Scientific Agent Skills: 177 Research Skills for Claude Code, Codex and Cursor

K-Dense's Scientific Agent Skills repository was trending on GitHub on October 14, 2026. The project packs 177 scientific and research skills into one MIT-licensed collection that works with Claude Code, Codex, Cursor, Google Antigravity and any other host that supports the open Agent Skills standard. It is the new name for the project formerly called Claude Scientific Skills, renamed to show it is no longer Claude-specific.

This guide explains what is inside the Scientific Agent Skills repository, how to install it, what it is good for, and what to check before you let an agent run it unsupervised. We read the README and release notes; we did not run the skills on research data, so we make no claim about how accurate the outputs are.

Scientific Agent Skills at a glance

What is in the 177 skills?

The README organizes the collection by domain. A sample of what it lists:

• Bioinformatics and genomics: sequence analysis, single-cell RNA-seq, pooled CRISPR screens, primer design, amplicon microbiomes, variant annotation and phylogenetics.
• Cheminformatics and drug discovery: molecular property prediction, virtual screening, ADMET analysis, docking, lead optimization and calibrated 1D NMR processing.
• Proteomics and mass spectrometry: LC-MS/MS processing, peptide identification and protein quantification.
• Clinical and healthcare research: clinical trials, pharmacogenomics, pharmacokinetic modelling and evidence workflows. The README stresses these are for research and retrospective validation, not patient-specific diagnosis, treatment or deployment decisions.
• Imaging and neuroscience: quantitative microscopy, privacy-aware DICOM processing, research-only whole-slide image analysis, BIDS datasets and NWB conversion.
• Physics, materials, geoscience and engineering: crystal structure analysis, phase equilibria, battery cycling models, geospatial remote sensing, fluid dynamics and discrete-event simulation.
• Data and communication: statistics, network analysis, publication-quality figures, literature synthesis, posters, slides and citation management.

Release notes show how the collection is maintained. Version 2.71.0 refreshed all 181 skills, including a Database Lookup skill covering 80 sources with access requirements and provenance checks, and updates such as PathML 3.0.8 and GeoPandas 1.2.0. Version 2.72.0 removed four document skills (docx, pdf, pptx and xlsx) that had been vendored from Anthropic's skills repository under Anthropic's own license, bringing the count to 177; the README tells users to install those four from Anthropic's repository if they rely on them. A "recently added workflows" table lists assay design (Primer Design, FlowKit, MAGeCK, CellProfiler), spectra and metabolism (nmrglue, 13C Metabolic Flux, Tellurium), structure and neural data (RELION, NWB Conversion), microbiome and seawater chemistry, and energy and materials simulation (PyBaMM, Cantera, pycalphad).

How do you install and use it?

The README offers several routes.

1. npx (supported hosts). Run npx skills add K-Dense-AI/scientific-agent-skills. It uses the common installer for Agent Skills hosts; confirm install paths in your host's documentation. We cover that tool in our guide to installing skills with npx skills.
2. GitHub CLI. With v2.90.0 or later, gh skill can install skills too, following GitHub's skill-management changelog.
3. Plugin. The collection is also a portable Agent Plugins package (plugin.json plus skills/), so clients supporting that standard can load it as one plugin.
4. K-Dense BYOK. A free desktop co-scientist that runs locally with your own API keys, offers 40+ models and web search, and can use all 177 skills. Data stays on your computer, with optional cloud compute through Modal.

After install, ask your agent for a task in plain language, such as "analyze this single-cell RNA-seq dataset" or "design primers for this target", and the agent loads the relevant skill. For background on how skills are discovered and loaded, see our complete guide to agent skills.

What are people asking about it?

Is this just documentation? Largely, yes, and that is the point. The README says the agent can use any Python package or API on its own, and the skills add curated documentation, examples, scientific conventions and validation checks to make results more reliable. Each skill records its own compatibility requirements and validation scope, and the README notes that local tests, live-service checks and illustrative examples have different coverage.

Does it work outside Claude? Yes. The rename exists because the skills follow the open standard and the README lists Cursor, Codex and Antigravity alongside Claude Code.

Why was it renamed? The notice says "same skills, broader compatibility" for any agent supporting the Agent Skills standard.

Is there evidence it improves results? The project points to a paper on arXiv, but we have not reviewed its experiments, and we cannot say what it shows. Treat any improvement claim as unverified until you test on your own tasks.

A practical first session

A flask beside a sheet with a spinning gear, illustrating scientific agent skills running a research analysis

If you want to judge the collection quickly, a small, checkable task beats a grand one. Here is a way to run a fair trial.

1. Pick a task whose answer you already know. Reuse a dataset from a past paper or course, such as a differential expression analysis or a molecular property table, so you can compare the agent's output with a trusted result.
2. Install in a scratch project. Use a fresh directory and a fresh agent session so the skills do not mix with other tooling, and note which skills the agent says it loaded.
3. Ask for provenance. Request that the agent list the databases, package versions and parameters it used. The README stresses provenance checks and validation steps in several skills, so a good run should show them.
4. Compare with and without skills. Run the same prompt once with the collection disabled. If the answers match, the skills add little for that task; if the skilled run catches an edge case the plain run missed, that is real evidence.
5. Record cost and time. Skills add context tokens, so measure the extra spend against the quality gain.
6. Share the failure cases. The project accepts contributions and issues, and a reproducible bug report on a skill is more useful to everyone than a star.

This kind of paired test is also how you should evaluate any skills library, whether it covers science, design or DevOps. Leaderboards and star counts say little about whether a skill helps with your data.

How does it compare with other AI-for-science tooling?

This collection is general open tooling, while several labs ship dedicated science stacks. We have covered NVIDIA's BioNeMo agent toolkit for drug discovery, Novo Nordisk's adoption of Claude Science and Claude designing protein binders. Those are vendor platforms or lab results. Scientific Agent Skills is a free, host-neutral layer you can bolt onto the agent you already use, with the trade-off that quality varies by skill and you carry the validation burden.

Safety and trust: what to check before you enable skills

A skill is text and scripts an agent will follow with your permissions. For a collection of this size:

• Read before enabling. Open the SKILL.md for each skill you plan to use, plus any scripts it runs.
• Enable only what you need. 177 skills in context can bloat prompts and cause the agent to pick the wrong one.
• Sandbox execution. Run analysis in a container or throwaway environment, and keep credentials out of it.
• Check data handling. Some skills call external databases and APIs. Do not send patient, proprietary or regulated data to services you have not vetted.
• Respect the research-only limits. The clinical and healthcare skills explicitly exclude patient-specific diagnosis or treatment decisions.
• Pin a version. The collection is updated frequently, and a refresh can change behavior. Record the version (2.72.0 at the time of writing).

The repository advertises a security-scan workflow and includes a security disclaimer section, which is a good sign of hygiene but not an audit. For teams running agents on real systems, a runtime control layer such as AgentBeam, from the explainx.ai team, can block dangerous actions before they run, regardless of which skills you load.

Who should use it?

Researchers and data scientists who already work with Claude Code, Codex or Cursor and repeat the same bioinformatics, chemistry or statistics workflows will get the most value. It is also a useful reference for anyone writing skills, since 177 working examples show how to encode domain conventions, validation checks and provenance requirements. If you work in a regulated clinical setting, treat it as a prototyping aid and keep a human expert in the loop.

Bottom line

Scientific Agent Skills turns the Agent Skills standard into something practical for science: a large, maintained, MIT-licensed catalog you can install in one command. Its trending spike shows demand for domain-specific agent knowledge. Its limits are the usual ones for third-party skills: you must review them, verify outputs, and keep sensitive data out of loops you cannot inspect.

Related reading

• What are agent skills? The complete guide
• Install Claude Code skills with npx skills
• NVIDIA BioNeMo agent toolkit for drug discovery
• Novo Nordisk adopts Claude Science
• Claude designed protein binders for 14 of 15 targets
• Human in the loop: when to let an agent run

Version numbers and counts are as shown in the project's README on October 14, 2026 and change often.