Anthropic

This provider calls Claude models on Anthropic's own API, through Anthropic's Java SDK. Put the nucleo-provider-anthropic library on the classpath and provide one credential, and every catalog entry under the provider key anthropic-direct can serve work.

The credential

The API key, from ANTHROPIC_API_KEY, stored under the credential id anthropic-api-key. It is the only thing the provider needs: Set up and run the demo says how a credential reaches the runtime.

What it serves

Claude, with the features the Messages API carries: pictures and PDF documents sent as themselves, tool calls in the API's own format, the model's extended or adaptive thinking (set per catalog entry by thinking_mode, and passed back to the model on the next turn exactly as it came), and prompt caching, with the number of cache breakpoints the entry allows (cache_breakpoints).

The library ships a small catalog of its own: the current Claude model of each family on this API, with its list prices and the limits of Anthropic's Start tier, used when the deployment has no models.json of its own; the discovery writes the rest. Its entries are named for the channel, as in claude-opus-5-direct, so the same model on Bedrock is a separate entry (Grades, the catalog and the picker explains entries).

The discovery (Discovering an account's models) reads the API's model listing, which says which models the key may call. Anthropic lists no rate limits; each response carries the account's limits in its anthropic-ratelimit-* headers, and the discovery's one call per model writes them into the catalog.

How it works inside

How each part of a message becomes Anthropic's native content blocks is in Inside the Anthropic provider, for those working on the runtime itself.