Skip to main content
Kylon exposes provider-compatible model routes under /proxy. Use the same Kylon API-key authentication as other proxy calls.

OpenAI-compatible routes

Example:

Anthropic-compatible routes

Example:

Gemini-compatible routes

Gemini helper requests

For file or audio/video understanding, first request upload URLs:
POST /proxy/gemini/files/upload-urls accepts up to 20 files, each at most 1 GiB. Upload each file’s bytes to the returned uploadUrl using its declared MIME type, then pass the returned gcsUri values to either understanding helper:
An understanding request accepts 1–10 URIs and a non-empty prompt; optional mimeTypes must have one entry per URI. The response includes model, text, and token usage. Convert Office documents to PDF before this flow; exact spreadsheet values should be read with spreadsheet tools. /proxy/gemini/nano-banana-pro requires prompt and filename. It accepts optional resolution (1K, 2K, or 4K) and base64 or data-URL image input through input_image or input_images. It returns image data and text in JSON; filename labels the result and does not write a local file.

Cerebras-compatible route

OpenRouter route

Use the OpenRouter route to access a broad catalog of third-party models with a single Kylon API key. Example:

OpenRouter video

The current supported model is minimax/hailuo-3, with whole-second durations from 5 to 15 and resolution 2K (the default). The proxy validates the model, duration, resolution, and reference types before dispatch.
Reference images, videos, and audio must first be uploaded through /proxy/tools/upload. In frame_images or input_references, supply the returned staging key in the nested url field, for example:
Public HTTP URLs are rejected in these reference fields. frame_images accepts images only, with first_frame or last_frame as frame_type. Audio references must be paired with at least one image or video reference. The proxy signs workspace staging references for the provider. Generation is charged before dispatch and is nonrefundable once dispatched. Polling and content downloads have no separate request charge. A completed poll can reconcile a higher provider-reported cost with an additional charge; it does not refund an overestimate. See Proxy Billing.

ElevenLabs-compatible routes

Provider routes use Kylon proxy authentication and the workspace’s available credits. A matching agent skill is not required to reach the route.

Streaming

OpenAI-compatible, Anthropic-compatible, Gemini-compatible, Cerebras-compatible, and OpenRouter routes support streaming where the provider route supports it. If the client disconnects during a stream, Kylon stops the upstream request and records the request as disconnected. Alternate provider routes use their own configured model catalog. A model name accepted by one route is not a promise that another route accepts it.

Usage accounting

Model routes compute credits from usage when provider usage data is available. See Proxy Billing.