InternScience/MolClaw

InternScience profile

Latest recorded source activity: September 24, 2026 at 07:26

SkillsMP catalog refreshed: October 17, 2026 at 00:09

skills collected: 67

GitHub stars: 35
GitHub forks: 3

Repository README

Repository README source: README.md

Read the source README

Expand full README

<a id="top"></a> # <img src="assets/molclaw_logo.png" alt="MolClaw Logo" width="42" /> MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization <div align="center">  [![arXiv](https://img.shields.io/badge/arXiv-2604.21937-red?style=flat-square)](https://arxiv.org/abs/2604.21937) [![License](https://img.shields.io/badge/License-MIT-green?style=flat-square)](./LICENSE) [![Python 3.12+](https://img.shields.io/badge/Python-3.12%2B-blue?style=flat-square)](https://www.python.org/downloads/) **A reproducible pathway for autonomous molecular evaluation, screening, and optimization.** ✨ If MolClaw helps your research or engineering workflow, we'd really appreciate a star. </div> --- ## 📖 Overview Computational drug discovery is fundamentally a **long-horizon workflow orchestration problem**: each decision depends on upstream tool outputs, scientific constraints, and strict quality verification. MolClaw is designed for this setting as a practical and reproducible framework for **molecule evaluation, screening, and optimization**. This open-source release combines a benchmark suite, a hierarchical skill library, and runnable inference/evaluation pipelines: - **Hierarchical skill architecture (`L1`/`L2`/`L3`)** that separates atomic tool operations, workflow composition, and methodology-level scientific governance. - **Model- and runtime-agnostic design philosophy**: skill definitions are intended to remain stable as backbone models and agent runtimes evolve. - **Unified benchmark stack (`molbench`)** covering molecular screening and molecular optimization settings under standardized evaluation protocols. - **Runnable execution pipelines in `molclaw_run`** for direct baseline LLM inference and Claude Code skill-driven execution. <div align="center"> <img src="assets/molclaw_architecture.png" alt="MolClaw Architecture" width="920" /> </div> --- ## 📊 MolBench `molbench` is a **multi-dimensional benchmark suite for drug discovery agents**, designed to evaluate not only final answers but also the reliability of structured, tool-augmented scientific workflows. - **`molbench-ms-1`**: molecular property filtering. - **`molbench-ms-2`**: target-aware binding-affinity comparison QA. - **`molbench-ms-3`**: virtual-screening style ranking and hit identification. - **`molbench-mo`**: molecule optimization tasks, including editing and physicochemical optimization tracks. `molbench/eval` contains unified evaluators (`eval_runner.py`, `run_eval_bench.py`) together with ChemCoTBench-related evaluation dependencies, enabling **comparable cross-run analysis** under one consistent evaluation interface. <div align="center"> <img src="assets/molbench_leaderboard.png" alt="MolBench Leaderboard" width="920" /> </div> --- ## 🤖 Skills `skills/` is the reusable skill library for agent execution and workflow standardization: - **`skills/L1_tools`**: tool-level operation skills (atomic actions, interface conventions, output checks). - **`skills/L2_workflows`**: multi-step workflow skills (task decomposition, step ordering, failure recovery). - **`skills/L3_methodology`**: high-level scientific methodology and governance principles. - **`skills/LR_research`**: literature review and research-related skills. - **`skills/auto-generated-skills`**: skills automatically generated during agent execution. - **`skills/system_prompt`**: system prompts defining the agent's core behavior. This hierarchy is built to be **modular, composable, and auditable**, so complex workflows remain maintainable as tasks, tools, and model backbones evolve. For Claude Code usage, copy skills into your task workspace: ```bash mkdir -p ./.claude/skills cp -R <YOUR_CLONED_MOLCLAW_DIR>/skills/* ./.claude/skills/ ``` --- ## 🚀 Quick Start ### 1) Clone Repository and Create Conda Environment ```bash git clone https://github.com/InternScience/MolClaw.git cd MolClaw conda env create -f environment.yaml conda activate molclaw ``` This environment is the recommended baseline for `molbench` inference and evaluation dependencies. ### 2) Configure Environment ```bash cp .env.template .env # edit .env and fill required keys/endpoints for your model provider # load env vars into current shell set -a source .env set +a ``` ### 3) Connect Claude Code to the MolClaw MCP Server Register the deployed DrugSDA-Tool endpoint in Claude Code: ```bash claude mcp add \ molclaw-scp \ --transport http \ 'https://scp.intern-ai.org.cn/api/v1/mcp/2/DrugSDA-Tool' \ --header "SCP-HUB-API-KEY: <YOUR-API-KEY>" ``` Replace `<YOUR-API-KEY>` with the SCP Hub key configured through `.env.template`. ### 4) Claude Code Interact with Skills Copy the Skills into your Claude Code workspace and then you can interact with them in your conversations. ```bash # in your task workspace mkdir -p ./.claude/skills cp -R <YOUR_CLONED_MOLCLAW_DIR>/skills/* ./.claude/skills/ claude # then run /init and ask your task ``` ### 5) Benchmark Runs Use the commands below for baseline and Claude benchmark runs; each command performs inference and automatic `molbench` evaluation. ```bash # Baseline benchmark run (direct LLM inference + auto evaluation) bash molclaw_run/infer/baselines/launch_baseline.sh --cfg config/baseline_molbench-ms-1.yaml # Claude benchmark run (Claude Code workflow + auto evaluation) bash molclaw_run/infer/claude_agent/launch_claude.sh --cfg config/claude_template.yaml --claude-mode both ``` You can switch datasets/tasks by replacing the config file with other templates in `config/`. ### 6) Evaluate an Existing Run Directory If you have an existing run directory with generated outputs and want to evaluate it under `molbench` protocols, you can use the unified evaluation script: ```bash python molbench/eval/run_eval_bench.py <RESULTS_DIR> --cfg <CONFIG_PATH> ``` --- ## 📄 License This project is licensed under the MIT License - see the `LICENSE` file for details. --- ## 📝 Citation If you use MolClaw in your research or projects, please consider citing our paper: ```bibtex @article{zhang2026molclaw, title={MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization}, author={Zhang, Lisheng and Wang, Lilong and Sun, Xiangyu and Tang, Wei and Su, Haoyang and Qian, Yuehui and Yang, Qikui and Li, Qingsong and Tang, Zhenyu and Sun, Haoran and others}, journal={arXiv preprint arXiv:2604.21937}, year={2026} } } ``` --- <div align="center"> **Made with ❤️ by the MolClaw Team** [GitHub: InternScience/MolClaw](https://github.com/InternScience/MolClaw) • [⬆ back to top](#top) </div>

Loading the repository README…

Skills in this repository

Biochemists & Biophysicists Biological Scientists, All Other Software Developers Chemists

5 occupation categories · 21% classified

skill

occupation

description

updated

molclaw-admet

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__pred_mol_admet`. Predict the ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties of the input molecules.

2026-09-24

molclaw-boltz2-affinity

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__pred_binding_affinity_boltz2`. Predict binding affinity between target protein sequence and small molecule SMILES using Boltz-2.

2026-09-24

molclaw-chai1-predict

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__chai1_predict`. Predict protein structures with Chai-1 from sequence or FASTA input and return model scoring summaries.

2026-09-24

molclaw-chroma-toolkit

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__chroma_complex`, `mcp__molclaw-scp__chroma_monomer`, `mcp__molclaw-scp__chroma_symmetry`. Chroma toolkit skill covering chroma_monomer for single-chain generation, chroma_complex for multi-chain…

2026-09-24

molclaw-compound-retrieve

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__retrieve_smiles_by_compoundname`. Retrieve SMILES strings by compound name using PubChem with an NCI resolver fallback.

2026-09-24

molclaw-denovo-sampling

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__reinvent_denovo_sampling`. Generate new molecules de novo.

2026-09-24

molclaw-dleps

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_dleps_score`. Calculate disease reversal scores for the provided molecules relative to a specific disease.

2026-09-24

molclaw-drug-likeness

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_drug_chemistry`. Compute the drug-likeness metrics (QED score and Number of violations of Lipinski's Rule of Five) of the input candidate molecules (SMILES format).

2026-09-24

molclaw-equiscore-docking

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__convert_complex_cif_to_pdb`, `mcp__molclaw-scp__convert_pdb_to_pdbqt_dock`, `mcp__molclaw-scp__convert_smiles_to_format`. End-to-end docking-score ranking using EquiScore for candidate molecules…

2026-09-24

molclaw-equiscore-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__equiscore_pipeline`, `mcp__molclaw-scp__equiscore_pocket`, `mcp__molclaw-scp__equiscore_screen`. Unified EquiScore skill for pocket extraction, pocket scoring, and end-to-end docking-to-score…

2026-09-24

molclaw-esmfold

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__pred_protein_structure_esmfold`. Use ESMFold model to predict 3D structure of the input protein sequence.

2026-09-24

molclaw-evobind-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__evobind_tool`. Design linear or cyclic peptide binders from receptor FASTA sequences using EvoBind2 with structured result outputs.

2026-09-24

molclaw-extract-chains

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__extract_and_save_chains`, `mcp__molclaw-scp__extract_pdb_chains`. Extract protein sequence of each chain from the protein structure file (pdb format).

2026-09-24

molclaw-file-transfer

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__base64_to_server_file`, `mcp__molclaw-scp__server_file_to_base64`. Transfer files through the MCP client using a local path when supported, or Base64 directly.

2026-09-24

molclaw-fix-pdb

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__fix_pdb`. Repair and clean PDB or mmCIF structures with PDBFixer, returning a repaired PDB path and topology counts.

2026-09-24

molclaw-foldx-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__foldx_tool`. FoldX protein stability and mutation analysis tool. Supports 8 modes: structure repair (RepairPDB), stability calculation (Stability), mutation ΔΔG (BuildModel), complex interface…

2026-09-24

molclaw-fpocket

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__fpocket_toolkit`. Use fpocket to detect binding pockets and output their detailed properties for the input protein. This offers a more concise approach to pocket identification.

2026-09-24

molclaw-goca-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__goca_pipeline`. Run GoCa coarse-grained protein MD pipeline and collect key simulation artifacts from a unified run directory.

2026-09-24

molclaw-hdock-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__hdock_tool`. Run HDOCKlite docking for protein complexes and return run directories with ranked models.

2026-09-24

molclaw-interaction-visualizer

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__interaction_visualizer`. **PRIMARY tool for all single-structure interaction analysis.** MCP-exposed protein–ligand / peptide / protein–protein interaction analysis and Schrödinger-style…

2026-09-24

molclaw-karmadock-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__karmadock_tool`. Run KarmaDock graph generation and virtual screening to produce ranked ligand poses and summary metrics.

2026-09-24

molclaw-linker-sampling

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__linkinvent_linker_sampling_by_warhead_pair_name`, `mcp__molclaw-scp__linkinvent_linker_sampling_by_warheads`. Generate new molecules sampling from the input two warhead fragments.

2026-09-24

molclaw-mmpbsa-toolkit

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__analyze_mmpbsa`, `mcp__molclaw-scp__gmx_mmpbsa_propro`, `mcp__molclaw-scp__prepare_complex`, `mcp__molclaw-scp__prepare_protein_md`, `mcp__molclaw-scp__run_mmpbsa`. Prepare, run, and analyze…

2026-09-24

molclaw-mol-basic-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_basic_info`. Compute a set of basic molecular properties for a given list of SMILES strings, returning the molecular formula, exact and average molecular weights, counts of heavy and…

2026-09-24

molclaw-mol-charge-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_charge`. Compute Gasteiger partial charges and formal charge for a list of SMILES strings, returning the minimum, maximum, average, and range of the Gasteiger charges alongside the…

2026-09-24

molclaw-mol-complexity-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_complexity`. Compute custom molecular complexity-related descriptors for a given list of SMILES strings, returning the molecular complexity score, aromatic proportion, and asphericity…

2026-09-24

molclaw-mol-hbond-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_hbond`. Compute hydrogen bonding-related properties for a list of SMILES strings, specifically determining the number of hydrogen bond donors and acceptors for each input molecule.

2026-09-24

molclaw-mol-hydrophobicity-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_hydrophobicity`. Computes hydrophobicity-related molecular descriptors for a given list of SMILES strings, returning the octanol-water partition coefficient (logP) and molar…

2026-09-24

molclaw-mol-similarity

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_common_fragments`, `mcp__molclaw-scp__calculate_morgan_fingerprint_similarity`. Calculate both Tanimoto similarities and the count of shared structural fragments between a target molecule…

2026-09-24

molclaw-mol-structure-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_structure_complexity`. Compute a set of molecular structure complexity descriptors for a list of SMILES strings, returning detailed metrics for each molecule including the number of…

2026-09-24

molclaw-mol-topology-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_mol_topology`. Compute a comprehensive set of topological descriptors for a list of SMILES strings, returning the Topological Polar Surface Area (TPSA), a series of valence and non-valence…

2026-09-24

molclaw-mol2mol-sampling

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__reinvent_mol2mol_sampling`. Generate new molecules sampling from the input molecule.

2026-09-24

molclaw-openawsem-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__openawsem_sim`, `mcp__molclaw-scp__openawsem_traj_extract`. Runs OpenAWSEM simulations and extracts representative trajectory frames for downstream ensemble analysis.

2026-09-24

molclaw-p2rank

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__pred_pocket_prank`. Use P2Rank to locate binding pockets in the input protein. Unless specified by the user, prioritize using fpocket.

2026-09-24

molclaw-pack-sidechains

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__pack_sidechains`. Predicts full-atom sidechain conformations from backbone PDBs using AttnPacker for structure preparation workflows.

2026-09-24

molclaw-peptide-sampling

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__get_pepinvent_info`, `mcp__molclaw-scp__pepinvent_peptide_sampling_by_peptide`, `mcp__molclaw-scp__pepinvent_peptide_sampling_by_template`. Generate new peptide molecules sampling from the input…

2026-09-24

molclaw-prolif-tool

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__prolif_docking`, `mcp__molclaw-scp__prolif_md`, `mcp__molclaw-scp__prolif_pdb`, `mcp__molclaw-scp__prolif_protein_protein`. Unified ProLIF analysis skill covering MD trajectories, docking poses,…

2026-09-24

molclaw-protein-ligand-interactions

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__analyze_protein_ligand_interactions`. Analyze protein-small-molecule contacts, interaction types, pocket composition, and steric clashes from a complex PDB.

2026-09-24

molclaw-protein-metrics

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__calculate_pdb_basic_info`, `mcp__molclaw-scp__calculate_pdb_composition_info`, `mcp__molclaw-scp__calculate_pdb_quality_metrics`, `mcp__molclaw-scp__calculate_pdb_structural_geometry`,…

2026-09-24

molclaw-protein-openmm

unclassified

Use when calling these MolClaw MCP tools: `mcp__molclaw-scp__openmm_extract_frames`, `mcp__molclaw-scp__protein_openmm_md`. Run OpenMM protein MD and extract evenly spaced trajectory frames for downstream structural analysis.

2026-09-24

Showing 40 of 67 collected skills.

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/InternScience/MolClaw.