Skip to main content

​
Where credentials live

A provider credential is a row under Settings > Model Providers. The row belongs to an organization, a team or a project, and every virtual key whose scope can see the row can dispatch through it. The same rows serve evaluators, the playground and the rest of LangWatch.
The Model Providers page with the Add Model Provider menu open, listing every provider

Settings > Model Providers, the Add Model Provider menu.

Each row has a Name, the credential fields listed below, optional extra headers, a Routing handle and an Advanced (Gateway) section with a fallback priority, a provider config JSON, and requests per minute, tokens per minute and requests per day fields. The gateway does not enforce those three; they are recorded on the row only. Cap traffic with the requests per minute and requests per day settings on a virtual key. See Rate limits.
The Bedrock provider drawer with the Advanced (Gateway) section expanded

Advanced (Gateway) on a Bedrock row: fallback priority and provider config.

See Model providers for the row itself and Fallback chains for the order the gateway tries rows in.

​
Supported providers

Model names are provider/model. The provider segment is the id in the first column. See Model naming. The provider ids azure_openai, aws_bedrock, vertex_ai, google_vertex and google_gemini are accepted as aliases of azure, bedrock, vertex and gemini.

​
What each endpoint family means

A request to an endpoint the provider does not serve gets the provider’s own error back. For example, POST /v1/embeddings with an Anthropic model returns Anthropic’s rejection.

​
More than one row of the same provider

Create one row per account, region or endpoint, for example an Azure OpenAI row per Azure resource. Give each row a Routing handle, for example eu, and pin a request to that row with eu/gpt-5-mini. A routing policy orders the rows a key tries. See Routing policies.
Also check: Model naming for how a model name selects a provider, and Fallback chains for what happens when a provider fails.
Last modified on September 6, 2026