Installation

NEExT requires Python 3.9–3.12 and is published on PyPI as NEExT. The core install is intentionally lightweight — heavier capabilities (PyTorch GNNs, the Workbench, visualization) live behind optional extras.

Core install

shell

pip install NEExT

This pulls in everything needed for the standard pipeline: NumPy, pandas, NetworkX, python-igraph, scikit-learn, XGBoost, imbalanced-learn, the vectorizers library (for Wasserstein embeddings), SciPy, and pydantic.

Optional extras

Install extras with the NEExT[extra] syntax. You can combine several at once, e.g. pip install "NEExT[gnn,advanced]".

Extra	Adds	Use it for
`gnn`	PyTorch (`torch>=2.0`)	GNN embeddings (GCN / GraphSAGE / GIN)
`advanced`	UMAP, Optuna, Plotly, Matplotlib, Seaborn	dimensionality reduction, tuning, plots
`workbench`	FastAPI, Uvicorn, HTTPX, PyArrow, UMAP, torch	the local Workbench UI
`workbench-mcp`	everything in `workbench` plus the `mcp` SDK	Workbench + the MCP endpoint for agents
`experiments`	Jupyter, Notebook, ipykernel	running notebooks
`dev`	pytest, ruff, black, isort, mypy, pre-commit	contributing to NEExT
`docs`	Sphinx + extensions	building the Sphinx API docs
`all`	`advanced`, `dev`, `docs`, `experiments`, `workbench`, `gnn`	everything

Note: the gnn extra is pure PyTorch — there is no DGL or PyTorch Geometric dependency, so it installs reliably across platforms. The mcp SDK in workbench-mcp requires Python 3.10+.

GNN embeddings

shell

pip install "NEExT[gnn]"

The Workbench (local UI + MCP)

shell

# Workbench UI only
pip install "NEExT[workbench]"

# Workbench UI + MCP endpoint for agents
pip install "NEExT[workbench-mcp]"

Once installed, launch it with the neext-workbench command. See Workbench.

Verify the install

verify.py

import NEExT
from NEExT import NEExT as NEExTFramework

print(NEExT.__version__)   # e.g. 0.3.1
nxt = NEExTFramework()      # logs "NEExT framework initialized"

Next: walk the Quickstart for a complete graph-classification run.