Use with Claude Code

The Claude Code plugin lets you run oh-my-mlip by asking in plain language. It works from any directory.

Add the plugin

/plugin marketplace add JinukMoon/oh-my-mlip

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

The repository is its own marketplace (.claude-plugin/marketplace.json), so no hosted registry is involved. To use a local clone instead, add its directory as the marketplace:

/plugin marketplace add ~/oh-my-mlip

Where it installs

The plugin carries only the instructions. Envs, weights and the machine-local records live in a clone of this repository, the hub:

• $OH_MY_MLIP_HOME if you set it, otherwise ~/.oh-my-mlip, which Claude clones on first use;
• about 5 to 15 GB per installed model.

To use a clone you already have, set OH_MY_MLIP_HOME to it before starting Claude Code. The copy of the repository inside the plugin directory is never used as the hub: it is replaced on every plugin update, so the install scripts refuse it.

Keep it up to date

An installed plugin stays on its version until you update it:

/plugin marketplace update oh-my-mlip
/plugin update oh-my-mlip@oh-my-mlip

Then restart Claude Code. The hub is a normal git clone: update it with git -C ~/.oh-my-mlip pull (see Help).

What it adds

You rarely need to type these; Claude picks the right one from your request.

Command	What it does
/oh-my-mlip:setup <model>	install a model's env, compile it for your GPU if needed, and verify energy and forces on the GPU
/oh-my-mlip:run <model>	single-point energy and forces, or a structure relaxation
/oh-my-mlip:catbench	CatBench adsorption benchmark across models, including your own VASP data
/oh-my-mlip:finetune	fine-tune a model on your data with its framework's own trainer
/oh-my-mlip:distill	distill a model into an NN-MTP student for LAMMPS, with the onthefly-distill engine

The skills are thin: model facts live in models.json and the procedures in AGENTS.md and recipes/, so the plugin and a Codex session follow the same steps.

Check that it works

Ask:

Install MACE-MPA-0 and check it runs on my GPU.

It is done when Claude reports energy and forces computed on your GPU.

Notes

• Gated models (UMA, eSEN): Claude shows the license page to accept and asks you to log in to Hugging Face yourself; it never asks you to paste a token. See Gated models.
• An optional MCP server (python -m oh_my_mlip.mcp_server, from an interpreter with requirements-mcp.txt installed) offers the registry and the run, install-check and verify steps as typed tools; the plugin does not need it.