FiftyOne

FiftyOne is Voxel51's set of skills that teach your agent to work with computer vision datasets in FiftyOne, the open-source tool for building and curating visual datasets and evaluating models. Paired with the FiftyOne MCP server, your agent can load a dataset, clean it, run models on it and explain where the model fails, from a plain-language request. What you can do with it - Import datasets in any common format (COCO, YOLO, VOC, CVAT, KITTI and more), multimodal groups, LeRobot episodes and Hugging Face Hub datasets - Find and remove duplicate or near-duplicate images - Run zoo models for detection, classification, segmentation and embeddings - Evaluate predictions with mAP, precision, recall and confusion matrices, and look at false positives and negatives - Visualize embeddings in 2D to find clusters and outliers, and run end-to-end curation: quality checks, annotation audits, class balance and splits - Export to COCO, YOLO, VOC, CVAT, CSV or the Hugging Face Hub - Build FiftyOne plugins, notebooks and remote zoo models, and troubleshoot common setup problems How it connects - Install as a Claude plugin, or as a Gemini CLI extension from voxel51/fiftyone-skills - Skills that touch datasets need the FiftyOne MCP server (pip install fiftyone-mcp-server), running in the same Python environment as FiftyOne - Development skills (plugins, notebooks, code style) work without the MCP server Good to know - FiftyOne runs locally on your machine; datasets stay where you keep them - Some skills need FiftyOne plugins installed, such as @voxel51/brain for duplicate search - Apache 2.0 licensed