Ecosystem#
KonfAI is the core, but several packages and tools sit around it. This page maps how they relate and — importantly — what is shipped versus what is external or experimental, so you know what you can rely on today.
The map#
Resolved YAML and medical data enter the KonfAI core. KonfAI Apps and KonfAI MCP both build on the core; MCP also operates Apps. Hugging Face artifacts feed Apps and the task-specific command-line tools. Slicer clients operate Apps and ImpactReg, while ONNX to konfai-rs remains experimental.
Data, transforms, models, losses, metrics, and workflow intent.
App configs, checkpoints, metadata, and demo datasets.
publishedkonfai
One declarative, patch-native engine for medical-imaging workflows.
konfai-apps
Resolve local or Hub Apps and run the same workflow through Python, CLI, REST, or SSE logs.
konfai-mcp
Operates core workflows and Apps through structured scientific tools.
Local and remote execution share one App contract.
ImpactSynth · ImpactSeg · ImpactReg · MRSegmentator · TotalSegmentator
Inspect, configure, validate, execute, compare, and resume.
Runs KonfAI Apps from 3D Slicer.
externalDrives the complete ImpactReg orchestrator.
externalkonfai-rs
Native and WebAssembly inference path.
experimentalReleased platform#
Piece |
Status |
What it is |
|---|---|---|
|
✅ Shipped (PyPI) |
The core framework: config-by-reflection, lazy patch-based data, model graphs, and the three YAML workflows. This is what the rest of this documentation is about. |
|
✅ Shipped (PyPI, own CI) |
Packages a mature workflow as a reusable app — see Using KonfAI Apps. |
|
✅ Shipped (PyPI, own CI) |
Operates KonfAI workflows and Apps through structured tools for dataset inspection, config authoring and validation, job execution, monitoring, metrics, and comparison — see Agent workflows (MCP server). |
App bundles ( |
✅ Shipped (thin wrappers) |
Ready-to-use CLI shims: |
|
🟡 Shipped, heaviest |
A full multi-preset registration orchestrator (Elastix + the IMPACT semantic metric), not a thin wrapper. The most moving parts of the five. |
Demo data & models (HF) |
✅ Published |
|
Challenge repos |
✅ External |
Top-ranking MICCAI-challenge projects built on KonfAI (SynthRAD 2025 T1/T2, TrackRAD 2025, Panther, CURVAS, CURVAS-PDACVI). Referenced from the README; not in this tree. |
Ready-to-use apps#
App CLI |
Task |
Modality |
Models |
|---|---|---|---|
|
Synthetic CT (sCT) |
MR → CT, CBCT → CT |
|
|
Multimodal body segmentation (11 labels) |
CBCT / MR / CT |
|
|
Whole-body MRI segmentation |
MRI |
folds 1–5 |
|
Whole-body CT/MRI segmentation |
CT / MRI |
|
|
Multimodal deformable registration |
MR/CT, CBCT/CT |
presets |
Each also exposes eval and (except TotalSegmentator) uncertainty; the thin
wrappers add pipeline. See Using KonfAI Apps for how to run them and
CLI reference for the full flag reference.
External clients and experimental edge#
Piece |
Status |
Notes |
|---|---|---|
SlicerKonfAI |
✅ External GUI |
A 3D Slicer client of the |
SlicerImpactReg |
✅ External GUI |
A dedicated 3D Slicer client for the complete |
ONNX export → |
🧪 Experimental |
|
Note
Trust model. Resolving a konfai-apps app copies and imports its .py
files (arbitrary code) and pip-installs its requirements.txt by default
(opt out with KONFAI_APPS_INSTALL_REQUIREMENTS=0), so only resolve apps
from sources you trust — see Python API (apps).
See also#
Using KonfAI Apps — what an app is and how to run apps from the CLI
CLI reference — the full flag reference for the app CLIs