Skip to content

Get started

Before you start: Linux with an NVIDIA GPU, and conda (or mamba). Gated models (UMA, eSEN) also need a Hugging Face account; your agent asks when it gets there. Details: Host requirements.

1. Connect your agent

Add the plugin once:

/plugin marketplace add JinukMoon/oh-my-mlip

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

The plugin works from any directory. On first use Claude clones the hub into ~/.oh-my-mlip (or uses $OH_MY_MLIP_HOME if you set it) and builds envs there, about 5 to 15 GB per model. Details: Use with Claude Code.

Tell your agent:

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.

git clone https://github.com/JinukMoon/oh-my-mlip.git && cd oh-my-mlip
source env.sh

2. Install your first model

Ask your agent:

Install MACE and check it runs on my GPU.

or run it yourself:

./install.sh MACE
python scripts/setup_verify.py MACE-MPA-0 --json

The model counts as installed only when it has computed energy and forces on your GPU.

Two kinds of model need one more thing, and your agent handles both:

  • UMA and eSEN are gated. Accept the license on the model's Hugging Face page, then run hf auth login in your own terminal (Gated models).
  • NequIP and Allegro are compiled for your GPU during install; to use them on a different GPU, compile them there (GPU-architecture compilation).

3. Do something with it

Relax POSCAR with MACE-MPA-0 and give me the relaxed structure.

From here: