Amazon Bedrock

What it is

Amazon Bedrock is a fully managed AWS service for building with foundation models from several providers, including Anthropic's Claude. You call a model through an AWS API instead of hosting it yourself, with the same IAM permissions, logging and billing as the rest of the account. There are no servers to run and you pay per token.

Around the models, Bedrock adds features for production use: guardrails that filter inputs and outputs, knowledge bases for retrieval, agents, and cross-region inference profiles that route a request to the model in whichever region has capacity. Model access is granted per account, so a team can control exactly which models its code is allowed to call.

How I use it

• Radar, the help assistant inside Rental Pilot, and the assistants on three company websites run on Claude through Bedrock, under a daily spend cap with every question logged.
• The agent on mervinc.com is a Lambda function calling Claude Sonnet on Bedrock through a cross-region inference profile, streaming each step to the page.
• The build crew runs five Claude agents on Bedrock: Haiku for the planner, designer, data and reviewer stages, Sonnet for the builder.
• Each function's IAM role names the inference profile and the model, and nothing else.
• Bedrock Guardrails checks every question before the model sees it and every answer after.

Bedrock is the right choice when the rest of the system already lives on AWS. Calls are authorized by an IAM role instead of an API key in a config file, costs land on the same bill, and guardrails sit next to the model. That is why my agents on this site and in Rental Pilot call Claude through Bedrock: one less secret, one less vendor, and the same least-privilege rules as every other function.

Questions