[Plugin Submission] protein-design-skills — agent-agnostic end-to-end protein design

Plugin Submission — Updated

Repo renamed from kimi-protein-design to protein-design-skills. Major architecture overhaul since initial submission.

Repository

https://github.com/devxia/protein-design-skills

One-line Description

Agent-agnostic protein design plugin — 15+ pipelines from backbone generation to structure validation, entirely through natural language conversation.

What It Does

protein-design-skills is an agent-agnostic protein design plugin that orchestrates 15+ open-source ML tools through Skills + Hooks + Scripts (no MCP server needed). Works with Kimi Code, Claude Code, Codex CLI, and any coding agent that reads skills.

5-stage pipeline:

Stage	Primary Tool	Alternatives	Purpose
Stage 0	PDBFixer	—	Mandatory PDB repair
Stage 1	RFdiffusion	Chroma, FoldFlow, DiffPepBuilder, RFpeptides, Genie 3	Backbone generation
Stage 2	ProteinMPNN	LigandMPNN, ESM-IF1, PiFold	Sequence design
Stage 3	AlphaFold3	Boltz-1, Chai-1, OmegaFold, ESMFold, Protenix, OpenFold3	Structure validation
Stage 4	Filtering	Cross-validation, Score-first screening	Quality ranking

15+ Design Pipelines

Pipeline	Stage 1	Stage 2	Stage 3	Best For
Standard	RFdiffusion	ProteinMPNN	AlphaFold3	General purpose
Fast Screening	RFdiffusion	ProteinMPNN	ESMFold/OmegaFold	No databases needed
Ligand-Aware	RFdiffusionAA	LigandMPNN	AlphaFold3	Small molecules, cofactors
Peptide	DiffPepBuilder	Built-in	AlphaFold3	8-30aa peptides
Cross-Validation	RFdiffusion	ProteinMPNN	Boltz-1 + Chai-1 + OmegaFold	Most robust ranking
Chroma	Chroma (joint)	—	AlphaFold3	All-atom, natural language
Antibody	IgDiff/RFdiffusion	AbMPNN/ProteinMPNN	AlphaFold3	Antibodies, nanobodies
Enzyme	RFdiffusionAA	LigandMPNN	AlphaFold3	Active sites, catalysis
... and 7 more

Architecture: Skills + Hooks + Scripts

Layer	What	Count	Location
Skills	Markdown knowledge for the LLM	76	skills/
Hooks	Automation scripts (auto-trigger)	22	protein_design/hooks/
Scripts	Standalone execution	19	scripts/

How it works: Skills teach the agent → Hooks fire automatically → Scripts run tools directly.

What Hooks Do

Hook	Trigger	What It Does
user-onboarding	First protein prompt	Welcome message + tool status + quick start guide
session-health-check	Protein prompts	Checks installed tools, suggests alternatives
tool-recommender	Design requests	Recommends scripts and parameters
error-recovery	Tool failures	Suggests fixes, alternative tools, install commands
progress-reporter	Long jobs	ETA estimation, file counting, progress updates
pipeline-orchestrator	Stage completion	Auto-detects next step, suggests what to run
quality-gate	Validation results	Pass/fail decisions with thresholds
design-report	Filtering complete	Auto-generates summary with rankings
gpu-check-hook	Before GPU jobs	Checks VRAM, warns if insufficient

Installation (Kimi Code)

/plugins install https://github.com/devxia/protein-design-skills
/new

Example Usage

User: Design a binder targeting PD-L1
→ Stage 0: PDBFixer preprocesses target.pdb
→ Stage 1: RFdiffusion generates 10 binder backbones
→ Stage 2: ProteinMPNN designs 8 sequences per backbone
→ Stage 3: AlphaFold3 validates top designs
→ Stage 4: Filter by ipTM > 0.8 and pLDDT > 80

Why It Should Be Curated

1. Agent-agnostic — works with Kimi Code, Claude Code, Codex CLI, and any agent that reads skills. Not a single-agent plugin but a universal protein design framework
2. 15+ pipelines — from fast screening (ESMFold, no GPU) to full cross-validation (Boltz-1 + Chai-1 + OmegaFold). Covers monomers, binders, peptides, antibodies, enzymes, and more
3. 76 workflow skills — comprehensive coverage of every pipeline stage, tool alternative, design pattern, and troubleshooting scenario
4. 22 auto-triggering hooks — context injection, GPU safety checks, error recovery, pipeline orchestration, desktop notifications. No manual setup
5. Zero vendor lock-in — all underlying tools are open-source and run locally. Missing-tool errors return structured messages with install guides
6. Production-ready — conda environment isolation, cross-tool execution, batch pipeline runner, job management, progress monitoring with ETA
7. Fills a gap — the only agent-agnostic computational biology / structural bioinformatics plugin in the coding agent ecosystem

What Changed Since Initial Submission

• Architecture: MCP Server → Skills + Hooks + Scripts (agent-agnostic)
• Repo: kimi-protein-design → protein-design-skills
• Skills: 7 → 76 (covering 30+ design pipelines)
• Hooks: recommended → 22 auto-triggering automation scripts
• Scripts: — → 19 standalone Python scripts
• Tools: 4 → 15+ (added Boltz-1, Chai-1, ESMFold, OmegaFold, Chroma, FoldFlow, LigandMPNN, Protenix, OpenFold3, etc.)
• Pipelines: 1 → 15+ design pipelines
• Agents: Kimi Code only → Claude Code + Codex CLI + Kimi Code + any agent

Checklist

• kimi.plugin.json manifest is valid
• Bilingual README (English + Chinese)
• 76 workflow skills covering all pipeline stages
• 22 auto-triggering hooks
• 19 standalone scripts for direct execution
• Bilingual documentation under docs/ (en + zh)
• Installation guide, quick start, pipeline architecture, troubleshooting
• Cross-conda environment support
• Agent-agnostic design (Claude Code, Codex CLI, Kimi Code)
• MIT licensed

Start by reviewing the linked protein-design-skills repository, especially kimi.plugin.json, README, skills/, protein_design/hooks/, scripts/, and docs/. Verify the manifest and stated installation flow, then assess whether the external plugin meets the repository's curation requirements; done means reaching a clear accept-or-reject decision.