jeonghyeonkim8652/protein-design-mcp repository overview

⁠Protein Design MCP Server

An MCP⁠ server that gives LLM agents access to computational protein design tools. Ask your LLM to design binders, generate de novo folds, predict structures, score interfaces, or relax with Rosetta — it calls the right tool automatically.

19 tools total. This image bundles 13 of 19; the 6 license/conflict-gated tools (rosetta_*, predict_*_boltz) install separately via pip — see below.

Source: github.com/jasonkim8652/protein-design-mcp⁠

⁠Quick start

⁠GPU

docker volume create protein-design-models

docker run --rm -i \
  --gpus all \
  -v protein-design-models:/models \
  -v $(pwd):/data \
  jeonghyeonkim8652/protein-design-mcp:latest

Requires NVIDIA Container Toolkit⁠. Model weights (~10 GB) download lazily on first use and persist in the named volume.

⁠CPU

docker run --rm -i \
  -e DEVICE=cpu \
  -v protein-design-models:/models \
  -v $(pwd):/data \
  jeonghyeonkim8652/protein-design-mcp:latest

CPU mode disables design_binder, design_fold, generate_backbone (RFdiffusion is GPU-only) → 10 tools available.

⁠MCP client config

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "protein-design": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm", "--gpus", "all",
        "-e", "SKIP_MODEL_DOWNLOAD=true",
        "-v", "protein-design-models:/models",
        "-v", "/absolute/path/to/your/pdbs:/data",
        "jeonghyeonkim8652/protein-design-mcp:latest"
      ]
    }
  }
}

Claude Code — same structure in .mcp.json; swap the data volume for ${workspaceFolder}:/data.

⁠Tools bundled in this image (13)

Backends in the image: RFdiffusion, ProteinMPNN, ESMFold, ColabFold/AlphaFold2, OpenFold, ESM2, OpenMM.

⁠Tools NOT bundled (6)

These require extra dependencies that can't ship in a single image:

⁠PyRosetta tools — rosetta_score, rosetta_relax, rosetta_interface_score, rosetta_design

PyRosetta cannot be redistributed in a container. Install after accepting the Rosetta license⁠:

pip install "protein-design-mcp[gpu,rosetta]"
python -c "import pyrosetta_installer; pyrosetta_installer.install_pyrosetta()"

⁠Boltz-2 tools — predict_structure_boltz, predict_affinity_boltz

Boltz-2 needs torch>=2.2, which conflicts with RFdiffusion's torch==2.0.1. Install in an isolated virtualenv:

python -m venv ~/.venvs/protein-design-boltz
source ~/.venvs/protein-design-boltz/bin/activate
pip install "protein-design-mcp[boltz]"

Then run two MCP servers side-by-side — one Docker (RFdiffusion stack), one venv (Boltz). Your LLM sees all 19 tools.

⁠Tags

⁠Links

• GitHub: https://github.com/jasonkim8652/protein-design-mcp⁠
• PyPI: https://pypi.org/project/protein-design-mcp/⁠
• Release notes: https://github.com/jasonkim8652/protein-design-mcp/releases/tag/v1.0.0⁠
• Issues: https://github.com/jasonkim8652/protein-design-mcp/issues⁠

Apache License 2.0.