The problem: When an AI agent needs to use API tokens, passwords, or keys on a remote server, the standard approach exposes secrets in the LLM's context window. The ...