GitHub Trending has turned into a marketplace for AI agent skills and harnesses

GitHub Trending as a storefront in waiting

According to a dev.to post, GitHub Trending has quietly changed character: instead of surfacing unrelated projects, the current leaderboard reads like a catalogue of tooling for AI coding agents. The post points to four heavily starred repositories — ECC, awesome-design-md, scientific-agent-skills and Orca — and argues they are not isolated hits but distinct layers of one emerging stack: what a marketplace looks like before anyone builds the storefront.

The numbers cited are striking. ECC, a harness optimizer, sits at 245,643 stars and ships 277 skills plus 67 agents that run across Claude Code, Cursor, Codex and other tools. awesome-design-md, at 112,505 stars, is a library of DESIGN.md files that give agents the visual context needed to generate UI matching a brand's specifications. scientific-agent-skills has reached 41,399 stars with 165 validated skills spanning biology, chemistry, medicine and drug discovery, and is reportedly used by 190,000 scientists. Orca, an agent development environment at 59,072 stars, coordinates fleets of 25 or more coding agents working in parallel git worktrees.

The four-layer stack

The dev.to analysis breaks the ecosystem into layers, each solving a different problem.

ECC anchors the foundation layer. With 289 contributors, it packages skills, agents, hooks and security tooling so a single repository can shape agent behavior across every major coding tool. The post frames it as the community-driven counterpart to the harness-as-moat thesis — following DeepSeek open-sourcing its own harness in August — with its 277 workflow skills representing accumulated configuration no individual operator would rebuild alone.

awesome-design-md represents the design-spec layer. Drop a DESIGN.md file into a project root and the agent picks up colors, typography, spacing and interaction patterns without Figma exports or JSON schemas. An OSS Insight analysis cited in the post shows a fork-to-star ratio of 12.6%, which the author reads as developers integrating the library into production rather than starring out of curiosity.

scientific-agent-skills proves the vertical layer. Drawing on more than 100 scientific databases, it turns a coding agent into a research assistant that knows how to query PubMed, parse crystallography data and order statistical analyses correctly. The post argues this is where real value accrues, because domain expertise is hard to replicate.

Orca completes the stack at the fleet layer, orchestrating Claude Code, Codex, OpenCode and 20-plus other agents against users' own subscriptions. Per the post, it went from first commit to 59,000 stars in five months.

The marketplace signals

The distribution infrastructure already exists, according to the post: searches for "agent skills" grew 19-fold over two years, from 21,000 to over 400,000 per month; more than 20 coding agents support the SKILL.md standard; the SkillsMP index tracks 1.9 million public skills; and paid marketplaces such as Agensi offer creators a 70/30 revenue split.

Security researcher Daniel Cuthbert warned that "MCP is the new App Store," likening the moment to "the iPhone App Store in July 2008" because the agent selects a tool and the user never sees the alternatives. The dev.to post accepts the analogy but highlights its unflattering side: early app stores shipped thousands of trivial apps before quality gates emerged, and skills are in that phase now.

The gatekeeping gap

Here the comparison breaks down, the post argues. iOS and Google Play launched with review processes and kill switches; skills directories have GitHub stars and README files. An audit of 22,511 skills cited in the post found 140,963 issues, prompt injection in 36% of tested skills, and an average quality score of 6.2 out of 12. An unaudited skill is effectively an unaudited dependency with code-execution rights. The question is not whether curation emerges, but whether it arrives before a high-profile supply-chain attack forces it.

How practitioners are reacting

Hacker News threads referenced in the post show a split. One camp treats skills as the natural evolution of linting rules and style guides — codified team practices that agents enforce automatically. The other expects the abstraction to fragment when the next generation of tools arrives. A separate debate asks whether skills should remain free open-source artifacts or become paid products, echoing earlier tensions over VS Code extensions, npm packages and GitHub Actions.

Cole Medin's "Harness Engineering" video, cited by the post, claims the most valuable engineering skill of 2026 is assembling and tuning the harness layer between model and codebase. Karpathy's line from his Sequoia Ascent talk — "you can outsource your thinking, but you can't outsource your understanding" — applies directly: installing a skill outsources execution, but understanding what it assumes about a codebase cannot be delegated.

Why it matters

GitHub Trending is a lagging indicator of where developer attention is heading, and it is currently pointing at a distribution layer for agent capabilities rather than standalone applications. For teams adopting coding agents, the practical decisions are which layer to invest in — harness, design specs, vertical packs or orchestration — and how to vet what they install, because the ecosystem's scale has outrun its quality controls. If the app-store parallel holds, a curation and trust layer is coming. Whether it arrives before a serious supply-chain incident will determine how quickly organizations can safely consume community-built skills.