Amazon Bedrock

​

Set up

import { BedrockRuntimeClient, ConverseCommand } from '@aws-sdk/client-bedrock-runtime';

const bedrock = stigg.wrap(new BedrockRuntimeClient({ region: 'us-east-1' }), {
  stiggContext: { customerId: 'customer-acme', entitlement: { currencyId: 'credits' } },
});

await bedrock.send(
  new ConverseCommand({
    modelId: 'anthropic.claude-sonnet-4-5-20250929-v1:0',
    messages: [{ role: 'user', content: [{ text: 'Summarize this ticket' }] }],
    inferenceConfig: { maxTokens: 1024 },
  }),
);

• Use the Converse API. ConverseCommand and ConverseStreamCommand are checked and metered, for every model Bedrock serves. Other commands sent through the same client, such as InvokeModelCommand, go through without a check and are not metered. InvokeModel returns each model vendor’s own response format, so the SDK cannot read its usage.
• Set the stiggContext in wrap(). A Bedrock call takes a command object rather than a parameter object, so it has no place for a stiggContext. To serve several customers, wrap a client for each.
• A streamed call is reported once, when the stream finishes. Iterate the stream property of the response with for await.
• Credentials come from the standard AWS credential chain, as for any AWS SDK client.

import { bedrockRecipe } from '@stigg/agent-sdk-js';

const bedrock = stigg.wrap(new BedrockRuntimeClient({ region: 'us-east-1' }), {
  recipe: bedrockRecipe({ estimate: (command) => command?.input?.inferenceConfig?.maxTokens ?? 1 }),
  stiggContext: { customerId: 'customer-acme', entitlement: { currencyId: 'credits' } },
});

​

What is reported

• No cost. Unlike OpenRouter and LiteLLM, Bedrock does not return a price, so you can bill tokens but not what the call cost. The price appears only on your AWS bill.
• No model dimension. Bedrock serves exactly the model you name and does not fall back to another, so there is no routing drift to track.
• Cache tokens are separate. Like Anthropic, Bedrock counts cache reads and writes separately from inputTokens. AWS defines totalTokens as the input and output tokens, so on a call that uses the prompt cache it can be lower than the sum of the four counters. Price the total or the parts, never both.