HighFold3 MCP Server

AI-powered cyclic peptide structure prediction via Docker

An MCP (Model Context Protocol) server for HighFold3 structure prediction with Cyclization Switch and CycPOEM technology. Core capabilities:

• Predict structures for linear, cyclic, and disulfide-bonded peptides
• Predict peptide-protein complexes (built-in protein targets or custom sequences)
• Parse HELM notation for complex peptide topologies
• Batch process peptide libraries for virtual screening
• Monitor long-running prediction jobs

Quick Start with Docker

Approach 1: Pull Pre-built Image from GitHub

The fastest way to get started. A pre-built Docker image is automatically published to GitHub Container Registry on every release.

# Pull the latest image
docker pull ghcr.io/macromnex/highfold3_mcp:latest

# Register with Claude Code (runs as current user to avoid permission issues)
claude mcp add highfold3 -- docker run -i --rm --user `id -u`:`id -g` --gpus all --ipc=host -v `pwd`:`pwd` ghcr.io/macromnex/highfold3_mcp:latest

Note: Run from your project directory. `pwd` expands to the current working directory.

Requirements:

• Docker with GPU support (nvidia-docker or Docker with NVIDIA runtime)
• Claude Code installed

That's it! The HighFold3 MCP server is now available in Claude Code.

---

Approach 2: Build Docker Image Locally

Build the image yourself and install it into Claude Code. Useful for customization or offline environments.

# Clone the repository
git clone https://github.com/MacromNex/highfold3_mcp.git
cd highfold3_mcp

# Build the Docker image
docker build -t highfold3_mcp:latest .

# Register with Claude Code (runs as current user to avoid permission issues)
claude mcp add highfold3 -- docker run -i --rm --user `id -u`:`id -g` --gpus all --ipc=host -v `pwd`:`pwd` highfold3_mcp:latest

Note: Run from your project directory. `pwd` expands to the current working directory.

Requirements:

• Docker with GPU support
• Claude Code installed
• Git (to clone the repository)

About the Docker Flags:

• -i — Interactive mode for Claude Code
• --rm — Automatically remove container after exit
• --user `id -u`:`id -g` — Runs the container as your current user, so output files are owned by you (not root)
• --gpus all — Grants access to all available GPUs
• --ipc=host — Uses host IPC namespace for better performance
• -v — Mounts your project directory so the container can access your data

---

Verify Installation

After adding the MCP server, you can verify it's working:

# List registered MCP servers
claude mcp list

# You should see 'highfold3' in the output

In Claude Code, you can now use the HighFold3 tools:

Structure Prediction:

• submit_linear_peptide_prediction
• submit_cyclic_peptide_prediction
• submit_disulfide_cyclic_peptide_prediction
• submit_cyclic_peptide_protein_complex_prediction
• submit_helm_peptide_protein_complex_prediction
• submit_batch_peptide_prediction

Job Management:

• get_job_status / get_job_result / get_job_log
• list_jobs / cancel_job / cleanup_old_jobs

Utilities:

• validate_peptide_sequence
• validate_helm_notation
• get_server_info

---

Next Steps

• Detailed documentation: See detail.md for comprehensive guides on:
  • Available MCP tools and parameters
  • Local Python environment setup (alternative to Docker)
  • Script-based usage without MCP
  • Example workflows and use cases
  • Configuration file formats

---

Usage Examples

Once registered, you can use the HighFold3 tools directly in Claude Code. Here are some common workflows:

Example 1: Cyclic Peptide Structure Prediction

Submit a cyclic peptide structure prediction for sequence "CFWKYKYK" with job name "RGD_analog".
Monitor the job and show me the results when complete.

Example 2: Disulfide-Bonded Cyclic Peptide

Predict the structure of disulfide cyclic peptide CWKDGKYKWC with a disulfide bond between positions 1 and 10.

Example 3: Peptide-Protein Complex

Predict the structure of cyclic peptide CFWKYKYK bound to a PDZ domain protein.

Example 4: HELM Notation Pipeline

I have a cyclic peptide in HELM notation: PEPTIDE1{C.F.W.K.Y.K.Y.K}$PEPTIDE1,PEPTIDE1,1:R1-8:R1$$$V2.0
Predict its structure in complex with protein sequence MKLIVQPGET...

Example 5: Batch Virtual Screening

Submit batch predictions for these cyclic peptides:
- CFWKYKYK (RGD analog)
- CWKDGKYKWC (disulfide variant)
- RGDRGD (integrin binder)

---

Troubleshooting

Docker not found?

docker --version  # Install Docker if missing

GPU not accessible?

• Ensure NVIDIA Docker runtime is installed
• Check with docker run --gpus all ubuntu nvidia-smi

Claude Code not found?

# Install Claude Code
npm install -g @anthropic-ai/claude-code

---

License

CC-BY-NC-SA 4.0 (Google DeepMind)

Credits

Based on HighFold3 — Advanced AI model for cyclic peptide structure prediction with Cyclization Switch and CycPOEM technology.