Playground

Two chat UIs: the org playground for any OpenAI-compatible endpoint, and the project playground for your fine-tuned adapters.

Arkor Cloud has two playgrounds with different jobs.

Org playground

arkor.ai/<org>/playground chats with any OpenAI-compatible endpoint: your Arkor endpoints, a local vLLM, or a third-party provider. Point it at a base URL and go.

• Requests go straight from your browser to the endpoint. Your API key is held in memory only, never stored, and never sent to the dashboard backend — it reaches only the endpoint you point the playground at (when that target is an Arkor endpoint, its *.arkor.app ingress receives the key to authenticate the request, like any other client would send it).
• The endpoint must allow browser (CORS) requests. Because the call comes from page JavaScript, the target has to permit cross-origin requests from arkor.ai. Arkor endpoints already do; for a local vLLM or a third-party provider that blocks browser calls, enable CORS on that server (for vLLM, e.g. --allowed-origins '["https://www.arkor.ai"]') or the request will fail even though the API itself is compatible.
• Endpoint picker. A dropdown lists your org's enabled Arkor endpoints, grouped by project. Selecting one fills in the base URL (and marks open endpoints as needing no key).
• Controls for model override, temperature, max tokens, system prompt, and thinking mode (reasoning effort) on supported models.
• Code example. The current configuration can be exported as a ready-to-run snippet.
• Frontier compare. An optional second pane streams the same conversation through a frontier model, so you can judge your model side by side. The frontier call is proxied server-side; your endpoint key is not involved.

Settings persist per organization in your browser, except the API key, which is deliberately forgotten.

Project playground

arkor.ai/<org>/<project>/playground chats with your fine-tuned adapters. Pick a completed training job, then:

• Final — the adapter from the end of the run.
• Checkpoint N — any mid-run checkpoint, to compare training stages.
• Base only — the job's base model with no adapter, for a baseline.

Responses stream token by token. The adapter loads dynamically on Arkor's inference workers, so switching between jobs and checkpoints takes seconds and needs no deployment. When a checkpoint chats well, serve it properly with an endpoint.