MLIP Agent Skills
Japanese version / 日本語版
A Claude Code plugin marketplace for open-source machine-learning interatomic potentials (MLIPs).
This repository focuses on practical ASE-based workflows for four open-source MLIP families:
Based On Matlantis Agent Skills
This repository was created based on Matlantis Agent Skills. The original Matlantis-oriented content is kept in this repository as reference material only.
By default, agents should use the open-source MLIP skills in the mlip plugin. Matlantis/PFP-specific skills are legacy reference material and should be used only when the user explicitly asks about Matlantis, PFP, pfp_api_client, or matlantis_features.
Overview
MLIP Agent Skills helps Claude Code generate Python code for atomistic simulations using ASE-compatible MLIP calculators. The current scope is calculator setup, model/task selection, structure relaxation, molecular dynamics, NEB, phonons, and common analysis workflows.
The skills emphasize portable ASE workflows. When a workflow can be expressed with standard ASE, prefer that form so the same code can be reused across CHGNet, SevenNet, MatterSim, and UMA with only the calculator factory changed.
Skills Catalog
| Category | Skill | Summary |
|---|
| Foundation | mlip-overview | MLIP family selection, domain coverage, authentication, and key tradeoffs |
| Foundation | mlip-setup | Calculator initialization for CHGNet, SevenNet-Omni, MatterSim, and UMA-s-1p2 |
| Workflow | mlip-workflows | ASE relaxation, MD, NEB, phonon, and screening patterns shared across MLIPs |
| Reference | matlantis-reference | Matlantis/PFP legacy notes; reference only, not the default recommendation |
Getting Started
Prerequisites
- Claude Code CLI with plugin support
- Python environment suitable for the selected MLIP
- ASE for the common simulation interface
Installation
Run the following commands in the Claude Code REPL:
# Register the marketplace
/plugin marketplace add https://github.com/wakamiya0315/mlip-agent-skills
Then enable the mlip plugin from the plugin manager:
/plugin
OpenCode Installation
OpenCode does not read the Claude Code marketplace layout directly. To use these skills in OpenCode, copy the SKILL.md directories into an OpenCode skill search path. See the OpenCode Agent Skills documentation for the current search-path rules.
Global install, recommended for your own machine:
scripts/install_opencode_skills.sh
Project-local install:
scripts/install_opencode_skills.sh --project
By default this installs only the OSS MLIP skills:
mlip-overview
mlip-setup
mlip-workflows
The Matlantis/PFP reference skill is intentionally not installed by default. Install it only if you want legacy reference access:
scripts/install_opencode_skills.sh --include-matlantis-reference
OpenCode searches these locations among others:
~/.config/opencode/skills/<name>/SKILL.md
.opencode/skills/<name>/SKILL.md
Optional OpenCode permission example:
{
"permission": {
"skill": {
"mlip-*": "allow",
"matlantis-reference": "ask"
}
}
}
MLIP Selection Guide
| MLIP | Default Use | Notes |
|---|
| CHGNet | Inorganic crystals, Materials Project-like bulk systems, fast pre-relaxation | Good first choice for crystalline inorganic structures and magnetic-moment-informed workflows |
| SevenNet-Omni | Broad PBE-level materials workflows | Use model="7net-omni" and choose modal by target fidelity/domain |
| MatterSim | Broad materials simulations across elements, temperatures, and pressures | Default checkpoint is the 1M model; use the 5M checkpoint when accuracy matters more than speed |
| UMA-s-1p2 | Multi-domain materials, catalysis, molecules, MOFs, molecular crystals | Requires Hugging Face model access; choose task_name by application domain |
Calculator Setup Examples
CHGNet
from chgnet.model.dynamics import CHGNetCalculator
calc = CHGNetCalculator(use_device="cuda") # use "cpu" or "mps" when needed
atoms.calc = calc
energy = atoms.get_potential_energy()
CHGNet also provides StructOptimizer for structure relaxation and MolecularDynamics for MD. Prefer these CHGNet-native helpers when the user specifically wants CHGNet's built-in relaxation or MD utilities.
SevenNet-Omni
from sevenn.calculator import SevenNetCalculator
calc = SevenNetCalculator(model="7net-omni", modal="mpa", device="auto")
atoms.calc = calc
