Open Source Ranking: November 11, 2026
This week, GitHub Trending painted a sharp picture of where the development ecosystem is heading — and the answer comes down to one word: agents. Seven of the ten repositories listed add autonomous AI capabilities to the developer workflow, while the remaining three reveal how the underlying infrastructure (design systems, personal finance, and security) is adapting to this new reality.
The agent tooling ecosystem
The dominant pattern is fragmentation into specialized tools. Rather than a single super-agent that does everything, we're seeing the emergence of modular skills that developers install on demand.
addyosmani/agent-skills (115,804 stars) is the reference implementation. Created by Addy Osmani (ex-Google, ex-Netflix, a historic figure in web development), the repository packages engineering workflows that senior engineers apply intuitively — when to write specs, what to test, how to review, when to ship — as skills that any agent can follow. With 439 commits and support for Claude Code, Gemini CLI, Codex, Antigravity, Cursor and more, it's the closest thing to an "industry standard" for AI-assisted development discipline. It's no coincidence that it accumulated nearly 120,000 stars in less than a year: it solves the hardest problem with coding agents, which is making them think like senior engineers rather than just writing code.
shadcn/improve (10,714 stars) applies the same principle of specialization with a distinct strategy: use the most expensive model only for the cognitive part of the work. It audits the codebase, generates a prioritized improvement plan, and delivers specs in markdown format so cheaper models can execute. The result is a "brain + hands" architecture that dramatically reduces the cost of auditing large codebases. The core idea is simple but powerful: intelligence compounds when you understand project context; execution is cheap because it follows instructions.
vercel-labs/agent-browser (49,776 stars) solves the problem of how coding agents interact with web interfaces. Instead of heavy dependencies like Playwright or Puppeteer, it packages a Rust CLI that captures pages as accessibility trees with compact refs (`@e1`, `@e2`). The agent acts on a ref, re-reads the page, and repeats — all in a single shell command. With support for Chrome via CDP, cloud servers (Browserbase, Browserless, AgentCore) and even Electron apps (VS Code, Slack, Discord, Figma), agent-browser has become the de facto standard for agent-driven browser automation.
homeassistant-ai/ha-mcp (7,208 stars, 640 stars just today) extends the MCP (Model Context Protocol) pattern to the home automation universe. With it, Claude, ChatGPT and other assistants can configure and automate a Home Assistant using natural language. Adoption in the community is already substantial: on the Home Assistant subreddit, the Claude integration receives unanimous praise as "next level" for managing smart homes remotely. ha-mcp shows how home automation was ready to be controlled by agents — it just needed the right protocol.
Tools that make agents possible
Two repositories aren't agents themselves, but they're the infrastructure that makes them viable.
meta/astryx (24,695 stars, 2,842 stars today) is the design system that Meta released in September and has seen exponential growth since. With over 150 accessible UI components, customizable CSS tokens, and dedicated CLI and MCP tooling, Astryx was built to be consumed by both humans and AI agents. It's the first major design system that includes specifications in a format a large language model can read and execute directly — generating themes, scaffolding projects, and producing documentation ready for agent consumption.
openai/codex-security (11,608 stars) is the other side of the same coin: instead of facilitating construction, it helps validate what was built. The CLI and TypeScript SDK that OpenAI released as open source use frontier models to find, validate, and fix security vulnerabilities in repositories. With CI support, batch scans, cost controls, and structured exports, codex-security is the tool that security teams will need as agents write increasingly more production code. The release follows the Hugging Face incident, where frontier models were able to find real vulnerabilities — and codex-security institutionalizes that capability.
Applications beyond the development workflow
The remaining three repositories show that the agent automation wave isn't limited to software development.
ZhuLinsen/daily_stock_analysis (70,069 stars) is a complete multi-market stock analysis system powered by large language models. It collects market data, real-time news, technical indicators, sentiment signals, and automated pushes — all at zero cost for scheduled execution. With 70,000 stars, the project illustrates how AI agents are penetrating highly regulated domains: the promise of automated financial analysis has already attracted hundreds of thousands of users, even with non-trivial reliability risks in real markets.
we-promise/sure (12,789 stars, 1,079 stars today) is an open source alternative to Mint and other personal finance SaaS products. Unlike experimental projects, it's a fully functional application — with web, desktop (macOS), and mobile apps — that runs via Docker and gives users full control over their own financial data. The rapid growth (1,000 stars in a single day) reflects market fatigue with SaaS models that sell financial data: the pitch is simple but powerful — you host it, you control it, you trust the code.
What the collection reveals
Today's GitHub Trending edition tells a coherent story. AI agents are no longer in the prototype phase — they're fragmenting into specialized tools, each solving a concrete step in the workflow: code auditing (improve), browser automation (agent-browser), engineering discipline (agent-skills), infrastructure configuration (ha-mcp). Meta is providing the design layer that agents will consume (Astryx), OpenAI is providing the security validation layer (codex-security), and users outside the dev world are applying the same logic to personal finance (sure) and stock analysis (daily_stock_analysis).
What's clear is that the next wave won't be about agents that "do everything" — it'll be about ecosystems of specialized agents connected via standard protocols (MCP, skills format). The winner won't be the most intelligent model, but the platform that best orchestrates specialization.
Sources: Meta Astryx Blog, InfoQ, GitHub - meta/astryx, GitHub - shadcn/improve, GitHub - addyosmani/agent-skills, GitHub - vercel-labs/agent-browser, GitHub - we-promise/sure, GitHub - homeassistant-ai/ha-mcp, GitHub - openai/codex-security, Trilogy AI Substack, GitHub - ZhuLinsen/daily_stock_analysis, ToSea AI Blog, Reddit - Home Assistant, ChatGPT Learn, OpenAI Developer Community
✓ Independent sources cross-checked and verified before publishing
