Claw4Science: A Dataset and Platform for the OpenClaw Scientific Agent Ecosystem

This paper introduces Claw4Science, a curated dataset and public platform that unifies the fragmented OpenClaw scientific agent ecosystem by cataloging 91 projects and 2,230 skills to facilitate discovery, analysis, and standardized development in bioinformatics and scientific workflows.

The Gist

Imagine you are a scientist trying to build a robot assistant to help you discover new medicines or analyze DNA. Before this paper, building these robots was like trying to construct a custom car engine from scratch every single time you wanted to drive. If you wanted to add a new feature, like a GPS or a better radio, you had to be a master mechanic, rewrite the engine code, and hope it didn't break the whole car.

Because of this, every scientist had their own isolated robot. They couldn't easily share parts, and if you wanted to use someone else's robot, you often had to rebuild it yourself.

Enter "OpenClaw": The LEGO Box for Science
The paper introduces a new system called OpenClaw. Think of OpenClaw not as a single robot, but as a universal LEGO baseplate. Instead of writing complex code to add a new feature, scientists can now write a simple "recipe" (called a Skill) in plain text (Markdown).

• The Skill: Imagine a recipe card for making a sandwich. You don't need to be a chef to write it; you just list the steps: "Get bread, add peanut butter, add jelly." In OpenClaw, a "Skill" is a recipe for a scientific task, like "Analyze this DNA sequence" or "Find all papers about cancer."
• The Magic: Because these recipes are simple text files, any compatible robot can read and follow them. You can write a recipe once, and it works on Robot A, Robot B, and Robot C without you having to change a thing. This made the ecosystem explode with creativity.

The Problem: The "LEGO Store" Became a Mess
While this was great, it got out of hand very quickly.

1. Too many boxes: There are now over 91 different robot projects and 2,230 different "recipes" (skills).
2. Name Confusion: It's like having 20 different stores all named "Best Toy Store." Some are great, some are junk, and you can't tell which is which just by the name.
3. No Quality Control: Anyone can write a recipe. Some are perfect, but others might have the wrong ingredients (bad code) that could ruin your experiment.
4. Lost in the Wild: The recipes are scattered across the internet in different folders. There is no central map to find the best tools.

The Solution: Claw4Science (The "Yelp" for Science Robots)
The authors of this paper decided to clean up the mess. They built Claw4Science, which is like a centralized library and review board for this entire ecosystem.

Here is what they did, using simple analogies:

• The Dataset (The Master Inventory): They went out and cataloged every single robot project and every single recipe. They organized them into neat categories, like "Genomics" (studying genes), "Drug Discovery" (finding new cures), and "Writing Papers." They found 2,230 skills across 34 different scientific fields.
• The Platform (The User-Friendly App): They built a website (claw4science.org) that acts as a Google Maps for these tools.
  • Disambiguation: If you search for "ScienceClaw," the site tells you, "Wait, there are four different projects with that name. Here is the link to the real one you want."
  • Aggregation: Instead of visiting 50 different websites to find a tool, you visit this one site to see them all.
  • Focus: It specifically highlights tools for biology and medicine, making it easy for scientists to find what they need without getting lost in the noise.

What They Learned (The "State of the Union")
By looking at all this data, the authors found some interesting patterns:

• The "Long Tail": A few big, popular tools get all the attention, but there are hundreds of tiny, specialized tools for very specific tasks (like studying a specific type of bacteria) that are just as important.
• The Shift: Science is moving from "closed, secret labs" to "open, shared kitchens" where everyone can share their recipes.
• The Warnings: Just because a recipe exists doesn't mean it's safe to eat. Some tools are buggy, some rely on outside services that might change tomorrow, and there is no official "seal of approval" yet to guarantee they work perfectly.

The Bottom Line
This paper is a guidebook and a map for a rapidly growing world of AI scientists. It says: "Look how far we've come! We have a massive library of tools now. But we need to organize it, check the quality, and build better rules so that when a scientist uses these tools, they can trust the results."

The Claw4Science platform is the first step in turning a chaotic pile of LEGOs into a well-organized, easy-to-use construction kit for the future of science.

Technical Summary

1. Problem Statement

The paper addresses the fragmentation and lack of standardization in the rapidly growing ecosystem of OpenClaw, a framework for scientific AI agents. While OpenClaw introduced a modular, skill-based architecture that lowered the barrier to entry for scientific automation, its rapid expansion has led to several critical issues:

• Fragmentation: Projects and skills are scattered across independent repositories, making discovery and comparison difficult.
• Naming Collisions: Multiple independent projects share identical names (e.g., "ScienceClaw," "PaperClaw"), causing user confusion.
• Lack of Standardization: There is no unified directory, no standardized evaluation framework for "skills" (the core unit of extension), and no agreed-upon benchmarks for scientific capability.
• Reproducibility & Quality: Skills vary wildly in quality, and the reliance on external dependencies (APIs, models) creates reproducibility risks.
• Isolation: Unlike the pre-OpenClaw era where agents were closed systems, the current ecosystem lacks a coherent structure to systematically study or compare tools across domains.

2. Methodology

The authors employed a multi-faceted approach combining data engineering, systematic analysis, and platform development:

• Data Collection & Curation:

  • Source Identification: Used GitHub keyword searches, social media monitoring, community recommendations, and cross-references to identify candidates.
  • Manual Review: Manually reviewed projects to ensure relevance to scientific workflows and public availability.
  • Skill Extraction: Developed an automated pipeline using the GitHub API to scan 12 science-focused repositories, identifying and classifying 2,230 skills into 34 scientific categories (96.0% classification accuracy).
  • Dataset Construction: Curated a dataset of 91 projects (those with >50 GitHub stars) organized by functional role and structural relationship to OpenClaw.

• Structural Analysis:

  • The authors analyzed the ecosystem using two complementary views: a functional taxonomy (grouping by purpose, e.g., core runtime, orchestration, domain application) and a structural map (distinguishing direct descendants from independent but compatible systems).
  • They examined cross-cutting patterns including programming language usage, institutional participation, and the distribution of "stars" (popularity) vs. scientific relevance.

• Platform Development:

  • Built Claw4Science (https://claw4science.org), a public platform designed to aggregate the distributed ecosystem into a unified interface.

3. Key Contributions

A. The Claw4Science Dataset

The first curated dataset of the OpenClaw ecosystem, comprising:

• 91 Projects: Categorized by functional roles (Core Platform, Orchestration, Research Systems, Domain Applications, Skill Libraries).
• 2,230 Skills: Spanning 34 categories, with Genomics & Sequencing (31.4%) and Research Tools (25.8%) being the dominant categories.
• Metadata: Includes GitHub stars, activity status, and structural relationships between projects.

B. Systematic Ecosystem Analysis

The paper provides the first structural breakdown of the OpenClaw ecosystem, revealing:

• Layered Architecture: The ecosystem is not a flat list but a layered structure: Core Platform $\rightarrow$ Lightweight Variants $\rightarrow$ Orchestration Systems $\rightarrow$ Research Workspaces $\rightarrow$ Domain Applications.
• The "Skill" Paradigm: Identified the "skill" (a structured Markdown file containing natural language instructions and optional code blocks) as the primary engine of growth. Skills allow researchers to contribute workflows without modifying agent source code, enabling portability across different runtimes (e.g., OpenClaw, NanoBot, PicoClaw).
• Growth Patterns: Observed a shift from isolated, closed-loop systems to a modular, shareable model. However, growth is uneven, with a "long tail" of specialized tools and high concentration of attention on core platforms.

C. The Claw4Science Platform

A public web platform that serves as a unified entry point for the ecosystem:

• Project Directory: Organizes projects by function and resolves naming collisions by linking entries to specific repositories.
• Skill Hub: Aggregates distributed skill repositories into a single navigable layer.
• Disambiguation: Explicitly distinguishes between similarly named projects (e.g., different "ScienceClaw" implementations).
• Content: Includes tutorials, updates, and ecosystem insights.

4. Results & Findings

• Ecosystem Scale: As of March 2026, the ecosystem includes 91 major projects with over 450,000 total GitHub stars and 2,230 curated skills.
• Domain Distribution: While Genomics is the largest category, the ecosystem covers a broad spectrum including clinical/medical (8.0%), drug discovery (5.7%), and general research automation.
• Technical Diversity: Core platforms utilize a mix of TypeScript, Python, Rust, and Go, while science-facing tools are predominantly Python-based.
• Governance Gaps: The study highlights a lack of formal governance, leading to naming collisions (23 cases identified in 91 projects) and inconsistent quality control.
• Evaluation Void: No widely adopted benchmarks exist for scientific skill-based systems; existing general agent benchmarks do not capture domain-specific scientific rigor.

5. Significance

• Foundation for Benchmarking: The dataset provides the necessary ground truth for developing standardized benchmarks to evaluate scientific AI agents, moving beyond general capabilities to domain-specific correctness and reproducibility.
• Shift in Scientific Computing: The paper argues that OpenClaw represents a paradigm shift from "closed-loop" discovery systems to a modular, shareable ecosystem where workflows are decoupled from specific agent runtimes.
• Community Infrastructure: Claw4Science acts as critical infrastructure to prevent the ecosystem from becoming unmanageable due to fragmentation. It enables researchers to navigate, compare, and reuse tools without deep software engineering expertise.
• Future Directions: The authors identify critical open problems requiring community attention:
  • Governance: Establishing naming conventions and project lifecycle management.
  • Quality Control: Developing mechanisms for skill review and validation.
  • Reproducibility: Addressing issues related to model drift and external API dependencies.
  • Edge Deployment: Exploring lightweight agents for resource-constrained scientific environments.

In conclusion, this work transforms the OpenClaw ecosystem from a chaotic collection of repositories into a structured, analyzable, and accessible resource, laying the groundwork for the next generation of automated scientific discovery.