Your own models and keys
Besides the platform catalog, a workspace can bring its own models: a Model key of your own connection stores a service address, a key and the list of models you allow. One key serves models of all three kinds — chat, embeddings and rerank.
Connecting a provider
- Open Settings → Integrations and create a Model key of your own connection.
- Fill in the Service address (for example
https://api.example.com/v1) and the API key. - Press Load from provider — the platform asks that endpoint for its model list. If the provider does not serve one, use Add manually.
- Turn on the models you want: loaded models arrive disabled.
- Pick the Model purpose on each row — Chat, Embeddings or Rerank.
- Save the connection.
Integrations
Model key of your own
Text is always accepted; the modality chips restrict which attachments are sent to the model. With Documents off, a PDF fails the turn on this model instead of travelling unread.
How a model reaches the dropdowns
A connection model gets an id of the form ws:<connection id>:<model id> and sits next to
the catalog models — but only in the dropdown that matches its purpose:
| Purpose | Where it appears | Example label |
|---|---|---|
| Chat | the Model field of the llm node | Acme · gpt-4o |
| Embeddings | the embedding node, knowledge-base creation | Acme · text-embedding-3-small |
| Rerank | the rerank node | Acme · rerank-v3.5 |
The label is “connection name · model label”, so two connections carrying the same model stay distinguishable. In Settings → Models these rows live under the workspace group, and their visibility is toggled there too.
Fallback chains
A separate connection kind is the fallback chain. It is assembled from models you already connected, listed in the order they are tried, and appears in the LLM dropdown as a single entry. Checkboxes decide what makes the chain advance — request timeouts, HTTP 4xx, HTTP 5xx.
When the provider is not reachable directly
The connection has an optional Proxy field: http://, https://, socks5:// or
socks5h://, with credentials allowed inline. Empty means a direct connection.
The platform’s own proxy is never applied here, even when the operator has configured one. The key and the address are yours, so a route through infrastructure you did not choose is not ours to pick — and the common BYO case is an endpoint on your own network, which an external proxy does not reach anyway. The same address is used both when the platform asks the provider for its model list and when a node calls it during a run.
What this does to billing
Your model runs on your key and you are billed for it directly, so the platform does not meter its tokens — otherwise the same work would be paid for twice.
| Charge | Catalog model | Your model (BYO) |
|---|---|---|
| Scenario run | 0.4 quarks | 0.4 quarks |
| Each LLM node | 0.1 quarks | 0.1 quarks |
| LLM tokens | the model's tariff | not billed |
| Embedding and rerank tokens | billed | no ledger row at all |
The tokens of a BYO call stay visible under Usage — the node really did run — but their price is zero.