claude-api

Use the official Anthropic SDK for your language and enable streaming for requests with long input or high max_tokens. The Skill detects your project language from files like pyproject.toml or package.json and provides the current SDK pattern, including the .get_final_message() helper for complete responses.

Invoke the migrate subcommand, which reads the model-migration guide and applies per-target breaking changes in order. It also audits prompt text and tool descriptions written for the source model, since dated prompting is part of every migration.

The Skill defaults to claude-opus-5 with adaptive thinking unless you explicitly name another model. Model IDs must be copied exactly from the current models table without date suffixes, and older model IDs are resolved through the models reference file.

Yes, but feature availability differs per platform and is tracked in a dedicated platform-availability reference. Claude Platform on AWS offers same-day API parity with SigV4 authentication, while Bedrock and Vertex AI are partner-operated with separate pricing and feature subsets.

Use the cost-optimize subcommand, which profiles token usage via the Usage and Cost Admin API or application logs, then ranks savings levers. Free wins like prompt caching and batch processing come first, followed by tradeoff levers such as effort levels and model choice validated against an eval.

Managed Agents fits when you want Anthropic to run the agent loop and host a per-session sandbox with persisted, versioned configs. The Tool Runner is harness-only: it automates the tool-call loop for tools you define, but you host and deploy everything yourself.