This week the GitHub leaderboard stopped being about models, or even apps. Of the 25 fastest-growing repositories, roughly a third are “skills,” and most of the rest are tools that wrap the coding agents those skills run inside. Even the week’s single biggest gainer fits the shape: odysseus added 1,017 stars — nearly double the next repo — by selling a self-hosted AI workspace, somewhere to own the agent stack rather than rent it. The board reads less like a list of new software than a map of the agent-coding stack feeding on itself.
At least eight of the top 25 either carry “skill” in their name or are a bare instruction file. last30days-skill sits at #2 (+539) and taste-skill at #5 (+447); mattpocock/skills (#10) ships “straight from my .claude directory,” graphify (#11) is an “AI coding assistant skill,” obra/superpowers (#13) is a skills framework, and andrej-karpathy-skills (#15) is, literally, “a single CLAUDE.md file” — with ECC (#14) and Anthropic’s own harness (#8, more below) rounding out the eight. A skill is usually nothing more than a Markdown or shell bundle that reshapes how Claude Code or Codex behaves, which is exactly why they climb so fast: the cost to publish one is a file, not a codebase.
Below the skills sits a second layer aimed at the agents themselves. cc-switch (#12) manages and switches between Claude Code, Codex, and others; CodexPlusPlus (#24) exists only to make Codex nicer to use; codegraph (#7, +340) builds a pre-indexed knowledge graph those agents can query across Claude Code, Codex, Gemini, and Cursor. None of the three is a coding agent — each is scaffolding bolted onto one. When the tooling around a tool trends on its own, the tool has become a platform.
The clearest maturation signal is who showed up. Anthropic itself sits at #8 with defending-code-reference-harness (+328) — skills for threat modeling, scanning, and patching — the model vendor now shipping into the same skills ecosystem its users are building, and doing it on the security side rather than the demo side. Branding is racing ahead too: astrid (#22) bills itself as an “operating system for AI agents,” its JavaScript SDK trailing at #25. The OS label is marketing more than architecture for now, but a vendor and a self-styled agent OS arriving in the same week is what an ecosystem looks like once it stops being a novelty.
A quieter pattern runs underneath: open clones of closed products. open-notebook (#19) is an open-source take on Google’s NotebookLM; open-design (#23) pitches a local-first alternative to Claude’s design tool; and alibaba/open-code-review (#16) puts a major vendor’s name behind an open code-review agent. The lag between a closed AI product shipping and an open replica trending is now measured in weeks — and incumbents are increasingly on the publishing side of it.
Not everything fit the pattern — roboflow/supervision (#21) and a one-click AI video generator (#17) rode in on unrelated demand. But taken together the week marks a shift in the unit of open-source AI distribution: from a model, to an app, to a skill — a behavior patch small enough to live in one file. That low barrier is both the engine behind these velocities and their asterisk. These are rolling seven-day star deltas, so they track attention more than adoption, and “skill” is partly a naming signal here. The direction, though, is hard to miss: the fastest way to ship AI capability this week was to reshape an agent someone else already runs.