Scientific Agent Skills
K-Dense-AI/scientific-agent-skillsK-Dense's library of 177 science skills covering bioinformatics, chemistry, clinical research, data analysis, lab tooling and scientific writing.
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.
A large, actively maintained research toolkit from an academic lab with real code behind the skills and a stated no-telemetry policy; expect some setup and API keys.
ToolUniverse is a research platform for building "AI scientist" agents. The core is a Python package (tooluniverse on PyPI) that exposes a large catalog of tools, which the setup skill puts at more than 2,700, covering databases and services such as PubMed, UniProt, ChEMBL, FAERS and ClinicalTrials.gov. It can run as an MCP server, through a tu command-line tool with subcommands for listing, searching, running and testing tools, or as a Python SDK. In MCP mode it starts in a compact setup that exposes five core tools (list, grep, get info, execute, find) while keeping the rest reachable through execute_tool.
The repo's 556 skills are mostly workflow guides written for specific research tasks. Names in the sample include ACMG variant classification, ADMET prediction, adverse event detection, antibody engineering, binder discovery, cancer genomics with TCGA and biomedical fact lookup. They average about 14 KB, a number have reference files, and one skill, setup-tooluniverse, walks a user step by step through choosing between MCP, CLI and SDK, configuring more than a dozen clients, and adding API keys. There are about 40 eval files and CI.
Installation options include a Claude Code plugin marketplace, npx skills add, an MCP config using uvx, or pip-style installs through uv. Optional extras add local machine-learning and cheminformatics tools.
556 skills averaging 14 KB with reference files in many, backed by a real SDK, CLI and MCP server with CI and about 40 eval files.
Counted from the repository file tree. How we score depth.
This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.
A large set of scientific-database skills for ToolUniverse, with setup scripts and a few analysis pipelines. Local use runs the unpinned tooluniverse Python package and edits your MCP client configs. The optional hosted ToolUniverse Connect path has the agent fetch and follow guides from the vendor's server, pipes a vendor skills.sh into sh, and has you store database API keys on the vendor's website; the README says the local install collects no telemetry.
The repository (9958 files) on Oct 29, 2026. Not a full review: 9803 of 9958 files were not downloaded (large repository; we reviewed the skill folders); we read 10 of 26 code files; it relies on code outside the repository: PyPI package tooluniverse (the real ToolUniverse implementation, installed by uvx or pip; includes code-executor tools); hosted services mcp.aiscientist.tools and connect.aiscientist.tools, including skills.sh and the guides returned by get_skill; https://aiscientist.tools/setup.md and the uv installer from astral.sh.
NVIDIA SkillSpector static scan: 168 pattern hits (MCP Rug Pull 58, Tool Misuse 28, Data Exfiltration 23, Supply Chain 14, Agent Snooping 10, Privilege Escalation 10). Most of the 168 SkillSpector hits are documentation: unpinned uvx mentions, sudo/apt and rm -rf cleanup commands in troubleshooting docs, MCP config file paths, and example URLs. The hits that mattered were the curl-to-sh installers and the hosted-guide flow, which are recorded as flags. The scripts I opened contain no obfuscation or credential reads.
A review, not a certification. Skills change after we check them — read the files before installing.
Apache-2.0 — commercial use allowed.
Yes. The README describes async operations for long-running tasks such as docking and simulations with progress tracking, and the SDK is meant for building autonomous research pipelines.
K-Dense's library of 177 science skills covering bioinformatics, chemistry, clinical research, data analysis, lab tooling and scientific writing.
98 skills for AI research and ML engineering, plus an autoresearch orchestrator that runs a two-loop experiment and synthesis cycle.
Official Hugging Face skills led by hf-cli, a Hub CLI reference generated from your local install, plus AWS/SageMaker deployment and operations skills.
ToolUniverse is an open-source ecosystem from Harvard's MIMS lab that connects language-model agents to thousands of scientific tools and databases through an MCP server, a CLI and a Python SDK. The repo also ships 556 skills for research workflows.
Run claude plugin marketplace add mims-harvard/ToolUniverse and then claude plugin install tooluniverse@tooluniverse. You can also add the MCP config with uvx or run npx skills add mims-harvard/ToolUniverse.
The README states it runs locally with no telemetry or analytics, and API keys go only to the services they belong to. The package is unpinned by default in the suggested MCP config, so consider pinning a version for reproducible work.
ToolUniverse supplies a tool runtime (MCP, CLI, SDK) with a very large database catalog plus skills on top. K-Dense's scientific-agent-skills is a skills library that relies more on the agent's own code and packages, so it fits you if you do not want a separate runtime.