Import Models from a Provider

Registering models one route at a time is fine for a handful, but tedious when a provider serves dozens. Import from provider takes any OpenAI-compatible base URL (and an optional API key), asks the provider which models it serves, and lets you import the ones you don't already have — each as a normal model route, with batch defaults you set once and per-model overrides where you need them.

The catalog is fetched server-side, so the API key you supply never reaches the browser. The flow is additive: it only ever creates new routes — nothing existing is changed — so you can re-run it any time to pick up just what's new.

When to use it

• Onboarding a provider (vLLM, an OpenAI-compatible gateway, a hosted endpoint) that exposes many models at once.
• Periodically checking a provider for new models without re-reviewing the ones you already serve.

If you only need one or two routes, creating them directly on the Models page (or via the Management API) is quicker.

Walkthrough

Open the dashboard Models page and choose Import from provider. The wizard runs in three steps.

1. Connect

Enter the provider's API base URL and an optional API key.

• The base URL must be http(s) and is the OpenAI-style base — obleth fetches GET {base}/models from it. If you get "No /models endpoint here (HTTP 404)", check whether the base should include or drop the /v1 suffix.
• The key, if the provider requires one, is sent as a bearer token on the server-side catalog request only. A 401/403 is reported back as a rejected key.

obleth accepts either the OpenAI { "object": "list", "data": [...] } shape or a bare JSON array, keeps entries with a non-empty id, and dedupes and sorts them.

2. Select

You now see every model the provider returned, each classified:

• new — no route of yours matches it. Selected by default.
• existing — already imported. Shown (toggle Show existing) but not offered for import.

A model counts as existing when one of your routes matches it by normalized model name, or by the same provider base URL and upstream id together. That dual check is what makes re-running the importer surface only genuinely new models, even if you renamed a route after importing it.

Set batch defaults once — they apply to every selected model:

• Model type — chat/completions, embeddings, audio transcription (STT), text-to-speech (TTS), or image generation.
• Context window and costs (input_cost_per_token, output_cost_per_token, and the per-image / per-audio-second / per-character costs where the type uses them).
• Admission weight, tags, enabled, and an optional description.

Then, per row, edit the model name obleth will register and apply any per-model overrides on top of the batch defaults. Filter the list by id or name to find specific models.

3. Review

obleth builds an import plan from your selections and shows it before anything is written. Confirm to create the routes. Each imported model becomes a normal route pointing at the provider's base URL and upstream id, carrying the API key you supplied (if any).

What it does and doesn't change

• Create-only. The wizard excludes already-existing models from selection, so importing never edits or overwrites a route you already have.
• No client or gateway restart. Imported routes are live as soon as they're created, like any other route.
• Re-runnable. Run it again later and only genuinely new models are offered.

Two selected models that normalize to the same name are rejected at the review step — rename one before continuing.

Related

• OpenAI-compatible API — the request surface your imported models serve.
• Multi-modal Models — registering embeddings, audio, and image models (the model types this importer supports).
• Management API — creating and updating routes programmatically, including bulk model templates.