The authors built nninteractive-agent, an open-source system that exposes nnInteractive's segmentation capabilities as tools over the Model Context Protocol (MCP), together with a skill document encoding expert prompting heuristics.
AI agents act as zero-shot 3D annotators using a promptable model
The authors show that Claude Code and Codex, equipped with tools and a skill document, can locate, prompt, inspect, and correct segmentations from plain language, achieving mean Dice of 0.81–0.84 across diverse modalities and species.
Academic
Fabian Isensee · Constantin Ulrich Harsy · Maximilian Rokuss · Benjamin Hamm · Philipp Vollmuth · Klaus H. Maier-Hein
German Cancer Research Center (DKFZ) · National Center for Tumor Diseases (NCT) Heidelberg · Heidelberg University · University Hospital Bonn · Heidelberg University Hospital
Research Digest··3 min read
Isensee et al.
§