Provider recipes

One complete providers[] entry per vendor, ready to copy into a deployment document and edit. A recipe is a starting point. The saved document carries the whole entry, nothing in it refers back to the recipe, and the gateway reads no recipe at runtime, so a vendor moving an endpoint is an edit to your document like any other.

Every entry spells its credential as a ${secret:NAME} reference, resolved from a file, the environment or the secret store, in that order. Replace the models list with the names your callers ask for.

The console offers the same list under New provider, and GET /admin/v1/recipes/providers serves it as JSON with the YAML beside each decoded entry. All three read one embedded file, and this page is generated from it.

A vendor that serves both dialects has two recipes, one per surface. Declare both and a request goes to the one that speaks its dialect, passed through byte for byte; declare one and a request in the other dialect is translated to it. Fidelity covers what that costs.

recipe	vendor	dialect	endpoint
`anthropic`	Anthropic	`anthropic`	`https://api.anthropic.com`
`anthropic-openai`	Anthropic, OpenAI-compatible surface	`openai`	`https://api.anthropic.com`
`azure-openai`	Azure OpenAI	`openai`	`https://YOUR-RESOURCE-NAME.openai.azure.com/openai`
`bedrock`	Amazon Bedrock Mantle, Anthropic models	`anthropic`	`https://bedrock-mantle.us-east-1.api.aws/anthropic`
`bedrock-openai`	Amazon Bedrock, OpenAI-compatible models	`openai`	`https://bedrock-mantle.us-east-1.api.aws`
`bedrock-runtime`	Amazon Bedrock runtime, Anthropic models	`anthropic`	`https://bedrock-runtime.us-east-1.amazonaws.com/anthropic`
`cerebras`	Cerebras	`openai`	`https://api.cerebras.ai`
`deepseek`	DeepSeek	`openai`	`https://api.deepseek.com`
`fireworks`	Fireworks	`openai`	`https://api.fireworks.ai/inference`
`gemini`	Google Gemini	`openai`	`https://generativelanguage.googleapis.com/v1beta/openai`
`groq`	Groq	`openai`	`https://api.groq.com/openai`
`mistral`	Mistral	`openai`	`https://api.mistral.ai`
`openai`	OpenAI	`openai`	`https://api.openai.com`
`openrouter`	OpenRouter	`openai`	`https://openrouter.ai/api`
`perplexity`	Perplexity	`openai`	`https://api.perplexity.ai`
`together`	Together	`openai`	`https://api.together.xyz`
`xai`	xAI	`openai`	`https://api.x.ai`

Anthropic

The anthropic recipe, documented at https://docs.anthropic.com/en/api:

providers:
  - name: anthropic
    dialect: anthropic
    base_url: https://api.anthropic.com
    api_key: ${secret:ANTHROPIC_API_KEY}
    models: ["claude-*"]

Anthropic, OpenAI-compatible surface

Anthropic’s own translation of the Chat Completions API. Anthropic documents it as a migration aid rather than a production surface: strict tools, response formats, prompt caching, extended thinking and audio input are dropped, and usage reports no detail splits. An OpenAI-dialect client reaching Claude through the anthropic recipe goes through the gateway’s own translator instead, which keeps strict tools, structured output and thinking and refuses what it cannot carry.

The anthropic-openai recipe, documented at https://platform.claude.com/docs/en/api/openai-sdk:

providers:
  - name: anthropic-openai
    dialect: openai
    base_url: https://api.anthropic.com
    api_key: ${secret:ANTHROPIC_API_KEY}
    models: ["claude-*"]

Azure OpenAI

The v1 API, served under /openai/v1 in the OpenAI request shape with no api-version parameter. Replace YOUR-RESOURCE-NAME with the resource name. A request’s model is a deployment name, so models lists the deployment names your callers ask for. The key is sent as a bearer token, the way OpenAI’s own client sends it to this API. For Microsoft Entra ID, name an oauth credential instead; the custom provider how-to has the entry.

The azure-openai recipe, documented at https://learn.microsoft.com/en-us/azure/foundry/openai/api-version-lifecycle:

providers:
  - name: azure-openai
    dialect: openai
    base_url: https://YOUR-RESOURCE-NAME.openai.azure.com/openai
    api_key: ${secret:AZURE_OPENAI_API_KEY}
    models: ["gpt-*", "o3*", "o4*"]

Amazon Bedrock Mantle, Anthropic models

Bedrock Mantle serves the Anthropic Messages API under the /anthropic path, with Bedrock API keys and Workspaces. AWS recommends the bedrock-runtime recipe for new applications. Mantle also serves count_tokens, the only token-counting path for Claude models offered cross-Region only. Replace us-east-1 with your region. The key is a Bedrock API key; for SigV4 signing, name an aws credential instead and set region to the signing region.

The bedrock recipe, documented at https://docs.aws.amazon.com/bedrock/latest/userguide/inference-messages-api.html:

providers:
  - name: bedrock
    dialect: anthropic
    base_url: https://bedrock-mantle.us-east-1.api.aws/anthropic
    api_key: ${secret:AWS_BEARER_TOKEN_BEDROCK}
    models: ["anthropic.claude*"]

Amazon Bedrock, OpenAI-compatible models

Bedrock Mantle’s Chat Completions surface, for the model families it serves that way. Replace us-east-1 with your region.

The bedrock-openai recipe, documented at https://docs.aws.amazon.com/bedrock/latest/userguide/inference-chat-completions-mantle.html:

providers:
  - name: bedrock-openai
    dialect: openai
    base_url: https://bedrock-mantle.us-east-1.api.aws
    api_key: ${secret:AWS_BEARER_TOKEN_BEDROCK}
    models: ["openai.gpt-oss*", "minimax.*"]

Amazon Bedrock runtime, Anthropic models

The endpoint AWS recommends for new applications. It serves the Anthropic Messages API under /anthropic, with IAM and invocation logging. Replace us-east-1 with your region. The key is a Bedrock API key or a short-term Bedrock bearer token. Model IDs carry the inference profile prefix, such as us. or global., where a model is served cross-Region.

The bedrock-runtime recipe, documented at https://docs.aws.amazon.com/bedrock/latest/userguide/inference-messages-api.html:

providers:
  - name: bedrock-runtime
    dialect: anthropic
    base_url: https://bedrock-runtime.us-east-1.amazonaws.com/anthropic
    api_key: ${secret:AWS_BEARER_TOKEN_BEDROCK}
    models: ["anthropic.claude*", "us.anthropic.claude*", "eu.anthropic.claude*", "global.anthropic.claude*"]

Cerebras

The cerebras recipe, documented at https://inference-docs.cerebras.ai:

providers:
  - name: cerebras
    dialect: openai
    base_url: https://api.cerebras.ai
    api_key: ${secret:CEREBRAS_API_KEY}
    models: ["llama*", "qwen*"]

DeepSeek

The deepseek recipe, documented at https://api-docs.deepseek.com:

providers:
  - name: deepseek
    dialect: openai
    base_url: https://api.deepseek.com
    api_key: ${secret:DEEPSEEK_API_KEY}
    models: ["deepseek-*"]

Fireworks

The fireworks recipe, documented at https://docs.fireworks.ai/api-reference:

providers:
  - name: fireworks
    dialect: openai
    base_url: https://api.fireworks.ai/inference
    api_key: ${secret:FIREWORKS_API_KEY}
    models: ["accounts/fireworks/*"]

Google Gemini

Google’s OpenAI-compatible endpoint is served under /v1beta/openai and takes /chat/completions without the /v1 prefix, which strip_path_prefix removes from the incoming path.

The gemini recipe, documented at https://ai.google.dev/gemini-api/docs/openai:

providers:
  - name: gemini
    dialect: openai
    base_url: https://generativelanguage.googleapis.com/v1beta/openai
    strip_path_prefix: /v1
    api_key: ${secret:GEMINI_API_KEY}
    models: ["gemini-*"]

Groq

Groq rejects three Chat Completions fields rather than ignoring them, so they are removed from every request on the way out.

The groq recipe, documented at https://console.groq.com/docs/openai:

providers:
  - name: groq
    dialect: openai
    base_url: https://api.groq.com/openai
    remove_request_fields: [logit_bias, logprobs, top_logprobs]
    api_key: ${secret:GROQ_API_KEY}
    models: ["llama*", "mixtral*", "gemma*"]

Mistral

The mistral recipe, documented at https://docs.mistral.ai/api:

providers:
  - name: mistral
    dialect: openai
    base_url: https://api.mistral.ai
    api_key: ${secret:MISTRAL_API_KEY}
    models: ["mistral-*", "codestral-*", "magistral-*"]

OpenAI

The openai recipe, documented at https://platform.openai.com/docs/api-reference:

providers:
  - name: openai
    dialect: openai
    base_url: https://api.openai.com
    api_key: ${secret:OPENAI_API_KEY}
    models: ["gpt-*", "o1*", "o3*", "o4*"]

OpenRouter

The openrouter recipe, documented at https://openrouter.ai/docs:

providers:
  - name: openrouter
    dialect: openai
    base_url: https://openrouter.ai/api
    api_key: ${secret:OPENROUTER_API_KEY}
    models: ["*"]

Perplexity

Perplexity serves /chat/completions at the root of its host, without the /v1 prefix.

The perplexity recipe, documented at https://docs.perplexity.ai/api-reference:

providers:
  - name: perplexity
    dialect: openai
    base_url: https://api.perplexity.ai
    strip_path_prefix: /v1
    api_key: ${secret:PERPLEXITY_API_KEY}
    models: ["sonar*"]

Together

The together recipe, documented at https://docs.together.ai/reference:

providers:
  - name: together
    dialect: openai
    base_url: https://api.together.xyz
    api_key: ${secret:TOGETHER_API_KEY}
    models: ["meta-llama/*", "Qwen/*", "deepseek-ai/*"]

xAI

The xai recipe, documented at https://docs.x.ai/api:

providers:
  - name: xai
    dialect: openai
    base_url: https://api.x.ai
    api_key: ${secret:XAI_API_KEY}
    models: ["grok-*"]