Provider Support

Supported providers, API formats, and configuration details.

Overview

Your app authenticates to the gateway using a QuilrAI API key. Provider credentials are configured in the dashboard and never exposed to clients.

Capability Matrix

Provider	Chat	Embeddings	Rerank	TTS	STT	Responses	Realtime	Models	SDK / Webhook
OpenAI	✓	✓	-	✓	✓	✓	✓	✓	-
Azure OpenAI	✓	✓	-	✓	✓	✓	✓	✓	-
Anthropic (Chat Completions)	✓	-	-	-	-	-	-	✓	-
DeepSeek	✓	-	-	-	-	-	-	✓	-
Gemini (Chat Completions)	✓	-	-	-	-	-	-	✓	-
General LLM	✓	-	-	-	-	-	-	✓	-
Anthropic (Messages)	✓	-	-	-	-	-	-	✓	-
AWS Bedrock (OpenAI-compatible via Converse)	✓	-	-	-	-	-	-	✓	-
AWS Bedrock (Anthropic)	✓	-	-	-	-	-	-	✓	-
AWS Bedrock Runtime (boto3)	✓	-	-	-	-	-	-	✓	-
Azure (Anthropic Messages)	✓	-	-	-	-	-	-	✓	-
Vertex AI	✓	-	-	-	-	-	-	✓	-
AWS Bedrock (Embeddings)	-	✓	-	-	-	-	-	✓	-
Cohere Rerank	-	-	✓	-	-	-	-	✓	-
AWS Bedrock Rerank	-	-	✓	-	-	-	-	✓	-
Jina Rerank	-	-	✓	-	-	-	-	✓	-
Voyage Rerank	-	-	✓	-	-	-	-	✓	-
General Rerank	-	-	✓	-	-	-	-	✓	-
QuilrAI SDK	-	-	-	-	-	-	-	-	✓
Microsoft Copilot Studio	-	-	-	-	-	-	-	-	✓

Responses and Realtime are supported on dedicated provider types (openai_responses, openai_responses_azure, openai_realtime, openai_realtime_azure). A key configured for any other primary provider must add one of these as an additional provider on the key to access the Responses or Realtime endpoints.

Chat Completions

Endpoint: /openai_compatible/v1/chat/completions Auth: Authorization: Bearer sk-quilr-xxx

Provider	Auth Mode	Required Fields	Optional Fields
OpenAI	API Key	api_key	-
Azure OpenAI	API Key	api_key, azure_endpoint	azure_api_version
Anthropic (OpenAI-compatible)	API Key	api_key	-
DeepSeek	API Key	api_key	-
Gemini (OpenAI-compatible)	API Key	api_key	-
General LLM (vLLM, Ollama, etc.)	API Key	api_key, base_url	-
AWS Bedrock (Converse via OpenAI-compatible)	Static AWS Keys	aws_access_key, aws_secret_key	aws_region, aws_session_token
AWS Bedrock (Converse via OpenAI-compatible)	Assume Role	aws_role_arn, aws_external_id	aws_region, aws_role_session_name, aws_session_duration_seconds

The OpenAI-compatible chat endpoint is not limited to OpenAI-hosted models. In addition to providers that already expose an OpenAI-compatible upstream API, QuilrAI currently translates AWS Bedrock chat models into this surface by calling Bedrock Converse behind the scenes. Create a bedrock provider key, select any Bedrock model that supports Converse, and use that Bedrock model ID in the OpenAI SDK model parameter. For exact parameter, message, tool, and streaming coverage, see OpenAI to Bedrock Translation.

AWS Bedrock default region: us-east-1. For assume-role setup (trust policy, ExternalId, permissions), see AWS Bedrock - Assume Role Setup.

Anthropic Messages

Endpoint: /anthropic_messages/v1/messages Auth: x-api-key: sk-quilr-xxx

Provider	Auth Mode	Required Fields	Optional Fields
Anthropic (Native Messages API)	API Key	api_key	-
AWS Bedrock (Anthropic via Bedrock)	Static AWS Keys	aws_access_key, aws_secret_key	aws_region, aws_session_token
AWS Bedrock (Anthropic via Bedrock)	Assume Role	aws_role_arn, aws_external_id	aws_region, aws_role_session_name, aws_session_duration_seconds
Azure (Anthropic Messages)	API Key	api_key, azure_endpoint	azure_api_version

AWS Bedrock default region: us-east-1. For assume-role setup (trust policy, ExternalId, permissions), see AWS Bedrock - Assume Role Setup.

AWS Bedrock Runtime (boto3)

Endpoints: /model/{model_id}/converse, /model/{model_id}/converse-stream, /model/{model_id}/invoke Alternate prefix: /bedrock-runtime/model/{model_id}/... Auth: AWS SigV4 signed request using the QuilrAI key as both access key ID and secret access key

Use this surface when your application already calls Bedrock Runtime through boto3 or another AWS SDK. Configure a bedrock provider key in QuilrAI, then set the SDK endpoint_url to the closest regional endpoint, such as https://guardrails-usa-2.quilr.ai/bedrock-runtime.

If your application uses OpenAI-compatible clients instead, the same bedrock provider key can be called through /openai_compatible/v1/chat/completions; QuilrAI converts the OpenAI chat request to Bedrock Converse for you.

Provider	Auth Mode	Required Fields	Optional Fields
AWS Bedrock Runtime (boto3)	Static AWS Keys	aws_access_key, aws_secret_key	aws_region, aws_session_token
AWS Bedrock Runtime (boto3)	Assume Role	aws_role_arn, aws_external_id	aws_region, aws_role_session_name, aws_session_duration_seconds

converse supports any selected Bedrock model that supports the Bedrock Converse API. invoke_model schema coverage is limited to Amazon Nova, Anthropic, and OpenAI-style Bedrock models. Non-streaming converse and supported invoke_model calls run request and response DLP. converse_stream runs request-side DLP, then passes the AWS EventStream response through unchanged. invoke_model_with_response_stream is registered but returns ValidationException.

Only Bedrock Runtime is proxied. Bedrock control-plane APIs and Bedrock Agent Runtime APIs are not proxied. For setup and boto3 examples, see AWS Bedrock - boto3 Runtime.

Vertex AI

Endpoint: /vertex_ai/ Auth: Authorization: Bearer sk-quilr-xxx

Vertex AI supports multiple authentication modes. Select the mode when creating the key.

Auth Mode	Required Fields	Optional Fields	Notes
API Key	api_key, gcp_project_id	gcp_region	Default region: us-central1
Express	api_key	-	No project ID needed
Service Account	service_account_json	gcp_project_id, gcp_region	Project ID derived from JSON if omitted
ADC	gcp_project_id	gcp_region	Application Default Credentials from environment

Multimodal & image-capable Gemini models

Vertex is a native passthrough to generateContent, so any Gemini model configured on the key works - including multimodal models that accept image / audio / video inputs and image-output models like gemini-2.5-flash-image-preview. Add the model name to the key's selected_models list and call it the same way you would upstream.

Request-side DLP scans text parts of the request. Non-text parts (image / audio / video bytes) and image / audio outputs pass through without response-side DLP - the guardrails pipeline is text-focused today.

TTS & STT

Endpoints: /openai_compatible/v1/audio/speech and /openai_compatible/v1/audio/transcriptions

Provider	TTS	STT	Auth Mode	Required Fields
OpenAI	✓	✓	API Key	api_key
Azure OpenAI	✓	✓	API Key	api_key, azure_endpoint

STT also supports /v1/audio/translations. Azure deployments use the /openai/deployments/{deployment}/ path prefix.

Embeddings

Endpoint: /openai_compatible/v1/embeddings Auth: Authorization: Bearer sk-quilr-xxx

Provider	Auth Mode	Required Fields	Optional Fields
OpenAI	API Key	api_key	-
Azure OpenAI	API Key	api_key, azure_endpoint	azure_api_version
AWS Bedrock (Embeddings)	Static AWS Keys	aws_access_key, aws_secret_key	aws_region, aws_session_token
AWS Bedrock (Embeddings)	Assume Role	aws_role_arn, aws_external_id	aws_region, aws_role_session_name, aws_session_duration_seconds

AWS Bedrock default region: us-east-1. Supports Titan and Cohere Embed families on Bedrock; requests and responses follow the OpenAI embeddings shape. For assume-role setup, see AWS Bedrock - Assume Role Setup.

Rerank

Endpoints: /rerank/v2/rerank, /rerank/v1/rerank, /rerank/rerank Auth: Authorization: Bearer sk-quilr-xxx

All three paths are registered to match Cohere's upstream routes (v2, v1, and legacy). Accepts a Cohere-compatible body (model, query, documents, optional top_n, return_documents) and returns a Cohere-shaped response (id, model, results[], usage).

Provider	Auth Mode	Required Fields	Optional Fields	Notes
Cohere Rerank	API Key	api_key	-	-
AWS Bedrock Rerank	Static AWS Keys	aws_access_key, aws_secret_key	aws_region, aws_session_token	Cohere Rerank 3.5 and Amazon Rerank families; reuses bedrock:InvokeModel IAM permission
AWS Bedrock Rerank	Assume Role	aws_role_arn, aws_external_id	aws_region, aws_role_session_name, aws_session_duration_seconds	See AWS Bedrock - Assume Role Setup
Jina Rerank	API Key	api_key	-	-
Voyage Rerank	API Key	api_key	-	-
General Rerank	API Key	api_key, base_url	-	Self-hosted ColBERT / TEI / Infinity exposing a Cohere-shaped /rerank endpoint

Request-side DLP scans the query and documents fields. Response-side DLP is not applied - responses are scores and indices only.

Responses API

Endpoint: /openai_responses/v1/responses Auth: Authorization: Bearer sk-quilr-xxx

Native passthrough for OpenAI's Responses API. Create / retrieve / cancel / delete / list-input-items are all supported.

Provider	Auth Mode	Required Fields	Optional Fields
OpenAI (Responses)	API Key	api_key	base_url
Azure OpenAI (Responses)	API Key	api_key, azure_endpoint	azure_api_version

Azure-deployment-style aliases are also accepted: /openai_responses/openai/deployments/{deployment}/responses[/{response_id}[/cancel|/input_items]]. The deployment name goes in body.model regardless of which URL shape is used.

Request-side DLP scans free-form user text inside input_text parts of input plus the top-level instructions. previous_response_id and built-in tools (web_search, file_search, computer_use, code_interpreter) are passthrough. Response-side DLP scans output_text parts on non-streaming responses; streaming responses bypass response-side DLP by design (request-side DLP still runs).

Realtime API

Endpoint: wss://<base>/openai_realtime/v1/realtime Auth: Authorization: Bearer sk-quilr-xxx

Native passthrough for OpenAI's Realtime websocket API (voice and text).

Provider	Auth Mode	Required Fields	Optional Fields
OpenAI (Realtime)	API Key	api_key	base_url
Azure OpenAI (Realtime)	API Key	api_key, azure_endpoint	azure_api_version

Compatibility aliases for SDK and browser clients are registered at /openai/v1/realtime, /openai/realtime, /openai_realtime/openai/v1/realtime, and /openai_realtime/openai/realtime.

Quilr accepts the API key in any of the following forms (in priority order) to cover both server and browser clients:

1. Authorization: Bearer sk-quilr-xxx header
2. api-key: sk-quilr-xxx header
3. api-key / api_key query parameter
4. authorization query parameter
5. WebSocket subprotocol openai-insecure-api-key.sk-quilr-xxx

The openai-insecure-api-key.* subprotocol is stripped before forwarding and is never sent upstream.

Guardrails coverage

Realtime sessions are passthrough today - DLP is not yet applied to live Realtime events in either direction. Request-side and response-side guardrails on Realtime are planned; until then use Realtime for voice/text flows that do not require in-session redaction. Session-level logging (handshake status, byte counters, usage summary) is still written.

Selecting a Provider on Multi-Provider Keys

A key can have one primary provider plus any number of additional providers of the same or different kind. When more than one compatible provider is configured, you can pick which one handles a request; if you don't pick, QuilrAI uses the first compatible provider on the key.

Endpoint	Body field	Header	Query param
Chat Completions / Anthropic Messages / Vertex / Embeddings / Rerank	provider or provider_label	X-Provider-Name / X-Provider-Label	-
Responses	provider or provider_label	X-Provider-Name / X-Provider-Label	-
Realtime (websocket)	-	X-Provider-Name / X-Provider-Label	provider or provider_label

Match by either the provider type (bedrock, openai_responses_azure, openai_realtime, anthropic_messages_bedrock, bedrock_embeddings, cohere_rerank, bedrock_rerank, jina_rerank, voyage_rerank, general_rerank, etc.) or the label you assigned to the additional provider when you added it in the dashboard.

SDK

API Endpoint: /sdk/v1/check Auth: Authorization: Bearer sk-quilr-xxx

The SDK provides guardrails-only scanning - no upstream LLM provider needed. Check text for PII, PHI, adversarial prompts, and custom intents without forwarding to any model.

Python

pip install quilrai

JavaScript

npm install quilrai

LiteLLM Proxy Plugin

QuilrAI integrates as a plugin for LiteLLM's proxy gateway. Configure it in your LiteLLM proxy config to add guardrails to all LLM traffic.

Microsoft Copilot Studio

Endpoint base: /copilot_studio/{sk-quilr-xxx} Routes: /validate, /analyze-tool-execution Auth: QuilrAI key in the endpoint path

Copilot Studio support is SDK-style external threat detection, not LLM proxying. Create a key with provider copilot_studio, configure the endpoint base in Power Platform admin center, and Copilot Studio calls QuilrAI before tool execution.

QuilrAI scans recent user prompt context and tool inputValues. It returns blockAction: true for block/redact/partial-redact outcomes because Copilot Studio cannot accept rewritten tool input. DLP timeout or internal DLP errors fail open with blockAction: false so transient service issues do not break the agent flow.

For setup steps, see Copilot Studio.