Search skills/

ScienceClaw Claude skills

A research agent system built on the OpenClaw gateway whose Skills and typed Operators are updated only after a replay-verified execution; ships with 319 skills.

beita6969/ScienceClaw 319 skills Science & academiaSkill & agent tooling
Our verdict Recommended with caveats

An ambitious research system with a published method, but it is heavy to install, relies on the OpenClaw platform, and its benchmark is the authors' own.

3/5 Depth expertise shipped Note Security Review before installing 3d Maintenance Updated 3 days ago 1.0K Adoption stars OK License MIT, commercial OK

What ScienceClaw does

ScienceClaw is not a plain skill pack. It is an agent system for scientific work built on the OpenClaw gateway, with a Python engine (packages/scienceclaw) that solves each task as a typed, executable workflow. When a repair to a failed run reproduces under a clean replay, the system records it as a linked Skill (a strategy) and Operator (a typed capability), and keeps it only if it still solves its source task and improves independent validation tasks. Updates are versioned snapshots that you promote and can roll back. The README reports gains on its own ScienceClaw-Eval benchmark across 23 disciplines (mean out-of-distribution gain of 16.45%); the benchmark data is hosted on Hugging Face and is the authors' own, so treat the numbers as theirs.

The repo includes 319 skill folders averaging about 7 KB. Many look inherited from the OpenClaw base (Feishu documents, 1Password, Apple Notes and Reminders, an ACP router for coding harnesses), alongside science skills such as academic-deep-research, academic-literature-search and an arXiv database skill. Setup is heavy: a one-click setup.sh installs Node, Python, the engine, tools and MCP servers, and the full profile downloads about 19 GB of model weights; the engine refuses tasks until setup completes. You supply an OpenAI-compatible endpoint, key and model through environment variables.

Good for

  • Researchers evaluating self-improving scientific agents
  • Teams already running the OpenClaw gateway who want science workflows
  • Study of replay-verified skill evolution

Not for

  • Anyone wanting a small, easy-to-install skill pack
  • Machines without room for tens of gigabytes of weights and tools

What's inside depth 3/5

A full engine and 319 skills averaging 7 KB with CI, but only two eval files in the repo and many skills are inherited gateway utilities rather than science procedures.

Skills
319
Reference docs
130
Scripts
160
Tests / evals
13
Templates
82
Avg SKILL.md
6.8 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • The OpenClaw gateway
  • Node.js, pnpm and Python 3.11+
  • An OpenAI-compatible LLM endpoint and key (SCIENCECLAW_API_KEY)
  • Up to about 19 GB of tool weights for the full profile

Security review

Review before installing

A large OpenClaw-based research agent bundle; we only hold a small part of it (about 85 of ~8,950 files), so this is a partial review. The skills mostly call public science APIs (arXiv, ChEMBL, Census) through curl or small Python scripts, while the Lobster plugin runs a local lobster binary. OpenProse can fetch and execute remote .prose programs from URLs, ACP routing uses unpinned npx -y packages, and the full setup installs many packages and ~19 GB of weights.

Worth knowing

  • Runs remote code at an unpinned versionOptional: prose run <url> (or registry shorthand resolved to https://p.prose.md/...) fetches a .prose program you name at runtime and the agent loads and executes it.extensions/open-prose/skills/prose/SKILL.md:71
  • Runs remote code at an unpinned versionACP router maps harnesses to npx -y @zed-industries/claude-agent-acp, npx -y opencode-ai acp, etc. with no version pin; README also uses npx openclaw onboard.extensions/acpx/skills/acp-router/SKILL.md:194
  • Runs remote code at an unpinned versionClawHub skill installs npm i -g clawhub and then installs/updates skills from clawhub.com on the fly (clawhub install, update --all --force).skills/clawhub/SKILL.md:25
  • Installs packages or toolssetup.sh / pnpm install, pip install -e packages/scienceclaw and python -m scienceclaw.cli setup install packages and download up to ~19 GB of model weights; many skills also tell the agent to pip/brew install CLIs.README.md:66
  • Relies on a CLI or package we could not reviewThe Lobster plugin spawns the external lobster binary (resolved from PATH) with a user/agent-supplied pipeline; other skills depend on op, clawhub, acpx, memo, remindctl, etc.extensions/lobster/src/lobster-tool.ts:285
  • Makes network requestsQueries export.arxiv.org and downloads PDFs from arxiv.org; other skills call ChEMBL (ebi.ac.uk), the Census API (key from CENSUS_API_KEY sent to api.census.gov), bioRxiv, etc.skills/arxiv-database/scripts/arxiv_search.py:23
  • Needs an API key for its own serviceEngine needs SCIENCECLAW_API_KEY for the user's own OpenAI-compatible endpoint; the Census skill uses CENSUS_API_KEY only against api.census.gov.README.md:89
  • Contains compiled binariesPer the dossier the full repo contains 11 binaries (gradle-wrapper.jar, tar/tgz test fixtures) that we cannot read; they are outside the files we hold.test/fixtures/hooks-install/tar-hooks.tar

What we checked

The repository (8952 files) on Oct 29, 2026. Not a full review: 8764 of 8952 files were not downloaded (large repository; we reviewed the skill folders); it contains compiled binaries or archives we cannot read; we read 7 of 9 code files; it relies on code outside the repository: The repository has 8952 files but only ~85 are in our snapshot (8764 readable files missing): the Python engine packages/scienceclaw, setup.sh, apps/, tests and most extensions are not available to us; 11 binary files in the full repo (e.g. apps/android/gradle/wrapper/gradle-wrapper.jar, tar fixtures under test/fixtures/hooks-install); npm packages openclaw, lobster, clawhub, @zed-industries/claude-agent-acp, opencode-ai, acpx (run via npx/npm by skills).

NVIDIA SkillSpector static scan: 53 pattern hits (Privilege Escalation 18, Data Exfiltration 13, MCP Rug Pull 7, Excessive Agency 4, Supply Chain 3, Memory Poisoning 2). SkillSpector hits are mostly documentation: 1Password CLI examples (printenv/.env), curl examples against public APIs (arXiv, ChEMBL) piped to a Python parser, apt/brew install steps in open-prose postgres docs, and 'reset state'/'maintain state' phrases. The real findings are the open-prose remote .prose fetching and unpinned npx use.

A review, not a certification. Skills change after we check them — read the files before installing.

License

MIT — commercial use allowed.

Can it run unattended?

Yes. It is designed to run tasks as replayable workflows with validation gates and versioned updates, although it needs your own model endpoint and a full setup first.

Alternatives to ScienceClaw

ToolUniverse

mims-harvard/ToolUniverse

Harvard MIMS lab's Python SDK, MCP server and 556 skills that give agents access to thousands of scientific databases and tools for biomedical research.

Recommended Broad access Depth 4/5

Scientific Agent Skills

K-Dense-AI/scientific-agent-skills

K-Dense's library of 177 science skills covering bioinformatics, chemistry, clinical research, data analysis, lab tooling and scientific writing.

With caveats Broad access Depth 4/5

AI Research Skills (Orchestra)

Orchestra-Research/AI-Research-SKILLs

98 skills for AI research and ML engineering, plus an autoresearch orchestrator that runs a two-loop experiment and synthesis cycle.

With caveats Broad access Depth 4/5

NanoClaw

nanocoai/nanoclaw

A small self-hosted assistant that runs Claude agents in Docker containers and connects to chat apps, extended by about 35 Claude Code setup skills.

With caveats Broad access Depth 4/5

FAQ

What is ScienceClaw?

ScienceClaw is an AI-for-science agent system on the OpenClaw gateway. It turns replay-verified repairs into versioned Skills and typed Operators, so the agent improves with use while you control which updates are promoted.

How do I install ScienceClaw?

Clone the repository and run setup.sh, or install manually with pnpm install, npx openclaw onboard and pip install -e packages/scienceclaw followed by python -m scienceclaw.cli setup. You also set an OpenAI-compatible endpoint, key and model; the full profile downloads about 19 GB.

Is ScienceClaw safe to run?

It is MIT-licensed and open, but it installs many tools and ships gateway skills that can reach credentials and files, so run it in a controlled environment. We could not review the installer script in the copy we inspected.

ScienceClaw vs ToolUniverse: which should I use?

ToolUniverse supplies a catalog of scientific tools through MCP, CLI and SDK. ScienceClaw is a whole agent runtime that evolves its own skills, so it is a research platform rather than a drop-in toolkit.

Sources: GitHub repository (stars, license, file tree, last push Oct 26, 2026). Review based on the README and SKILL.md files as published.