MCP server that experiments with cheaper LLM policies (model, prompt, reasoning, retry, escalation) against project tasks to recommend the lowest-cost policy that meets quality constraints.
