English | 简体中文

---

Agent skills for working with AlphaFold 3 — Google DeepMind's biomolecular structure prediction pipeline — from building input JSON files to running and interpreting results.

Built on top of the af3cli library for input generation and the official AlphaFold 3 inference pipeline.

Compatible with all tools supporting the Agent Skills standard: opencode, Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, and 50+ others.

💡 Also useful as a learning reference for understanding AlphaFold 3 — from model architecture and data pipeline internals to confidence metrics and troubleshooting — even without an AI agent.

---

Skills

	🧬 af3cli	🔬 alphafold3	🖥️ remote-server
What it does	Generate input JSON files	Run inference + interpret results	Manage remote servers via SSH
Input	FASTA · SMILES · SDF · CCD	Input JSON	Server address + credentials
Output	.json file	.cif structure + confidence scores	File transfers, background jobs, results
Dependencies	pip install af3cli + RDKit/Biopython	Docker + NVIDIA GPU	SSH client (built-in on macOS/Linux)
Key features	Chainable CLI · Python API · MSA · Templates · Bonds · Modifications	Confidence metrics · Performance tuning · Troubleshooting · Model internals · Codebase navigation	SSH config · scp/rsync · screen/tmux/nohup · Progress monitoring

🧬 af3cli

CLI and Python library for generating AlphaFold 3 input JSON files.

Based on SLx64/af3cli

• Sequences — Add protein / DNA / RNA from FASTA or inline, auto-detect reverse complement
• Ligands — Add from SMILES, CCD codes, or SDF files (RDKit)
• Modifications — Residue and nucleotide modifications at specified positions
• Templates & MSA — Structural templates (mmCIF) and paired/unpaired MSA data
• Bonds — Define bonded atom pairs between chains and ligands
• Chainable pipeline — Compose commands with - delimiter: af3cli config ... - protein add ... - ligand add ... - debug --show

🔬 alphafold3

Run and interpret AlphaFold 3 inference — from Docker commands to understanding output metrics.

Based on google-deepmind/alphafold3

• Running — Docker/Singularity commands, staged pipeline (data-only / inference-only), batch processing
• Input format — Complete JSON reference for sequences, ligands, bonds, modifications, MSA, templates
• Confidence metrics — pLDDT (per-atom), PAE (per-pair), pTM (overall fold), ipTM (interface), ranking scores
• Performance tuning — Compilation buckets, sharded sequence alignment databases (10-30x speedup), JAX persistent cache, unified memory
• Troubleshooting — V100 issues, SMILES two-letter atoms, MSA discrepancies, RDKit conformer failures
• Internals — Evoformer trunk, Diffusion Head, Confidence Head, data pipeline architecture, full codebase map

🖥️ remote-server

General-purpose remote Linux server management via SSH — the bridge between your local agent and remote GPU/compute servers.

• SSH connection — Key authentication, ~/.ssh/config aliases, jump hosts, connection testing
• File transfer — scp, rsync, direct file writing via heredoc
• Background jobs — screen, tmux, nohup for long-running tasks
• Progress monitoring — Check process status, tail logs, detect output file completion
• Result retrieval — Download output files, clean up remote temp files

---

Install

skills CLI (recommended)

npx skills add ByteTora/alphafold3-skills       # npm
bunx skills add ByteTora/alphafold3-skills      # bun
pnpm dlx skills add ByteTora/alphafold3-skills  # pnpm

The skills CLI automatically detects installed agent tools and installs to the correct directories for 55+ agents.

gh CLI

gh skill install ByteTora/alphafold3-skills

Requires GitHub CLI v2.90.0+.

Manual

git clone https://github.com/ByteTora/alphafold3-skills.git

# Copy to your agent's skills directory:
cp -r alphafold3-skills/skills/* ~/.opencode/skills/   # opencode
cp -r alphafold3-skills/skills/* ~/.claude/skills/     # Claude Code
cp -r alphafold3-skills/skills/* ~/.codex/skills/       # Codex
cp -r alphafold3-skills/skills/* ~/.cursor/skills/     # Cursor

---

Usage Examples

After installing, the skills trigger automatically when you mention relevant tasks.

Build an input file

"Create an AlphaFold3 input for myoglobin with a heme ligand"

→ The agent uses af3cli to generate the JSON with protein sequence, SMILES ligand, and bonds.

Run prediction

"Run AlphaFold3 on this input file on my GPU server at 192.168.1.100"

→ The agent uses remote-server to SSH into the server, upload the file, start a background job with screen/tmux, and report the estimated completion time.

Interpret results

"What does an ipTM of 0.85 mean for my predicted complex?"

→ The agent uses alphafold3 to explain: ipTM > 0.8 indicates high-confidence interface prediction.

Debug failures

"AlphaFold3 produced a clash score of -99, how do I fix this?"

→ The agent uses alphafold3 to identify the V100/XLA incompatibility and suggest XLA_FLAGS workaround.

---

Remote Servers

Use the remote-server skill to run AlphaFold 3 on remote GPU servers. Include the server address in your prompt:

"Fold this protein on my lab server (user@10.0.0.5)"

The agent will handle SSH connection, file upload, and background job management automatically.

---

Dependencies

The skills themselves are documentation with no runtime dependencies. However, the tools they guide you to use require:

Before using: AlphaFold 3 must be fully installed on the target machine (cloned, Docker image built, sequence alignment databases downloaded, model parameters obtained). This skill collection does not install AlphaFold 3 — it guides you through using an existing installation.

Dependency	Required by	Details
pip install af3cli[biopython,rdkit]	af3cli	Python 3.10+
AlphaFold 3 Docker image	alphafold3	docker build -t alphafold3 -f docker/Dockerfile .
NVIDIA GPU (A100/H100)	alphafold3	Compute Capability >= 8.0
Sequence alignment databases	alphafold3	~252 GB download, ~630 GB decompressed
Model parameters	alphafold3	Via Google Form

---

Contributing

Contributions are welcome! If you'd like to improve a skill or add new AlphaFold 3 related skills:

1. Fork the repo
2. Add or modify skills under skills/
3. Ensure each skill has a valid SKILL.md with YAML frontmatter (name, description)
4. Open a PR

---

Star History

---

References

• SLx64/af3cli — AlphaFold 3 input JSON generation tool
• google-deepmind/alphafold3 — Official AlphaFold 3 inference pipeline

---