Skip to main content
Run a standalone gateway when you need an inference endpoint without the full platform. You manage its models, keys, and provider credentials locally. For the webapp, workspaces, and managed usage history, use the full platform.

​
Install

You need Python 3.13+ and uv. Clone the repository and sync the CLI package:
Run CLI commands from this checkout with uv run airmux.

​
Start the gateway

Set a provider key, then initialize and start the gateway. The initializer creates a local configuration and inference key:
The gateway runs in the foreground at http://127.0.0.1:8080. Keep the provider key available to the process. To use another provider, export its key before starting the gateway. The initializer prints ready-to-run commands for your setup.

​
Send a request

In another terminal, copy the inference key created by initialization and send a request using a model ID from the local taxonomy:
Change the model value to another ID in your taxonomy. See the inference reference for other request formats.

​
Run the gateway in a container

Clone the source and build the image, then run the gateway with the local configuration mounted read-only:

​
Customize local setup

gateway init creates the starter configuration and taxonomy in .airmux. Use --directory to choose another location or --taxonomy to use your own model catalog. See the configuration reference for advanced settings. Standalone mode has no management API, webapp, or managed usage history. Its keys share one local workspace, and provider credentials come from environment variables. Usage events are discarded by default; see the event outbox settings to write them to a local file.