Today's GitHub trending page tells a clear story: the community is hardening the scaffolding around AI agents. The projects climbing fastest aren't models themselves — they're the runtimes, permission systems, and tooling that determine what agents can do, where they run, and how safely they operate. That shift from raw capability toward production-grade infrastructure is the defining pattern of the week.
cloudflare/security-audit-skill: Agents That Audit Themselves
Cloudflare open-sourced a coding-agent skill designed for multi-phase security audits, with machine-readable, independently verified findings as the output. It landed 3,155 stars today — by far the day's biggest mover. The concept is notable: rather than a standalone scanner, this is a composable skill that an AI coding agent can invoke as part of a larger workflow, structuring security findings in a format other agents or systems can act on programmatically. As codebases increasingly touch both human and agent authors, having security tooling that speaks the same interface is increasingly important.
cloudflare/security-audit-skill on GitHub Trending
cactus-compute/needle: Inference on the Smallest Hardware
Needle is a lightweight foundation model engineered specifically for constrained hardware — think embedded systems and minimal devices — capable of tool calls and structured extraction in environments where running a standard model is impractical. It gained 234 stars today. Most of the LLM ecosystem is still optimized for GPUs and cloud instances; needle targets the opposite end of the spectrum, where memory and compute budgets are measured in megabytes. For developers building IoT applications or offline-capable agents, a model designed from the ground up for constraints is more useful than a quantized version of something that wasn't.
cactus-compute/needle on GitHub Trending
trycua/cua: Open-Source Fleet Management for Computer-Use Agents
Cua orchestrates computer-use systems through open-source drivers with cross-platform fleet management, designed to support model training at scale. It earned 859 stars. Computer-use agents — those that control a desktop or browser as a user would — have matured considerably, but the infrastructure for running and managing fleets of them remains fragmented. Cua's contribution is treating computer-use agents as a managed resource class, with the kind of coordination tooling that cloud VMs have had for years but agent fleets generally haven't.
addyosmani/agent-skills: A Production Registry for What Agents Can Do
This JavaScript project from Addy Osmani defines a structured, production-grade registry of engineering skills for AI coding agents — effectively a typed catalog of capabilities with validated contracts. It pulled in 556 stars today. The problem it addresses is subtle but real: as agents become collaborators in engineering workflows, the question of what a given agent is actually good at — and how that's expressed in a machine-readable way — becomes a coordination problem. A shared skill registry lets agents compose reliably rather than each project reinventing its own capability model.
addyosmani/agent-skills on GitHub Trending
asciimoo/hister: A Search Engine You Own
Written in Go, hister is a self-hostable personal search engine — crawl your own index, control your own results. It gained 420 stars today. As search quality debates continue and AI-generated content makes ranking harder to trust, the appeal of a search engine whose inputs and outputs you fully control is growing. Hister is lightweight enough to run on modest hardware, which means self-hosted search is now accessible to individual developers rather than just organizations with serious infrastructure budgets.