DINO-X

Detects and localizes objects in images using natural language text prompts (English noun phrases) by routing requests through the DINO-X API client, which performs open-vocabulary detection without requiring pre-defined class lists. The MCP server wraps the detect-objects-by-text tool, accepting image URIs and text queries, then returns bounding box coordinates, confidence scores, and optional region-level captions for each detected object.

Performs comprehensive object detection across an entire image without requiring text prompts, using DINO-X's open-vocabulary capabilities to identify all detectable objects in a scene. The detect-all-objects tool invokes the DINO-X API with only an image URI, returning a complete set of detected objects with categories, bounding boxes, confidence scores, and optional captions for all regions.

Estimates human body pose by detecting 17 keypoints (head, shoulders, elbows, wrists, hips, knees, ankles) and returning their normalized coordinates. The detect-human-pose-keypoints tool sends images to the DINO-X API, which performs pose estimation and returns keypoint coordinates, confidence scores per keypoint, and optional bounding boxes for detected persons.

Generates annotated images with visual overlays of detection results (bounding boxes, keypoints, labels) by accepting detection output and rendering it onto the original image. The visualize-detection-result tool processes detection JSON and returns a local file path to the annotated image in STDIO mode, enabling agents to produce human-readable visual outputs for debugging or reporting.

Implements the Model Context Protocol v1.17.1 specification through two mutually exclusive transport modes: STDIO (for direct client integration) and HTTP (for remote deployment). The entry point at src/index.ts parses command-line arguments and instantiates either MCPStdioServer or MCPStreamHTTPServer, both delegating protocol handling to the @modelcontextprotocol/sdk package while registering tool handlers that invoke DINO-X API methods.

Encapsulates HTTP communication with the DINO-X platform through the DinoXApiClient class, handling authentication via API key, request serialization (image URIs and parameters), response deserialization, and error handling. The client abstracts DINO-X API details from tool handlers, providing typed method interfaces for detect-objects-by-text, detect-all-objects, and detect-human-pose-keypoints operations.

Manages server configuration through environment variables (DINOX_API_KEY, DINOX_API_BASE_URL) and command-line arguments (--stdio, --http, --port) parsed by the parseArguments() function in src/index.ts. Configuration is validated at startup and used to instantiate the appropriate server transport and API client, enabling flexible deployment across different environments without code changes.

Accepts image URIs in multiple formats (HTTP/HTTPS URLs and local file paths in STDIO mode) and resolves them to image data for API requests. The utilities module handles URI parsing and format validation, enabling agents to reference images from web sources or local filesystem depending on transport mode, with automatic format detection and error handling for invalid or inaccessible images.

Standardizes detection output across all tools by serializing results to JSON with normalized bounding box coordinates ([x_min, y_min, x_max, y_max] in 0-1 range) and per-detection confidence scores. The utilities module handles coordinate normalization and JSON formatting, ensuring consistent output format across text-prompted detection, open-vocabulary detection, and pose estimation, enabling downstream tools to parse results without format-specific logic.

Implements error handling throughout the MCP server by catching API failures, validation errors, and transport exceptions, then returning structured MCP error responses with error codes and human-readable messages. Tool handlers validate input parameters (image URIs, text queries) and propagate DINO-X API errors to clients through the MCP protocol, enabling graceful failure handling without server crashes.