Chamfer treats a model call like any other query. You choose the model, write or generate the prompt, and decide which fields from your data it can see. The Bench can add model steps to an app from a sentence, such as “summarize each ticket in two lines” or “pull the invoice number and total from the uploaded PDF”.

Because model calls run inside the app’s permissions, a user can never send a model data they couldn’t see on screen. Specs can keep fields such as card numbers or diagnoses out of every prompt. Each call is in the Logbook with the prompt, response, token count and cost, and Operators can use OpenAI models to work a queue on a named person’s authority. Structured outputs let a model fill a form for a person to check, instead of writing straight to a table, and a budget per app or team keeps spend predictable.