridges the two, and generated media is offloaded to object storage so the small box stays off the hot path.

Consumers

OpenAI SDKs · curl · browser · agents

HTTPS
Hetzner VPS · control plane
CaddyTLS · SSE
Nuxtfrontend
DjangoAPI · routing
Celeryasync jobs
Postgressource of truth
Rediscache · broker
Tailscale sidecarSOCKS5 · WireGuard out
Tailscale tailnet · MagicDNS · no port-forward
Home k3s cluster · GPU compute

inference-club-agent (kubernetes discovery) routes /v1/* to in-cluster services across 3× RTX 4090 + DGX Spark

vLLMLM StudioFlux.2LTX-2DiaACE-StepMagpie TTSNemotron ASRTRELLIS
Generated media (images · audio · video · 3D) is written to Google Cloud Storage and served straight from Google's edge — off the VPS hot path.

The agent reports what it serves every ~30 seconds, so the platform's model list and routing always reflect what is actually online. For the full story — request path, the home k3s cluster, the Hetzner deployment, and storage — see Architecture.