Model integrations

Provider-specific integration work becomes easier to reason about when discovery, tool design, and execution stay separate. This collection uses fictional support, invoice, and product-catalog examples to examine those boundaries. A tool definition describes a capability; the application still decides whether and how it can run.

The Anthropic and OpenAI guides develop narrow read-only assistants. The OpenRouter guide proposes a stable fixture set for comparing candidate configurations without changing the task underneath the experiment. Each article points to a relevant primary reference and labels its surrounding architecture as a proposed design.

Choose the guide that matches your application rather than treating provider names as interchangeable. Review missing results, partial coverage, and failed requests before adding more operations. The integration overview connects these patterns, and the prompt evaluation collection provides a useful next step for checking final-answer fidelity.