The short answer

AgentPlat’s optional Anthropic adapter implements the provider-neutral ModelAdapter contract over Anthropic’s Messages API. The application supplies the API key and model, and AgentPlat normalizes the model call for the selected runtime path. Your application remains responsible for data sharing, provider account controls and authorization of external effects.

Use the native Messages API adapter

@agentplat/model-anthropic connects AgentPlat to Anthropic’s Messages API without requiring an Anthropic SDK dependency. It implements the shared ModelAdapter interface, so an application can choose Anthropic for a model invocation while retaining the same AgentPlat execution boundary and surrounding application code.

The application supplies a model identifier and the API key, either when constructing the adapter or through the execution-scoped credential path. The adapter does not read environment variables. See the first execution guide for the application pattern and package reference for the AgentPlat package surface.

Keep data and credentials in the host application

The application decides which prompt, conversation context and artifact content to send to Anthropic. Keep API keys in your server-side secret mechanism, resolve the right credential for the current tenant or account, and avoid putting secrets into agent definitions or Room artifacts. A provider adapter does not decide that the application is authorized to share a particular record.

Anthropic controls its endpoint, account-level data settings, model availability and rate limits. Review those settings for the exact deployment and model. Test the behavior that matters to your product, including errors, timeouts, streaming and cancellation, rather than treating a provider-neutral type as a promise that every model has identical behavior.

Place policy around the invocation and action

When an application routes a model call through AgentPlat’s integrated execution path, it can apply provider-neutral controls around that invocation. A separate call made directly through another SDK is outside that path unless the application wraps or governs it too. The Inference Control guide explains the scope of these controls.

Model output remains a proposal or input to application logic. The application owns any policy that determines whether to show the answer, request human review or continue to a protected action. A successful Anthropic response does not authorize an agent to send a message or change an external system.

Select models with application-specific evidence

Use a small representative evaluation to choose a model for the task: define expected behavior, include difficult inputs, review failures and measure the operational dimensions your application needs. Pin a model identifier when reproducibility matters and record the provider configuration with the run.

AgentPlat provides the adapter boundary; it does not claim that Anthropic is better or worse than another provider. Quality, cost, latency and service limits depend on model version, workload, account and deployment. Check the current AgentPlat release channels before installing and follow Anthropic’s current API documentation for provider-specific behavior.

Sources and further reading

Documentation reviewed . Consult the linked documentation for current implementation details.