Omni-LPR is a specialized license plate recognition server with 5 tools. All tools have clear action-verb names (recognize_, detect_, list_) and reasonable descriptions (80-150 chars each). Input schemas are present with proper enum constraints for model selection. However, there are significant gaps: (1) No output schemas documented, the server does not declare what fields will be returned from recognition results; (2) Parameter descriptions are minimal and do not explain formats, error cases, or what constitutes a valid image; (3) No error handling patterns, no guidance on what happens when an image is invalid, unreadable, or contains no plate; (4) No acknowledgment of dependencies between tools (e.g., when to use recognize_plate vs. detect_and_recognize_plate); (5) Tools 1-4 show high redundancy (recognize_plate and recognize_plate_from_path do nearly the same thing, as do detect variants) but lack composition guidance. The schema presence (enums for models) saves this from a lower score, but missing output documentation and error recovery patterns are material gaps for LLM reliability.

• Document output schema for each tool. For recognize_plate, explicitly declare: 'Returns object {plate_text: string, confidence: number (0-1), bounding_box?: {x, y, w, h}, model_used: string}' or equivalent. For list_models, declare the structure of model objects returned.
• Consolidate recognize_plate and recognize_plate_from_path into a single 'recognize_plate' tool with a flexible 'image' parameter that accepts base64 string OR file path OR HTTPS URL. Handle type detection internally. Reduces cognitive load on LLM and eliminates near-duplicate tools.
• Similarly, consolidate detect_and_recognize_plate and detect_and_recognize_plate_from_path into 'detect_and_recognize' tool with flexible image input.
• Expand parameter descriptions with concrete constraints: 'image_base64: Base64-encoded JPEG or PNG image (max 10 MB, min 100x100 pixels, containing at least one license plate for best results)'. Add guidance: 'If image contains multiple plates, detection variant returns all; recognition variant processes first detected plate only.'
• Add error recovery guidance to tool descriptions: 'Returns empty plate_text if no plate detected. Returns null confidence if OCR quality is too low. Retry with a clearer image or different lighting conditions.'
• Document model selection criteria in enum descriptions: 'yolo-v9-s-608: Highest accuracy, slower (500ms typical). yolo-v9-t-256: Fast, lower accuracy (100ms typical). Choose based on accuracy vs. latency tradeoff.' Add to parameter description, not enum label itself.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry