MolClaw: An Autonomous Agent with Hierarchical Skills for Drug Molecule Evaluation, Screening, and Optimization

A reproducible pathway for autonomous molecular evaluation, screening, and optimization.

✨ If MolClaw helps your research or engineering workflow, we'd really appreciate a star.

---

📖 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.

---

📊 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.

---

🤖 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.

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:

mkdir -p ./.claude/skills
cp -R <YOUR_CLONED_MOLCLAW_DIR>/skills/* ./.claude/skills/

---

🚀 Quick Start

1) Clone Repository and Create Conda Environment

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

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) Claude Code Interact with Skills

Copy the Skills into your Claude Code workspace and then you can interact with them in your conversations.

# 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

4) Benchmark Runs

Use the commands below for baseline and Claude benchmark runs; each command performs inference and automatic molbench evaluation.

# 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/.

5) 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:

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:

@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 Han, Yingnan and Jiang, Yankai and Lou, Wenjie and Zhou, Bowen and Wang, Xiaosong and Bai, Lei and Xie, Zhengwei},
  journal={bioRxiv},
  year={2026},
  url={https://www.biorxiv.org/content/10.64898/2026.04.03.716272v1}
}

---

Made with ❤️ by the MolClaw Team

GitHub: InternScience/MolClaw • ⬆ back to top