Scientific Agent Skills: Research Skills for AI Agents

Background and Context

Large language models have become good at writing code and searching for information. Real research is still a hard test for them. Scientific work depends on specialist software packages, domain databases and unwritten field conventions. A general model knows roughly which library handles single-cell data, how to query a compound database, or which checks a statistical result needs. The details are often wrong. Scientific Agent Skills, created by K-Dense, targets this gap. It is an open-source collection of research skills for AI agents, and it currently holds 177 ready-to-use scientific and research skills.

The project was first called Claude Scientific Skills. It was renamed because it no longer serves only Claude. It now targets any agent that supports the open Agent Skills standard. It is not another agent framework. It is a layer of procedural knowledge that sits on top of existing agent hosts and gives them domain ability. The team also published an arXiv paper, titled "Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents", and asks researchers to cite it when they use the library.

Deep Analysis

The main strength of the library is wide coverage combined with narrow focus. The skills span biology, chemistry, medicine, physics, engineering, earth science, data analysis and scientific communication. The project description also names drug discovery, bioinformatics, cheminformatics and clinical research. Yet each skill is tied to one concrete target: a specific software package, a data source or a workflow. Each one gives usage guidance and states the scientific conventions and validation checks of its field. This makes the skills different from a plain API manual. They teach an agent not only how to call a tool but also what counts as a correct result. The project is also designed to work with more than 100 scientific databases, so an agent can work against real data sources instead of guessing from memory. The packaging is practical. The library follows the open Agent Skills standard. It is also a portable package that meets the Agent Plugins standard, built from a plugin.json file and a skills directory. A client that supports this standard can load the whole collection as one plugin. The repository runs two continuous integration pipelines, one for security scanning and one for skill testing. For a library whose skills run code and reach external data, this signals real attention to quality and safety.

Use is direct. Compatible hosts such as Cursor, Claude Code, Codex and Google Antigravity can load the skills, so a researcher is not locked to one model vendor. The documentation includes an illustrated skills guide, a domain-based skill taxonomy and a full skills catalog. A user can pick the skills that match a research area instead of installing everything. For people who do not want to build their own environment, K-Dense offers K-Dense BYOK. It is a free, open-source AI co-scientist that runs on the desktop and is powered by these skills. Users bring their own API key and can choose from more than 40 models. The workspace includes web search, file handling, more than 100 scientific databases and all 177 skills. Data stays local, and heavy jobs can scale to cloud compute through Modal.

Industry Impact

The project shows a division of labour that is taking shape. Models and agent frameworks handle general reasoning and tool calling. Domain knowledge travels as reusable, reviewable skill packages. For a laboratory or an engineering team, this means local methods, checklists and conventions can be written down as skills. The team no longer has to restate them in a prompt each time.

The open standard also lowers the cost of switching. If a team changes its agent host, the skills still work. The repository has close to 50,000 stars, which shows strong interest from developers and researchers. K-Dense also keeps a community presence on X, LinkedIn, YouTube and Reddit, and it has published a recorded getting-started webinar.

The risks are real. A skill can shape how an agent works, but it cannot guarantee that a conclusion is right, so a human must still check research output. A large library changes fast. When a package or a database interface changes, a skill can go out of date. Running code and reaching external data also raise security questions. The project has security scanning, but users should still run skills in a controlled environment. The claim of use by more than 250,000 scientists comes from the project's own description. Readers should test it against their own setting.

Outlook

Three questions deserve attention. The first is quality control as the number of skills grows. A library of 177 skills is easy to review. A much larger one needs strong testing and clear ownership for each skill.

The second is how consistently different hosts behave. Support for Cursor, Claude Code, Codex and Google Antigravity is a strong promise. The real test is whether the same skill gives the same guidance and the same checks in each of them.

The third is reach. Local-first workspaces such as K-Dense BYOK may bring the library to researchers with little technical background. If that works, the skills could become a shared base for research agents. They would then act as a reviewable record of how each field expects its work to be done.