MCP server providing OpenCV computer vision capabilities
Server provides 12 tools with basic schemas and descriptions, but quality is inconsistent. Most tools have descriptions (10-50 words), but many lack depth on WHEN to use them or dependencies. Parameter descriptions are present but generic. No output schemas documented. Error handling is minimal, most tools raise ValueError with basic messages, offering no recovery guidance. Tool names follow verb_noun convention (good), but several parameters lack type constraints (e.g., 'filter_type', 'method' accept free-form strings instead of enums). No tool annotations (readOnlyHint, destructiveHint). Response objects are partially structured but include raw OpenCV metadata that wastes tokens. Composition is reasonable, tools are single-purpose, but chaining support is weak (e.g., detect_features returns 'output_path' for chaining, but most tools don't). Security: no secrets exposed, but no permission gates or audit logging. Baseline: 549 production tools average 18-char names (this server: 15-25 chars, acceptable) and 194-char descriptions (this server: 30-80 chars, below baseline). 94% of A+ tools have docstrings (this server: 100% have tool docstrings, but parameter descriptions are thin). 100% of A+ tools have documented return types (this server: 0% have explicit output schemas).
Apply various filters to an image
Apply threshold to an image
Convert image between color spaces
Crop a region from an image
Detect edges in an image
Detect features in an image
Detect motion between two frames
No output schemas documented. Tools return dicts with undocumented fields (e.g., detect_features returns 'keypoint_count', 'keypoints', 'method', 'info', 'path', 'output_path'). LLMs cannot plan downstream calls or extract the right data without knowing return structure.
Enum constraints missing. Parameters like 'method' (sift|orb|brisk|akaze), 'filter_type' (blur|gaussian|median|bilateral), 'threshold_type' accept free-form strings. LLMs hallucinate invalid values; no validation prevents 'method=xyz' from reaching OpenCV.
Error handling is minimal. Most tools raise ValueError with generic messages ('Failed to detect features: {str(e)}'). No recovery guidance, no categorization (retryable vs fatal), no invalid-value echoing. LLMs cannot self-correct.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | <=2025-11-25 | v2 |
Extract frames from a video file
Get statistical information about an image
Match features between two images
Resize an image to specific dimensions
Save an image to a file
Tool descriptions lack WHEN/WHY context. E.g., 'Detect edges in an image' does not explain when to use Canny vs Sobel, or what the output is used for. Descriptions average 40-60 chars; baseline is 194 chars. LLMs cannot distinguish similar tools.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot distinguish safe read-only tools from destructive ones (save_image_tool, extract_video_frames_tool write files). Spec 2026-07-28 expects these hints.