foundry-skills

Foundry Claude Skills

Claude Code skills for running Foundry protein design tools on Vast.ai GPU instances.

Uses a pre-built Docker image (see Dockerfile in the Foundry repo) with all tools, checkpoints, and dependencies pre-installed — including Enhanced MPNN weights. Instances are ready to use immediately after launch.

Skills

Skill	Command	Description
/rfd3	rfd3 design	Run RFdiffusion3 all-atom protein structure design
/rf3	rf3 fold	Run RosettaFold3 structure prediction
/mpnn	mpnn	Run ProteinMPNN/LigandMPNN/SolubleMPNN/Enhanced MPNN sequence design
/run-foundry-job	Multi-step	Run complete design pipelines (e.g., RFD3 + Enhanced MPNN + RF3)

Each skill launches a Vast.ai GPU instance with the Foundry Docker image, uploads input files, runs the job, downloads results, and destroys the instance.

Prerequisites

• Claude Code installed (docs)
• Vast.ai CLI installed and authenticated:

  pip install vastai
  vastai set api-key <YOUR_API_KEY>
  

• SSH keys configured on Vast.ai:

  vastai show ssh-keys
  # If empty, upload your public key:
  vastai create ssh-key "$(cat ~/.ssh/id_ed25519.pub)"
  

• Foundry Docker image built and pushed to a registry (see Dockerfile in the Foundry repo)

Installation

Option 1: Install from GitHub (recommended)

Step 1 — Add the marketplace.

From within Claude Code, run:

/plugin marketplace add LiorZ/foundry-claude-skill

Step 2 — Install the plugin.

/plugin install foundry@foundry-skills

This permanently registers the skills. They will be available in all future Claude Code sessions.

Step 3 — (Optional) Set project scope for team sharing.

To make the plugin available for everyone working in a specific project:

/plugin install foundry@foundry-skills --scope project

This writes the configuration to .claude/settings.json which can be committed to your repo.

Option 2: Install from a local path (for development)

Launch Claude Code with the --plugin-dir flag pointing to a local clone:

git clone [email protected]:LiorZ/foundry-claude-skill.git
claude --plugin-dir ./foundry-claude-skill

This loads all skills for that session. Useful for testing and development.

Alternatively, add a local clone as a marketplace:

/plugin marketplace add /path/to/foundry-claude-skill
/plugin install foundry@foundry-skills

Option 3: Manual configuration

Add the following to your Claude Code settings file (~/.claude/settings.json for user-wide, or .claude/settings.json in your project):

{
  "extraKnownMarketplaces": {
    "foundry-skills": {
      "source": {
        "source": "directory",
        "path": "/path/to/foundry-claude-skill"
      }
    }
  },
  "enabledPlugins": {
    "foundry@foundry-skills": true
  }
}

Verifying installation

After installation, you should be able to invoke skills directly:

/rfd3 Design binders for my target protein

If the skills are loaded correctly, Claude will launch a forked agent that handles the full Vast.ai workflow.

Usage Examples

# Design protein binders targeting PD-L1
/rfd3 Design binders for PD-L1 using target.pdb with hotspots at A56, A115, A123

# Predict structure of a designed sequence
/rf3 Fold this sequence: MTSENPLLALREK...

# Design sequences for a backbone with Enhanced MPNN (best designability)
/mpnn Design 8 sequences for backbone.cif using Enhanced MPNN at temperature 0.1

# Design sequences with standard ProteinMPNN
/mpnn Design 10 sequences for backbone.cif using ProteinMPNN at temperature 0.1

# Run a full binder design pipeline
/run-foundry-job Design binders for target.pdb, then design sequences with Enhanced MPNN, then validate with RF3

Skills Reference

/rfd3 — RFdiffusion3 Design

Designs protein structures under complex constraints. Supports protein binders, enzyme active sites, nucleic acid binders, small molecule binders, and symmetric assemblies.

/rf3 — RosettaFold3 Prediction

Predicts biomolecular structures for proteins, nucleic acids, small molecules, and their complexes. Supports templating, covalent modifications, and MSAs.

/mpnn — Sequence Design

Fixed-backbone inverse folding with four model variants:

• ProteinMPNN — standard sequence design
• LigandMPNN — ligand-aware design (small molecules, DNA/RNA, ions)
• SolubleMPNN — solubility-optimized design for E. coli expression
• Enhanced MPNN — designability-optimized (recommended for de novo design). Fine-tuned LigandMPNN using ResiDPO that achieves ~2.7x better enzyme design success and ~2.3x better binder design success by optimizing directly for structural validation confidence rather than native sequence recovery.