Skip to content

oh-my-mlip

oh-my-mlip — machine learning interatomic potentials

Every major MLIP in one place, run by your LLM. oh-my-mlip gathers 20 machine-learning interatomic-potential frameworks (32 model variants) — MACE, SevenNet, NequIP, ORB, UMA, … — behind one registry, and your coding agent (Claude Code or Codex) does the work when you ask:

  • install any of them, each in its own conda env from a pinned recipe;
  • use them in your own scripts with the exact interpreter and calculator lines;
  • benchmark them on adsorption energies with CatBench, including your own VASP calculations;
  • fine-tune them on your data through each framework's own trainer;
  • distill any of them into an NN-MTP student that runs in LAMMPS.

It drives the real upstream frameworks and never reimplements a model.

Ask your LLM

oh-my-mlip is built to be driven by your coding agent. Set it up once, then ask.

Claude Code — add the plugin:

/plugin marketplace add JinukMoon/oh-my-mlip

/plugin install oh-my-mlip@oh-my-mlip

Codex or any other coding agent — tell it:

Clone https://github.com/JinukMoon/oh-my-mlip, read its AGENTS.md and the
documentation at https://oh-my-mlip.org/, and use it for the
MLIP work I ask for.

Then ask in plain language, for example:

You say Guide
"Install MACE and SevenNet and check they run on my GPU." Install models
"Relax this POSCAR with UMA and give me the relaxed structure." Use a model
"Benchmark MACE, SevenNet and UMA on adsorption energies for CO2 reduction on Cu." Benchmark with CatBench
"Turn my VASP calculations in ./dft into a CatBench dataset." VASP results to CatBench
"Fine-tune MACE on frames.traj." Fine-tuning
"Distill MACE into an NN-MTP student for 300 K MD of this slab." Distillation

Your agent asks for what only you can decide — models, data, D3, reference coefficients, how long the MD must be stable — and shows its plan before running anything long. Every step it takes is written to a file you can rerun.

Without an agent

git clone https://github.com/JinukMoon/oh-my-mlip.git && cd oh-my-mlip
source env.sh
./install.sh MACE                                  # MACE's own conda env
python scripts/setup_verify.py MACE-MPA-0 --json   # energy + forces on your GPU
python run_examples/single_point.py MACE           # a first calculation

Each guide ends with the commands to run a workflow yourself.

How it works

  • One env per framework. Their torch and CUDA stacks conflict, so each framework lives in its own conda env; resolve() tells you which interpreter and which calculator lines belong together.
  • Recipes, not improvisation. Package sets are pinned in envs/<env>.yml, and verified builds are recorded as exact lock files in envs/locks/.
  • Every procedure is a file. Job scripts, training commands and reports are written to disk before they run, so you can rerun them without an agent.
  • Weights come from each framework's official channel; gated models use your own Hugging Face login. For CC-BY-4.0 weights whose official host is unreliable, a byte-identical, checksum-verified mirror is used only as a fallback (see model licenses).