Noket lets one person share an AI model with other people, without handing over the API key behind it. The owner connects a model — one served by a local runtime through the Noket connector, or any public OpenAI- or Anthropic-compatible API — and decides who can use it, how much each person may spend, and when access ends. Everyone else calls the model with their own Noket API key, using the OpenAI or Anthropic SDK they already have.
Inference stays wherever the model already runs. Noket carries the request, enforces budgets, and records who called what.
Start where you are
Use a model someone shared with you
Accept the invitation, create an API key, and make your first request.
Call the API from your app
Point the OpenAI or Anthropic SDK at Noket and keep the code you already have.
Share your models
Add a model, set a budget, and invite the people who should use it.
Share a model on your machine
Install the connector, expose a local server through the tunnel, and add it as a model.
The three surfaces
| Surface | Where | What it is for |
|---|---|---|
| Console | noket.dev | Models, members, budgets, API keys, usage, logs, and tunnel nodes. |
| API | https://api.noket.dev |
OpenAI- and Anthropic-compatible inference endpoints. |
| Connector | noket, noketd on your machine |
Keeps the tunnel open and serves the upstreams you expose. |
You only need the connector if you are sharing a model that runs on your own hardware. Using someone else’s model, or sharing a public API, needs nothing but the console and an API key.
How the pieces fit
- A model is the callable unit. It has a public id such as
alice/llama-3, a price, access rules, and limits. People call the model, never the thing behind it. - An upstream is an API the connector can reach from your machine: a local runtime, a corporate gateway, or a hosted provider. An upstream model id is one model on that API.
- A node is a machine running the connector. Models bound to a node reach their upstream through the tunnel the connector opens, so nothing on your network accepts inbound connections.
Read Models and access for the vocabulary, or How the tunnel works for the connection itself.