Official DINO-X Model Context Protocol (MCP) server that empowers LLMs with real-world visual perception through image object detection, localization, and captioning APIs.
DINO-X MCP has four well-structured vision detection tools with clear naming conventions (detect-*, visualize-*) that follow verb-noun patterns. All tools have descriptions and input schemas are visible in the source code. However, there are several quality gaps: (1) Parameter descriptions are present but lack specificity about constraints, ranges, and format expectations. For example, 'textPrompt' description says 'avoid adjectives' but doesn't explain what happens if adjectives are included or define the max length. (2) Output schemas are not formally documented, responses are hardcoded text concatenations without a documented return structure, forcing LLMs to parse free-text results. (3) No error recovery guidance, errors return raw API error messages without actionable next steps. (4) Three of four tools lack parameter validation details (e.g., imageFileUri URI format, boolean parameter defaults). (5) The visualize tool's 'detections' parameter references an array structure but doesn't document the expected shape of detection objects.
Analyze an image to detect all identifiable objects, returning the category, count, coordinate positions and detailed descriptions for each object.
Detects 17 keypoints for each person in an image, supporting body posture and movement analysis.
Analyze an image based on a text prompt to identify and count specific objects, and return detailed descriptions of the objects and their 2D coordinates.
Visualize detection results by drawing bounding boxes and labels on the original image. Images are saved to the directory specified by IMAGE_STORAGE_DIRECTORY environment variable.
Output schemas not formally documented. Tools return text responses via hardcoded JSON.stringify() without a documented return type contract. LLMs cannot introspect what fields to expect, forcing them to parse unstructured text.
Parameter constraints not specified. 'imageFileUri' must start with 'https://' or 'file://' but the Zod schema only declares z.string() with generic description. No length limits, format validation regex, or pattern constraints. 'textPrompt' has no max length, no guidance on multi-word separation beyond 'periods'.
visualize-detection-result 'detections' parameter is typed as array but no schema describes the shape. Expected properties (name, bbox, etc.) are inferred from tool code but not documented. LLMs cannot know what structure to provide.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 66 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 43 | - | v1 |
No error recovery guidance. Errors return raw API failures (e.g., 'Failed to detect objects from image: <error message>') without actionable guidance. LLM doesn't know if the error is retryable, if input validation failed, or what to try next.
boolean 'includeDescription' parameter in three detection tools lacks default behavior documentation. Unclear whether false omits the 'caption' field, returns null, or includes it anyway. No guidance on when to use each mode (performance vs. detail trade-off).
visualize-detection-result stores output to IMAGE_STORAGE_DIRECTORY but no return value documents the output path or URI. LLM cannot discover or reference the generated visualization image.