Agent turns follow:

1. User message → prepare LLM messages (tools + MCP).
2. Model returns text and/or a tool call (native or XML).
3. Elcro validates params, may request approval, executes tool.
4. Tool result is appended; loop continues until the model stops.

Local / OpenAI-compatible providers get a tool-calling mediator: native tool_calls preferred, with XML dual-parse and a one-shot retry when the model talks about tools but does not call them.