Advanced medical image analysis?
Just ask for it.

MedMCP is an agentic system that runs state-of-the-art medical imaging pipelines from a familiar chat interface. No terminal, no code, just plain language.

Explore the workspace View on GitHub

Demo

MedMCP in action.

A short walkthrough of the workspace and its current capabilities, from a plain-language request to finished, inspectable results.

01

Describe your analysis

Use plain language, the same way you'd explain it to a colleague. No commands, environment management, or programming experience needed.

02

MedMCP plans and executes

MedMCP selects the right tools, sequences them, and handles all the technical details automatically.

03

You approve every action

Before MedMCP touches your data, you see exactly what is about to happen. Approve or reject with a single click. You always stay in full control.

The Agentic System

MedMCP doesn't just answer.
It does the work.

MedMCP is an agent, not a chatbot. It reasons about your goal, sequences the right tools, executes them under your control, and keeps a complete record of everything it did.

Plans multi-step pipelines

From one plain-language goal, it sequences the whole image analysis chain in the correct order.

Runs validated tools

It orchestrates trusted, published methods. Nothing is generated at runtime. The model plans, proven tools execute.

Approval-gated execution

Every file operation and tool call is shown for review first. Nothing touches your data until you approve it.

Full provenance & audit

Every action is logged to an append-only record: a complete, reproducible trail of what ran on your data.

Learns reusable workflows

Distill any session into a deterministic, replayable pipeline. Run it again on new patients with no model in the loop.

Batch over cohorts

Apply a workflow to a whole folder of subjects and process an entire cohort in a single run.

Who It's For

Built for the people who
work with medical images.

In the clinic and in the lab, MedMCP makes the field's best imaging tools available to the people who should use them.

Clinician & Radiologist

Research-grade tools, no terminal.

Run validated segmentation, registration, and lesion analysis on your own patients.

"Have this patient's MS lesions grown since their last scan?"
Domain Researcher

Focus on the science, not the setup.

Skip the Python environments and DICOM wrangling, and batch a whole cohort in one request.

"Convert this DICOM series to NIfTI, skull-strip it, and register it to MNI space."

Privacy & Safety

Your data stays inside
your institution.

MedMCP was designed for clinical and translational research environments, where data governance is non-negotiable. By default, everything runs on your own hardware. Outside services are optional, meant for research on public or non-sensitive data, and never switch on without your consent.

On-premise by default

The AI model runs on your hardware. Unless you connect an outside service yourself, no imaging data, patient metadata, or results are sent anywhere.

You approve every action

Every file operation, every analysis step, every tool call is shown to you before it runs. You stay in full control.

Outside services are opt-in

Cloud models and external tool servers are off by default and meant for work on public or non-sensitive data. Each needs your explicit consent.

No internet required

After installation, analyses run fully offline. MedMCP goes online for downloads you start: a new model, a stack, or an update.

Signed releases

MedMCP releases are cryptographically signed. The workspace verifies the signature before it updates itself and refuses anything that does not check out.

Full audit trail

Every approved and rejected action is logged persistently, along with the model used and where it ran. You get a complete record of what ran on your data and when.

Always Getting Better

Built to ride the wave of
open-weight model progress.

MedMCP is built around open-weight models that run on your own hardware. Muse Glimmer is the default, and six more open models are one click away in the workspace.

As stronger open models are released, they join the list, and your tools and workflows keep working unchanged. For research on public or non-sensitive data, you can also connect a cloud model with your own API key. It stays off unless you switch it on and confirm.

Default model
Muse Glimmer Default
Or switch in one click
Devstral Small 2
Gemma 4 31B
Granite 4.2 30B
Mistral Small 3.2
Nemotron 3.5 Lightning
Qwen3.8-27B
Cloud model, your API key Off by default

Our Vision

Make advanced medical imaging
usable by everyone who needs it.

"The tools exist. The methods work. What is missing is the infrastructure to put them in the hands of the people who are best positioned to use them. MedMCP exists to close that gap."

We believe that one of the biggest problems in medical image analysis is the systematic disconnect between well-working tools and the people who need them. No clinician or researcher should need a software engineer to run these pipelines. Our goal is simple: make the field's best methods usable by anyone who needs them, reliably and safely, on their own infrastructure.

Paper

The ideas behind MedMCP,
written up.

Page 1 of the MedMCP preprint Page 2 of the MedMCP preprint Page 3 of the MedMCP preprint

Preprint · 2026

MedMCP: An Open Agentic Ecosystem for Medical Imaging Workflows

Julian McGinnis*, Paul Friedrich*, Mark Mühlau, Jan S. Kirschke, Martin Styner, Daniel Rueckert, Philippe C. Cattin, Benedikt Wiestler

*Shared first authorship

The paper introduces the project and the design principles behind it: a local agentic workspace that runs medical image analysis from natural language, and a community-driven tool ecosystem that lets developers distribute capabilities as validated, independently testable servers.

Read the paper (PDF)

Capability Stacks

A growing library of
imaging capabilities.

MedMCP's capabilities are organized into domain stacks. Activate one and its tools appear automatically in the workspace, ready for the agent to plan with. All of this is work in progress: the first stacks are available, they keep evolving, and new ones land regularly.

medmcp-dicom Available

DICOM discovery, NIfTI conversion, and BIDS organisation. The foundation layer every other stack builds on.

ghcr.io/medmcp/dicom
View repo →
medmcp-neuro-core GPUAvailable

Brain extraction (HD-BET), image registration (ANTs), and whole-brain segmentation (FastSurfer).

ghcr.io/medmcp/neuro-core
View repo →
medmcp-neuro-ms GPUAvailable

Multiple sclerosis white-matter lesion segmentation and lesion load (LST-AI).

ghcr.io/medmcp/neuro-ms
View repo →
medmcp-totalsegmentator GPUAvailable

Whole-body segmentation of over 100 anatomical structures in CT and MR images (TotalSegmentator).

ghcr.io/medmcp/totalsegmentator
View repo →
medmcp-cardiac GPUAvailable

Cine cardiac MRI ventricle segmentation and function (CineMA), AHA wall analysis, and CT calcium scoring.

ghcr.io/medmcp/cardiac
View repo →
medmcp-cohort In progress

Clinical cohort assembly, imaging-biomarker stratification, and association analysis (ehrapy).

ghcr.io/medmcp/cohort
In active development
medmcp-spine In progress

Spinal cord segmentation, cross-sectional area, and lesion metrics for spine imaging workflows.

ghcr.io/medmcp/spine
In active development
medmcp-microscopy Coming soon

Cell detection, tissue segmentation, and quantification pipelines for digital pathology and microscopy images.

ghcr.io/medmcp/microscopy
On the roadmap
Need a specific pipeline or modality?
We're expanding the stack library across modalities. Tell us what your group needs prioritized.
Get in touch →

The Team

Built by researchers,
for researchers.

Julian McGinnis

Julian McGinnis

PhD Student

TU Munich · TUM University Hospital · MCML

julian.mcginnis@tum.de
Paul Friedrich

Paul Friedrich

Postdoctoral Researcher

University of Basel

paul.friedrich@unibas.ch

Get involved

Contributions
are welcome.

MedMCP is built in the open, and all of it is work in progress. New tools, new stacks, bug reports, and ideas are all welcome. Reach out for a demo or a chat, or jump straight into the code.

Get in touch Contribute on GitHub