A low-hallucination vision MCP server that uses OpenAI-compatible multimodal models with structured prompts, confidence gating, and forced JSON to reduce false claims in image analysis.