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 still in-progress or external, so you know what you can rely on today.

The map#

                        ┌─────────────────────────────┐
   YAML configs  ─────► │   konfai (core, PyPI)        │  TRAIN / PREDICTION / EVALUATION
                        │   train · predict · evaluate │
                        └──────────────┬──────────────┘
                                       │ public API
                        ┌──────────────▼──────────────┐
                        │   konfai-apps (PyPI)         │  package a workflow as an "app"
                        │   CLI · HTTP server · Python │
                        └───┬───────────┬───────────┬──┘
                            │           │           │
             apps/* wrappers│   HTTP/SSE│    stdio  │
                            ▼           ▼           ▼
                  impact_synth,   SlicerKonfAI   konfai-mcp
                  impact_seg,     (3D Slicer GUI) (agent server)
                  totalseg, …       [external]     [private]

What ships#

Piece

Status

What it is

konfai

✅ 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.

konfai-apps

✅ Shipped (PyPI, own CI)

Packages a mature workflow as a reusable app — see Using KonfAI Apps.

App bundles (apps/)

✅ Shipped (thin wrappers)

Ready-to-use CLI shims: impact-synth-konfai, impact-seg-konfai, mrsegmentator-konfai, totalsegmentator-konfai. Config + weights live on Hugging Face and download on first run.

impact-reg-konfai

🟡 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

VBoussot/konfai-demo (demo dataset), plus per-app model repos (ImpactSynth, ImpactSeg, TotalSegmentator-KonfAI, MRSegmentator-KonfAI, ImpactReg) and impact-torchscript-models (the SAM2.1 backbone behind IMPACTSynth).

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

impact-synth-konfai synthesize

Synthetic CT (sCT)

MR → CT, CBCT → CT

MR, CBCT

impact-seg-konfai segment

Multimodal body segmentation (11 labels)

CBCT / MR / CT

body

mrsegmentator-konfai segment

Whole-body MRI segmentation

MRI

folds 1–5

totalsegmentator-konfai segment

Whole-body CT/MRI segmentation

CT / MRI

total, total_mr, 3 mm variants

impact-reg-konfai register

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.

What is in-progress or external#

Piece

Status

Notes

konfai-mcp

🧪 Working, not yet published

A fastmcp server that exposes KonfAI to LLM agents: dataset inspection, config authoring/validation, launching training/prediction/evaluation runs, and reading live metrics. Working and tested, but not yet published.

SlicerKonfAI

🧪 External GUI

A 3D Slicer client of the konfai-apps CLI/server. Real and useful, but coupled to konfai-apps through a brittle byte-level contract (progress format, InferenceStack.mha, app.json, checkpoint schema) that can break on a release without a contract test.

ONNX export → konfai-rs

🧪 Experimental

konfai/export.py produces ONNX + a manifest for a planned portable (native/WASM) inference engine. Python-API-only, single static-shape head, feed-forward models only. See Python API (apps).

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#