Primary question: Are automated scanning plus a written security policy enough to let a third-party skill library run scientific code through your agent?
RepoDaily adoption score
RepoDaily rates this as 91/100 (strong) for adoption: evidence, installation path, production risk, differentiation, license clarity, and AI/agent fit are scored from the article sources and adoption notes.
6 source(s) across 3 source category/categories, plus a RepoDaily-specific evidence module when available.
6 workflow step(s), 5 next-action step(s), and 2 command/install signal(s) were detected.
Trending momentum is +720 stars, with maintenance/release/issue signals counted when present.
Risk is marked medium, with 6 security note(s) and 4 explicit skip condition(s).
3 opportunity lens item(s), 3 alternative(s), and 4 type-specific section(s) support differentiation.
License source or license wording is present.
7 AI/agent-related signal(s) were detected in the article text and metadata.
Project overview
Scientific Agent Skills (K-Dense-AI/scientific-agent-skills) is a Python-packaged collection of 163 ready-to-use scientific skills and, per its badges, 100+ connected scientific databases, currently at version 2.64.0 under an MIT license. The coverage named in the README runs from cancer genomics, individual-level 1000 Genomes queries, hosted regulatory-sequence prediction, and live pathogen-variant surveillance to PK/PD modelling and dose selection, drug-target binding, molecular dynamics, RNA velocity, and microbiome foundation models. The project was renamed from Claude Scientific Skills to Scientific Agent Skills, and the README now badges Cursor, Claude Code, Codex, and Google Antigravity as supported hosts. It pulled 720 stars this period at rank 7, and its own description claims 175,000+ scientists in use — a figure we cannot independently verify and treat as the maintainer's claim, not a measured fact.
We read it as a security story because security is the deciding factor for this class of library. A skill is a bundle of instructions plus scripts that an agent may execute on your machine, and SECURITY.md says outright: treat skill content from any source — including this repository — as code review material, not trusted input. To back that stance, the repo runs a security-scan.yml workflow and a skill-tests.yml workflow in CI, performs automated scanning with cisco-ai-skill-scanner (pinned at >=2.0.12 in pyproject.toml), and publishes a vulnerability policy that names exact in-scope and out-of-scope cases rather than boilerplate.
The bargain on offer: you get spec-compliant, tested skill assets — SKILL.md files kept under the recommended 500 lines, a closed six-field frontmatter contract, tests colocated under tests/<skill-name>/ — and in exchange you accept third-party code execution inside your agent and the duty of moving your own version pins forward, since the project does not backport fixes to earlier tags. The rest of this review weighs both sides of that trade.
Why it is trending now
- The rename from Claude Scientific Skills to Scientific Agent Skills widened the audience: the README badge moved from one vendor to four listed hosts — Cursor, Claude Code, Codex, and Google Antigravity.
- Scale draws attention: version 2.64.0 bundles 163 skills and 100+ scientific databases, from cancer genomics and PK/PD dose selection to RNA velocity and microbiome foundation models.
- 720 stars in the period at rank 7 — timing that lines up with the rename and the companion-app announcement rather than a single new feature.
- The free, open-source companion app K-Dense BYOK runs these skills on the desktop with 40+ models and your own API keys, optionally scaling to Modal cloud compute, and funnels new users back to the library.
- Rigor that is rare for a prompt collection: CI security scanning, a dedicated skill-test workflow, and a security policy that explicitly lists prompt injection inside references/ and assets/ as in-scope.
- The project description's claim of 175,000+ scientists is unverifiable, but it signals how far the collection has already spread through research groups.
Problem it solves
- General-purpose agents give generic answers; procedures like individual-level 1000 Genomes queries, drug-target binding, or bounded biomedical knowledge-graph search need curated instructions and helper scripts.
- Labs rebuild the same prompts and scripts ad hoc, with no validation, no tests, and no version management.
- Installing third-party skills is inherently risky — they execute code and steer agent behavior — and most collections ship with neither a security policy nor any scanning.
- Bloated instructions degrade agent performance; the Agent Skills spec recommends keeping SKILL.md under 500 lines with progressive disclosure, which is hard to sustain without tooling.
- Version drift: pin a skill to a tag and it stops receiving fixes, because the repo explicitly does not backport to earlier tagged releases.
How it works
- Each skill is a directory under skills/ with one required file, SKILL.md (YAML frontmatter plus Markdown instructions), and three optional directories — references/, scripts/, assets/ — the only four a skill may contain.
- The agent loads SKILL.md; longer material stays in references/ and is read only when needed, keeping the main file under the spec's recommended 500 lines.
- Scripts in scripts/ are executable helpers the agent may run on your machine; the security policy states that scripts reaching the network or reading credentials are documented as such in the relevant SKILL.md.
- Frontmatter is strict: name must match the directory (1-64 chars, lowercase letters, numbers, hyphens, no consecutive hyphens), description max 1024 characters, metadata.version a quoted numeric string, and only six top-level fields survive the reference validator, which parses with strictyaml and rejects JSON-style flow mappings.
- The repository root is an Agent Plugins package: plugin.json must stay schema-valid with its version synced to pyproject.toml, currently 2.64.0.
- CI runs security-scan.yml and skill-tests.yml; tests live under tests/<skill-name>/ with fixtures in tests/<skill-name>/fixtures/, and pytest runs with --import-mode=importlib so skills named after the libraries they wrap (neurokit2, simpy, qutip) do not shadow real packages during collection.
Architecture read: a spec-validated skill package
The unit of distribution is the skill directory: skills/<name>/ holding SKILL.md plus optional references/, scripts/, and assets/. Anything else — tests, fixtures, scratch output — must stay outside skills/, which keeps an agent's load surface small and predictable. The root plugin.json makes the whole repo consumable as an Agent Plugins package, while pyproject.toml declares version 2.64.0, requires-python >= 3.13, and runtime dependencies on cisco-ai-skill-scanner>=2.0.12, firecrawl-py>=4.9.0, pytest>=9.1.1, and python-dotenv>=1.0.0.
A reference validator enforces format: the dev dependency group pulls skills-ref from the agentskills repository, and CONTRIBUTING.md documents the closed six-field frontmatter set, the strictyaml parsing rules, and the 500-line SKILL.md guidance. Tests are colocated at the repository level rather than inside skills, mirroring skill directory names so a failing skill is locatable from the test path alone.
Maintenance risk: pin carefully, and read the drift
- SECURITY.md supports only main and the latest tagged release; earlier tags receive fixes only through new releases, so a pinned install is the user's job to move forward.
- Doc drift is already visible in the README: the badge says 163 skills while the BYOK paragraph in the same file says 'all 161 skills', and the databases badge says 100+ while the skills list says '78+ scientific databases.'
- Two runtime dependencies hint at reach: firecrawl-py (web fetching) and python-dotenv (environment variable loading) install by default alongside the scanner.
- At version 2.64.0 the release cadence is high, so pinned subsets can age quickly against upstream database and API changes.
Try-it path: a three-skill pilot in one afternoon
- Start with SECURITY.md's 'Before you install' section; the maintainers' own advice is to install only the subset you need and review everything first.
- Pick three skills matching work you already do by hand — the README names analytical method validation, PK/PD modelling and dose selection, and time series forecasting among the 163.
- Open each SKILL.md and every file in scripts/; verify that documented network calls and credential reads match what the code actually does.
- Load only that subset into one host — Cursor, Claude Code, Codex, or Google Antigravity — and rerun an analysis whose correct answer you already know.
- If wiring an agent host is too much setup, the README links a recorded webinar walking through the K-Dense BYOK desktop route with your own API keys.
Integration surface: hosts, standards, and the desktop app
- Any agent implementing the open Agent Skills standard (agentskills.io); the README badges Cursor, Claude Code, Codex, and Google Antigravity.
- The repo root is also an Agent Plugins package (agent-plugins.org), so plugin-aware hosts can consume the collection as one unit with a synced version.
- K-Dense BYOK uses the library as its engine: a desktop app with 40+ models via your own API keys, web search, file handling, 100+ databases, local data storage, and optional Modal scaling for heavy compute.
- Skills document external services — 100+ databases plus Hugging Science for scientific ML resource discovery — so results depend on third-party endpoints remaining reachable.
Who should pay attention?
Good fit if
- Bioinformatics, genomics, proteomics, metabolomics, and drug-discovery groups already running Claude Code or Cursor in daily analysis.
- Labs that want tested, versioned scientific procedures instead of copy-pasted prompts; tests sit under tests/<skill-name>/ for the collection.
- Groups that need a security-reviewed on-ramp to agent skills: private vulnerability reporting, CI scanning, and a written scope policy already exist here.
- Researchers who want a local co-scientist through K-Dense BYOK with their own API keys and data kept on their own machine.
Skip for now if
- Anyone unwilling to read the scripts their agent will execute; the project's own security policy calls skill content code review material, not trusted input.
- Environments pinned below Python 3.13 — pyproject.toml sets requires-python >= 3.13.
- Regulated settings that need frozen, supported versions: earlier tagged releases get no backported fixes.
- Users whose agent host does not implement the Agent Skills standard.
Risks and cautions
Well-run for its category — scanning, tests, and a real security policy are all present — but the category itself means third-party code execution inside your agent, and the timeliness of fixes depends on you moving version pins.
- Skills bundle scripts the agent may execute; SECURITY.md lists credential reading, data exfiltration, and prompt injection in references/ and assets/ as in-scope vulnerability classes, which shows the risk model is concrete, not theoretical.
- No backports: only main and the latest tagged release receive fixes, so pinned installs age by design.
- Skills wrap external services (100+ databases and public APIs), so results depend on endpoints the project does not control; per the policy, bugs in libraries like RDKit or Scanpy belong upstream.
- Counterweights: MIT license, security-scan.yml and skill-tests.yml in CI, cisco-ai-skill-scanner>=2.0.12 pinned as a dependency, and a documented reporting path via GitHub private vulnerability reporting.
- Report vulnerabilities through GitHub's private vulnerability reporting (Security tab -> Report a vulnerability), never a public issue; accepted reporters are credited in release notes unless they decline.
- In scope: scripts that read credentials, files, or environment variables without cause or transmit data unexpectedly; SKILL.md instructions steering an agent toward destructive, exfiltrating, or unauthorized action; prompt-injection content in references/ or assets/; skills whose docs misrepresent bundled code; unsafe secret handling; and flaws in scan_skills.py, scan_pr_skills.py, or the GitHub Actions workflows.
- Out of scope: bugs in third-party libraries (RDKit, Scanpy, public APIs) and bugs in agent hosts like Claude Code or Cursor — those belong to the upstream vendor.
- Reports should name the affected skill, the version or commit observed, what an attacker could achieve and the access needed, a minimal reproduction, and the agent host and model used, because skill behavior varies by host.
- Automated scanning runs through cisco-ai-skill-scanner, and the security-scan.yml workflow badge is surfaced directly in the README.
- Maintainer guidance: install only the subset of skills you actually need, and treat skill content from any source — including this repository — as code review material.
Alternatives to compare
| Approach | When to use | Trade-off |
|---|---|---|
K-Dense BYOK | You want a ready desktop research workspace — web search, file handling, 100+ databases — powered by these same skills instead of wiring skills into your own agent. | Free and open source; you pay only for your own model API keys. |
Hand-built SKILL.md files on the Agent Skills spec | You need two or three bespoke lab procedures and want full control over every instruction and script. | Engineering time only; the reference validator (skills-ref) is available from the agentskills repository, which this project itself uses as a dev dependency. |
Direct library use (e.g., RDKit, with Scanpy and public APIs alongside) | An agent-free code path is acceptable or required and you would rather own every line that touches your data. | Open-source libraries plus internal engineering time; SECURITY.md itself directs library bugs upstream. |
What this trend reveals
A publishable template for domain skill packs
The repo's conventions — a closed six-field frontmatter, strictyaml-clean YAML, tests under tests/<skill-name>/fixtures/, scanning wired into CI — form a copyable blueprint for any organization shipping internal skills in other domains such as reservoir engineering or clinical reporting.
Clone the repo, follow CONTRIBUTING.md's template to write one internal skill, and push it through the same frontmatter rules before building any tooling of your own.
Skill review as a serviceable discipline
SECURITY.md's in-scope list reads like a ready-made review checklist — credential reads, exfiltration, prompt injection in references/ — and the repo already depends on cisco-ai-skill-scanner>=2.0.12, so third-party skill auditing is technically tractable today.
Take five skills, audit every file in their scripts/ directories against the in-scope list, and time the effort; that number prices an internal or commercial review service.
Curation layers over 163 skills
The maintainers themselves advise installing subsets, and the README's counts drift (163 on the badge, 161 in the BYOK paragraph; 100+ databases on the badge, 78+ in the skills list) — exactly the confusion a curated, domain-filtered index would remove.
Build a filtered list for one field, say proteomics plus metabolomics, and measure whether researchers choose skills faster than by browsing skills/ directly.
RepoDaily verdict
A rare agent-skills library that treats security as a first-class design constraint: MIT-licensed, scanned in CI, spec-validated, and candid about its own risks. Adopt it in reviewed subsets and pin versions deliberately, and it earns this period's 720 stars; install all 163 blind and you have outsourced code review to a badge.