Redouble AI Nucleo documentation
Part 4 · Models

Claude on Bedrock through the Anthropic SDK

ai.redouble.nucleo.providers.bedrock.anthropic · nucleo-provider-bedrock-anthropic

The Bedrock library already reaches Claude through the Converse API (Amazon Bedrock). This add-on, nucleo-provider-bedrock-anthropic, reaches Claude on Bedrock through Anthropic's own SDK instead, for what Converse does not carry: extended thinking, prompt caching, and Bedrock's Mantle endpoint. The Anthropic SDK has no native-image support, which is why the add-on is a library of its own: a native build leaves it out and keeps Converse.

The credentials

The same as the Bedrock library's: an AWS identity and a region (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION, or AWS's own resolution), shared by every Bedrock provider, so a deployment configured for one is configured for all of them.

What it serves

Two providers, both for Claude alone:

The library ships a small catalog of its own with the Claude entries of both, used when the deployment has no models.json of its own. Without the add-on those entries are absent, and Claude on Bedrock is served through Converse. A catalog the discovery wrote with the add-on on the classpath still loads in a build without it: its -bedrock Claude entries are served through Converse, and its -mantle entries are left out.

Mantle projects

Every Mantle call runs under a Mantle project, and the project decides whether prompts are retained. The deployment names two in its settings (ModelSettings, set in the configurator described in Settings and the configurator):

The application's compliance envelope decides whether such a model may serve at all (DefaultComplianceEnvelope.setAllowDataShare), and the Mantle client checks it once more as the last step before a request leaves the process.

How it works inside

The classes, how entries link to a provider when a classpath lacks one, and the tests are in Inside Claude on Bedrock through the Anthropic SDK, for those working on the runtime itself.

^