Introduction

KosCompute is an EU-hosted inference API with OpenAI-compatible request and response shapes. Existing OpenAI SDKs and many OpenAI-compatible tools can connect by changing the base URL and API key.

The canonical base URL is:

https://api.koscompute.com/v1

API families¶

• Chat Completions for conversational generation, tools, structured output, reasoning, streaming, and vision.
• Responses for stateless OpenAI Responses-compatible text and local function workflows.
• Anthropic Messages compatibility for clients that speak the Messages protocol.
• Audio transcription for multipart speech-to-text requests.
• Speech generation for text-to-speech with a raw audio response.

Endpoint families without a currently published model are not advertised. The public contract therefore describes what can be used now, not every handler that happens to exist.

Models are runtime data¶

Call GET /v1/models to discover canonical IDs, readiness, supported endpoints, modalities, context and output limits, features, pricing, and current sanitized capacity. The model catalog fetches this data once when opened and falls back to a timestamped build snapshot if the live request fails.

Compatibility, not identity¶

OpenAI-compatible means common SDKs and payloads work with a changed base URL. It does not mean every hosted OpenAI feature exists. KosCompute is stateless and does not provide hosted tools, stored Responses, conversations, or background jobs. Anthropic compatibility likewise describes the wire protocol; requests are served by KosCompute-hosted models and are not sent to Anthropic.

Next step¶

Complete the quickstart, then select optional features from the model catalog before enabling them.