Can I call AWS Bedrock with the OpenAI SDK?

Yes. Store AWS credentials in Settings -> AI Providers, then send a Bedrock model ID (for example anthropic.claude-3-5-sonnet-20241022-v2:0 or amazon.nova-pro-v1:0) to https://api.cognocient.com/v1 with the OpenAI SDK. Cognocient AI signs the request with SigV4 and calls Bedrock's Converse API.

Do Bedrock cross-region inference profiles work?

Yes. IDs prefixed with us., eu., apac., global., us-gov., jp., au. or ca. are routed to Bedrock and priced as their base model.

Does streaming work with Bedrock through the gateway?

Yes. Streamed calls use ConverseStream; the AWS binary event stream is decoded and returned as standard OpenAI server-sent events, including streamed tool calls and a final usage chunk.

How do I put a hard budget on Bedrock spend?

AWS Budgets alert after the fact, hours later. A Cognocient AI budget is checked before each call is forwarded, so a feature, team or agent run that hits its limit is blocked, degraded to a cheaper model, or alerted on in real time.

AWS Bedrock with the OpenAI SDK: an LLM gateway for Claude, Nova and Llama on Bedrock

Call AWS Bedrock models (Claude, Amazon Nova, Llama, Mistral) with the OpenAI SDK through one gateway. SigV4 signing, Converse API translation, streaming and tool calls handled for you, with per-feature cost tracking and budgets.

Bedrock speaks its own API (Converse) and needs every request signed with AWS Signature Version 4. Cognocient AI does both, so your application keeps using the OpenAI SDK and a Bedrock model ID.

Setup

In Settings → AI Providers → AWS Bedrock enter an access key ID, secret access key and the region where you have model access enabled.
Give that IAM user bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on the models you use, and enable those models in the Bedrock console for that region.
Send Bedrock model IDs through the gateway:

How model IDs are recognised

No prefix is needed. A model is sent to Bedrock when it has the Bedrock shape <vendor>.<model> with one of these vendors: anthropic, amazon, meta, mistral, cohere, ai21, deepseek, writer, luma, stability, openai, qwen. An optional inference-profile prefix (us., eu., apac., global., us-gov., jp., au., ca.) is allowed, and is ignored for pricing.

A Bedrock ID needs a Bedrock key

Bedrock-shaped IDs only route to Bedrock when a Bedrock key is connected. Without one, the ID falls through to the default (OpenAI) route, which rejects it as an unknown model; if no OpenAI key exists either, the error names Bedrock as the missing provider. Connect the Bedrock key first.

What works

Feature	Notes
Chat, streaming	Converse and ConverseStream; OpenAI-format responses and SSE
Tool calling	OpenAI `tools` translated to Bedrock tool config; streamed tool calls supported
Token usage and cost	From Bedrock's own usage report
Responses API	Translated to chat like every provider

Why route Bedrock through a gateway?

Real-time limits. AWS Budgets and Cost Anomaly Detection report after the spend has happened. A budget here is checked before the call.
Attribution below the account level. Bedrock's bill shows one line per model. Headers like X-Cost-Feature and X-Cost-User split it by feature, team, user and your customers (Growth).
One client for everything. The same OpenAI client reaches Bedrock, Anthropic direct, OpenAI and Gemini, so moving a feature between them is a model-name change.
Reconcile against the AWS bill. Shadow Spend Reconciliation (Growth plan) compares what passed through the gateway with what AWS actually billed.

AWS Bedrock with the OpenAI SDK: an LLM gateway for Claude, Nova and Llama on Bedrock

Setup

How model IDs are recognised

What works

Why route Bedrock through a gateway?

On this page