GitHub's 10 fastest-growing AI repositories this week: some teach agents to audit code, others give them memory

Someone on X compiled a top 10 of AI repositories rising fast on GitHub this week. The list covers security auditing, design systems, skill packs, browser automation, code maps, agent memory, game development teams, persistent terminals, and quick-start templates. Original post: link

At a glance, these projects share a common direction: for agents to get work done, they first need decent tools and processes. Ordinary users can start with claude-quickstarts and agent-browser; developers should look more at security-audit-skill and Agent Skills.

1. security-audit-skill

Cloudflare breaks code security auditing into six stages: reconnaissance, coverage-driven search, candidate validation, structured output, independent record validation, and target-neutral reporting. Problems found by the agent are classified into confirmed, needs_validation, and rejected, and a new validation agent tries to refute them. The repo also includes a coverage ledger and findings validation scripts. Cloudflare notes that repeated runs can improve coverage; in testing, a single run found roughly half of the total vulnerabilities found across repeated runs. This detail is more useful than the phrase “AI automated auditing.” Security auditing needs traceable evidence; it is unrealistic to expect a one-shot scan to be enough.

https://github.com/cloudflare/security-audit-skill

2. astryx

Meta's open-source design system, used internally for eight years, supports more than 13,000 applications and includes over 150 accessible components. It is based on React 19 and StyleX; themes are overridden through CSS custom properties, and if you do not want to be tied to its styles, you can directly use className with Tailwind or plain CSS. It also emphasizes that humans and AI use the same APIs, documentation, and CLI. Design systems are beginning to treat agents as formal users—this shift is worth noting.

https://github.com/meta/astryx

3. Agent Skills

Anthropic's public skill collection. Each skill is a folder whose core is SKILL.md, plus scripts and resources, letting Claude follow a fixed process for specific tasks. The repo includes skills for creativity, development, enterprise communication, and documents; the production-oriented docx, pdf, pptx, and xlsx skills are also included. Writing processes as files is more stable than repeatedly writing prompts.

https://github.com/anthropics/skills

4. agent-browser

A browser automation tool from Vercel Labs. Agents can open pages, type, and click, taking over web operations. It is suitable for repetitive actions such as form filling and page checks.

https://github.com/vercel-labs/agent-browser

5. graphify

Organizes code or materials into a graph that can track relationships, so agents can query where something is and what it connects to. When finding context in a large repository, this kind of map is closer to how the human brain works than full-text search.

https://github.com/Graphify-Labs/graphify

6. claude-mem

Saves memories of agent conversations for use next time. It solves the problem of having to explain the background from scratch every time a new session starts.

https://github.com/thedotmack/claude-mem

7. Claude-Code-Game-Studios

Configures Claude Code as a game development team, with agents in different roles dividing work and collaborating. The repo name is very direct; its real value is demonstrating how multiple agents can assign roles and hand off work.

https://github.com/Donchitos/Claude-Code-Game-Studios

8. herdr

Provides a runtime environment for coding agents. If the connection drops, the work terminal remains; after reconnecting, you return to the original layout. The pain point of long-running tasks is very concrete: don't start over every time the network drops.

https://github.com/herdrdev/herdr

9. gstack

Garry Tan's collection of Claude Code configurations, with tools set up by role, from business management to quality control. It is useful for comparing against your own agent workflow.

https://github.com/garrytan/gstack

10. claude-quickstarts

A collection of Claude application examples offering templates you can actually run. If you want to try the API or build a prototype, starting from a runnable skeleton is faster than reading docs.

https://github.com/anthropics/claude-quickstarts

Some commenters said they would try security-audit-skill and claude-mem first. Others asked whether the security audit output includes reproduction steps; a “verified” label alone is not enough. That is a reasonable request. How agents prove they did the work correctly after finishing will be a more practical question going forward.

Bookmarks easily pile up, but few things actually run. Among these 10 repos, ordinary users should start with claude-quickstarts; developers should start with security-audit-skill; and those who want to build long-term workflows should look at Agent Skills and claude-mem. Star counts can show popularity, but whether a tool can leave evidence and processes is what determines whether it can enter daily work.