Four Projects, One Name: Comparing the ScienceClaw Ecosystem

Four ScienceClaws Walk Into a Lab...

In the fast-moving OpenClaw ecosystem, naming collisions happen. But four independent teams — spanning MIT, the Chinese Academy of Sciences, and individual researchers — all choosing ScienceClaw as their project name isn't just a coincidence. It reflects a shared conviction that AI agents are ready to transform scientific research.

Despite sharing a name, these four projects take fundamentally different approaches. Here's how they compare.

---

Quick Comparison

---

ScienceClaw (Zaoqu-Liu): The Minimalist

GitHub: Zaoqu-Liu/ScienceClaw

The most elegant of the three. The entire system is a single SCIENCE.md file (~600 lines) plus 266 domain skills — all in plain markdown. Zero custom code, zero TypeScript, zero Python servers.

What makes it unique:

• Research Recipes — 6 pre-built workflows (gene-landscape, target-validation, literature-review, etc.) that auto-detect from a single prompt
• 77+ databases including UniProt, PDB, TCGA, GTEx, ChEMBL, STRING
• Export to Word/PowerPoint/LaTeX with one command
• Literature monitoring via /watch command

Best for: Bioinformatics researchers who want a plug-and-play pipeline with zero setup friction. If you just want to type "analyze THBS2 in tumor microenvironment" and get a 30-page report with 87 citations, this is your pick.

---

ScienceClaw (TaichuAI): The Enterprise Platform

GitHub: AgentTeam-TaichuAI/ScienceClaw

Built by the NLP Group at the Chinese Academy of Sciences, this is the most ambitious of the three. It deliberately breaks from the OpenClaw architecture, using LangChain DeepAgents + AIO Sandbox as its foundation.

What makes it unique:

• 1,900+ built-in tools via Harvard's ToolUniverse — the largest tool library of any ScienceClaw
• Full web UI (Vue 3 frontend + FastAPI backend) with login system, file management, and resource monitoring
• Security-first design — everything runs in Docker containers with no host system access
• Full transparency — every step (search → reasoning → tool calls → output) is visible and traceable
• Windows desktop app available — no Docker needed

Best for: Teams and labs that need enterprise-grade security, a polished web interface, and don't mind a heavier deployment. The Docker-based 10-service architecture means more resources but also more capabilities.

---

ScienceClaw (beita6969): The Self-Evolving Colleague

GitHub: beita6969/ScienceClaw

The most innovative of the three. Its killer feature is self-evolution — the agent creates new skills at runtime based on your usage patterns. By week 4, it has specialized skills tuned to your specific subfield.

What makes it unique:

• Self-evolving skills — new SKILL.md files are generated automatically without redeployment
• 4-layer persistent memory with temporal decay weighting and LanceDB vector storage. "Continue the review from last Tuesday" actually works
• 1-hour+ session timeout (vs. OpenClaw's default 600s) with mandatory depth thresholds — Quick=5, Survey=30, Review=60, Systematic=100+ tool calls
• Zero-hallucination protocol — hard rule: every citation must come from a tool result in the current conversation. No fabricated PMIDs.

Best for: Individual researchers who want an AI colleague that grows with them. The persistent memory and self-evolution mean it gets better the more you use it. The zero-hallucination protocol makes it the most trustworthy for citation-heavy work.

---

ScienceClaw (MIT LAMM): The Decentralized Research Network

GitHub: lamm-mit/scienceclaw | Paper: arXiv:2603.14312 | Platform: Infinite

The most academically rigorous of the four. Built by Markus Buehler's lab at MIT, this ScienceClaw takes a fundamentally different approach: instead of one agent doing everything, it creates a network of independent agents that coordinate without central control.

What makes it unique:

• Decentralized multi-agent coordination — no central planner. Agents discover and fulfill each other's information needs through pressure-based scoring
• DAG artifact lineage — every output is an immutable artifact with typed metadata and parent lineage, forming a directed acyclic graph. Full computational reproducibility baked in
• ArtifactReactor — peer agents trigger multi-parent synthesis across independent analyses. Schema-overlap matching finds connections humans would miss
• Infinite platform — a shared scientific discourse platform where agent outputs become auditable scientific records with provenance views
• Autonomous mutation layer — prunes the expanding artifact DAG to resolve conflicting or redundant workflows
• 300+ interoperable tools with scientific profile-based selection

Best for: Research groups and labs that want multiple AI agents collaborating on the same problem space — each approaching it from a different angle. The DAG lineage and Infinite platform make it the most publishable and reproducible of the four. The only one with a peer-reviewed paper.

---

So Which One Should You Use?

The beauty of open source: you don't have to choose just one. They're all MIT-licensed, and each brings something the others don't.

---

Last updated: March 22, 2026. All four projects are actively maintained.