MedMCP: An Open Agentic Ecosystem for Medical Imaging Workflows
- 1Technical University of Munich
Abstract
Deep learning has driven rapid progress in medical image analysis, with methods for tasks such as organ segmentation, anomaly detection, and radiology report generation now matching or surpassing expert performance on several benchmarks. While many methods are released as open source, their routine adoption by clinicians and medical researchers remains limited. The bottleneck is not a lack of capable tools, but the absence of infrastructure to make them accessible to non-technical users. To break this barrier, we present MedMCP, an agentic framework that solves this last-mile problem along two axes: (1) a local agentic workspace that lets users apply state-of-the-art image analysis methods through natural language, and (2) a community-driven tool ecosystem that lets developers deploy and distribute capabilities as validated, independently testable servers. MedMCP enforces a strict orchestration-execution boundary in which a local language model translates user intent into workflow plans, while all image analysis is delegated to domain-specific tools. The framework runs entirely on-premise, every side-effecting operation passes through a human-in-the-loop permission gate, and a full provenance trail links each output to the tools, parameters, and inputs that produced it. An inspectable workspace lets users verify intermediate results rather than blindly trusting the agent, and completed sessions distill into deterministic, shareable workflows that re-execute without the agent, turning one-off interactions into reproducible analyses. This paper introduces the project's general idea and describes the design principles that guide its development. Beyond a single toolkit, we envision MedMCP as a community-driven platform where each contributed tool becomes easily accessible, compounding individual efforts into shared capability.
Related
- TIAgent: From Natural Language to Grounded Agentic Workflows for Computational Pathology
- Structure Before Intelligence: Gating What Agentic Scientific Workflows Release
- Agentic transcranial functional ultrasound imaging: @fUS
- Agentic LLM reasoning in pediatric diabetes outcome prediction from multimodal medical records
- The electro-MICA toolbox for integrating electrophysiology within multimodal imaging and connectomics workflows
BibTeX
@misc{mcginnis2026medmcp,
title = {{MedMCP: An Open Agentic Ecosystem for Medical Imaging Workflows}},
author = {J. McGinnis and P. Friedrich and M. Muhlau and J. S. Kirschke and M. Styner and D. Rueckert and P. C. Cattin and B. Wiestler},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.12.758708},
url = {https://www.biorxiv.org/content/10.64898/2026.10.12.758708v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - MedMCP: An Open Agentic Ecosystem for Medical Imaging Workflows
AU - J. McGinnis
AU - P. Friedrich
AU - M. Muhlau
AU - J. S. Kirschke
AU - M. Styner
AU - D. Rueckert
AU - P. C. Cattin
AU - B. Wiestler
PY - 2026
DA - 2026/10/12
PB - bioRxiv
DO - 10.64898/2026.10.12.758708
UR - https://www.biorxiv.org/content/10.64898/2026.10.12.758708v1
AB - Deep learning has driven rapid progress in medical image analysis, with methods for tasks such as organ segmentation, anomaly detection, and radiology report generation now matching or surpassing expert performance on several benchmarks. While many methods are released as open source, their routine adoption by clinicians and medical researchers remains limited. The bottleneck is not a lack of capable tools, but the absence of infrastructure to make them accessible to non-technical users. To break this barrier, we present MedMCP, an agentic framework that solves this last-mile problem along two axes: (1) a local agentic workspace that lets users apply state-of-the-art image analysis methods through natural language, and (2) a community-driven tool ecosystem that lets developers deploy and distribute capabilities as validated, independently testable servers. MedMCP enforces a strict orchestration-execution boundary in which a local language model translates user intent into workflow plans, while all image analysis is delegated to domain-specific tools. The framework runs entirely on-premise, every side-effecting operation passes through a human-in-the-loop permission gate, and a full provenance trail links each output to the tools, parameters, and inputs that produced it. An inspectable workspace lets users verify intermediate results rather than blindly trusting the agent, and completed sessions distill into deterministic, shareable workflows that re-execute without the agent, turning one-off interactions into reproducible analyses. This paper introduces the project's general idea and describes the design principles that guide its development. Beyond a single toolkit, we envision MedMCP as a community-driven platform where each contributed tool becomes easily accessible, compounding individual efforts into shared capability.
ER -