AWS Bedrock

Install:

shell

pip install "nullrun[bedrock]"

bedrock_converse.py

import nullrun
import boto3

nullrun.init(api_key="nr_live_...")

client = boto3.client("bedrock-runtime", region_name="us-east-1")
resp = client.converse(
    modelId="anthropic.claude-3-sonnet-20240229-v1:0",
    messages=[{"role": "user", "content": [{"text": "Hello"}]}],
)

Patched at the boto3 transport layer; works for every model hosted on Bedrock (Anthropic, Mistral, Cohere, Meta, AI21, etc.). The extractor handles both the nested (response.usage with inputTokens / outputTokens) and the top-level (Anthropic-on-Bedrock flattens them onto the response body) shapes.

Test coverage

Per NullRun's testing policy (see the SDK README), AWS Bedrock patches have extractor unit tests only — no full transport-to-track integration test. The boto3 extractor shape is exercised by isolated tests, but no multi-roundtrip end-to-end test confirms the cost actually flows through /api/v1/track. The Use with Bedrock → Manual tracking recipe is the recommended pattern for production Bedrock usage. This is documented per §8 of the source-of-truth positioning.