K-Dense scientific-agent-skills passes 51,000 stars as agent skills move into science

A skills library built for the laboratory

A GitHub repository of reusable skills for scientific AI agents passed 51,000 stars in mid-October, signaling that the agent-skills pattern popularized by developer tooling is now being ported into research. According to a dev.to post, K-Dense-AI/scientific-agent-skills sat at 51,288 stars on October 16, 2026, gaining roughly 1,158 stars per day and ranking as the third-fastest mover on GitHub Trending, up from around 44,000 stars in September.

The post describes the repository not as a prompt collection but as a curated library of 177 validated skills spanning biology, chemistry, cheminformatics, proteomics, clinical research, materials science and lab automation. Each skill uses the SKILL.md format that Claude Code and Cursor already expect, but the contents go well beyond a coding assistant's guidance. A protein-docking skill, for example, specifies AutoDock Vina parameters, requires the agent to validate docking scores against the PDBBind v2020 benchmark, and flags binding energies below -7.0 kcal/mol as potential hits needing molecular-dynamics follow-up.

Per the post, four properties distinguish these skills from their coding counterparts: parameters vetted against established benchmarks rather than guessed defaults; explicit dependency chains (most skills require uv for Python environments, and some need toolchains such as BLAST, HMMER or RDKit); output-validation rules that let an agent self-check results before reporting them; and safety constraints, with some skills refusing to generate synthesis routes for controlled substances or recommend dosages.

The project started as "Claude Scientific Skills," a Claude-Code-specific collection, and was renamed in 2026 to reflect support for any agent implementing the open Agent Skills standard, including Cursor, Codex and Google Antigravity. It is MIT-licensed, and K-Dense ships a companion desktop app, K-Dense BYOK, that runs the skills locally with a researcher's own API keys.

Two caveats apply. The post notes the project's claim of more than 250,000 researchers is self-reported and cannot be verified from outside, even though the star trajectory is genuine. And each skill can execute code, install packages, make network requests and modify files, so the project's own README advises reviewing every SKILL.md before installation.

Agent skills go vertical

The post situates the repository in a broader wave, citing coding-focused skills repositories such as obra/superpowers at 300,000 stars and affaan-m/ECC at 281,000. Its central argument is about stakes: coding skills fail cheaply, wasting tokens at worst, while a flawed molecular-synthesis skill could order hazardous reagents, contaminate a cell culture, or produce irreproducible results that waste months of downstream work. On that reading, science is the first vertical forcing the skills ecosystem to build trust infrastructure — provenance, benchmark validation and explicit refusal rules — on top of the "just ship a SKILL.md" ethos.

Sakana ships AI Scientist v3

In the same 24-hour window, Sakana AI's David Ha announced AI Scientist v3, branded "The Agentic Scientist," alongside the MASS paper on recursive self-improvement via multi-agent self-supervision. According to the post, the original system from August 2024 automated hypothesis generation, coding, analysis and paper drafting, and its methodology paper appeared in Nature in March 2026 with collaborators from UBC, the Vector Institute and Oxford. Version 2 dropped human-written templates and introduced agentic tree search at the cost of lower per-run success rates; one generated manuscript scored 6.33 at an ICLR 2025 workshop, but Sakana acknowledged that none of three generated papers met the bar for a main-conference publication.

Version 3 restructures the system as a team rather than a single linear pipeline: a principal-investigator agent decomposes problems, delegate agents explore sub-problems, and a review agent evaluates quality before anything is written up. The announcement also drew commentary from Jürgen Schmidhuber, who reprised his 1987 priority claims on recursive self-improvement while disclosing that he has joined Sakana AI as Chief Scientific Advisor.

The robotic half of the loop

The third signal in the cluster is physical rather than digital. At the Karlsruhe Institute of Technology, the Energy Materials Acceleration Platform (E-MAP) runs inside a glovebox-sealed environment where robotic systems prepare materials, mix semiconductor inks, deposit thin films and characterize samples for photovoltaics and optoelectronics without human hands. Construction began in 2023, per the post, which links E-MAP to a wider wave: PNNL's AutoLabs, described in Scientific Reports in June 2026, uses a translation agent to convert experimental goals into instructions for a robotic platform, enabling five to ten times more experiments than manual practice; a July 2026 Nature Machine Intelligence paper covered an agent autonomously running X-ray experiments at SLAC; and the Helmholtz Association is planning an alliance to connect autonomous labs across German research centers.

Why it matters

The convergence sketched in the post suggests a division of labor: skills libraries supply the protocols, parameters and validation rules, while self-driving labs supply closed-loop robotic execution. If SKILL.md becomes a standard container for scientific know-how, the binding question shifts from capability to trust — who validated a protocol, against which benchmark, and what the skill will refuse to do. For agent builders, mature vertical libraries are a distribution channel: any harness that can install the K-Dense library can turn a coding agent into a research agent. For researchers, a local, MIT-licensed AI co-scientist is now effectively a download away. The gap between that convenience and the consequences of a mis-validated protocol is where the ecosystem's next hard work lies.