Your own inference provider.
K-AI turns the GPUs you own, and the ones you rent by the hour, into one OpenAI-compatible endpoint with a queue, access control, usage tracking and a built-in knowledge base.
What it does
One endpoint, many places
Chat front ends and software talk to a single API. K-AI decides whether a request runs on a GPU at home, a Docker host, or a rented pod, and queues it fairly when everything is busy.
Pay for GPUs only when used
Rented RunPod pods start on demand, load models from a network volume, and stop themselves after an idle period. A spend cap keeps the bill predictable.
Knowledge that follows the model
Attach one or more knowledge bases to a model profile. Passages are injected into the prompt or exposed as a search tool, and only if the caller's key may read them.
Your data stays yours
Models, documents and the knowledge graph live on infrastructure you control. Sign-in is your own SSO; software uses personal keys with narrow grants.
Profiles instead of command lines
A profile captures a model, its context length, parsers and where it may run. Switching between profiles that share the same load needs no restart.
Operable by one person
A control panel for profiles, nodes, RunPod, keys and usage, with guards against the expensive mistakes: duplicate pods, runaway spend, loading a model that will not fit.
The shape of it
Clients sign in or present a key, the router resolves the model to a profile, optionally grounds it in knowledge, and sends it to the node that already serves it. If none does, it loads the model, or tells you exactly why it cannot.
Built for
Home labs
Turn a Windows gaming PC and a few Linux boxes into a shared model service with real access control.
Private research libraries
Turn shelves of PDFs and ebooks into searchable, graph-linked knowledge for any model you run.
Small teams
Give each person and each tool its own key, see who used what, and cap what it can cost.