Agent skills for Claude Code and Codex
The MoleculeDesk skill teaches a coding agent how to use the moldesk CLI. Ask in plain language, for example "predict the structure of this sequence" or "which models can my Mac run?", and the agent:
- installs MoleculeDesk if it's missing (after asking you);
- checks your hardware with
moldesk doctorandmoldesk listbefore choosing a model; - prepares input files from your data or the shipped examples, and runs the model;
- summarizes outputs, confidence scores, and where the 3D report is;
- follows workflows such as design-then-fold (ProteinMPNN → Boltz-2) and ligand docking.
The skill is written to keep you in control. The agent asks before installing models (some are over 15 GB), before sending sequences to a public MSA server, and before anything that costs money. It never asks for or handles cloud API keys.
Claude Code
$ claude plugin install moldesk --marketplace pooriyapfn/moldeskOn Claude Code older than 2.1.275, use two steps instead:
claude plugin marketplace add pooriyapfn/moldesk
claude plugin install moldesk@moldeskStart a new session. Claude uses the skill automatically when a task matches; you can also call it directly with /moldesk:moldesk. Inside a session, /plugin install moldesk --marketplace pooriyapfn/moldesk does the same.
Update later with claude plugin marketplace update moldesk, then claude plugin update moldesk@moldesk.
Codex
$ codex plugin marketplace add pooriyapfn/moldeskThen open Codex, enter /plugins, choose the MoleculeDesk marketplace, and install moldesk. The skill is available in the next session. Mention it explicitly with $moldesk, or just describe the task.
Without a plugin manager
The skill is a plain folder with a SKILL.md file, so you can copy it directly. For Codex:
$ mkdir -p ~/.agents/skills && curl -fsSL https://codeload.github.com/pooriyapfn/moldesk/tar.gz/main | tar -xz -C ~/.agents/skills --strip-components=4 moldesk-main/plugins/moldesk/skills/moldeskFor Claude Code, use the same command with ~/.claude/skills instead of ~/.agents/skills. Run it again to update.
The source is in the repository under plugins/moldesk/skills/moldesk.
What the skill covers
| Topic | What the agent learns |
|---|---|
| Installation | The one-command installer, MOLDESK_HOME for GPU servers, updating |
| CLI | Every command and flag, --json output, exit codes |
| Compatibility | Reading moldesk doctor and moldesk list, which models run on Apple Silicon and which need Linux + NVIDIA |
| Models | Input formats, parameters, and example inputs for Boltz-2, OpenDDE, DiffDock-L, BindCraft2, ProteinMPNN, and LigandMPNN |
| Workflows | Smoke test, structure prediction, complex and affinity, docking, design-then-fold, binder campaigns, viewing structures |
| Cloud | Runpod quotes, explicit cost approval, and cleanup. Cloud execution is still disabled, and the skill says so |
Safety rules in the skill
- Ask before installing MoleculeDesk or a model, and show download and disk sizes.
- Ask before sending sequences off the machine (MSA server options, cloud uploads).
- For cloud runs: quote with
--dry-runfirst, show the cost and the shutdown warning, and only continue after an explicit yes. Never pass the risk acknowledgment flags on its own. - Never ask for, type, or store API keys. You run
moldesk cloud connect runpodyourself. - Report predictions as predictions, with confidence scores and the need for experimental validation.
- Respect model licenses, such as BindCraft2's hosting restriction.