Top 10 Fastest-Growing Open Source Projects on GitHub Last Week — All AI, 52,000+ New Stars

No surprises here: AI still rules. Last week, the ten fastest-growing open source projects on GitHub were all AI-related, adding a combined total of more than 52,000 new stars in just seven days.

The trend is remarkably focused — instead of piling up yet more models, the community is filling in the execution layer that lets AI agents actually get work done: skill packaging, parallel orchestration, context management, code graphs, plus real-world scenarios like office automation, education, and trading.

Below are the ten projects, ranked by new stars from highest to lowest.

1. mattpocock/skills

+10,651 stars — the fastest riser. Veteran engineer Matt Pocock distilled the real engineering practices from his own .agents directory into a composable set of Agent Skills. This is not just another prompt collection: it packages genuine engineering workflows — requirement alignment, TDD, debugging, architecture optimization, and code review — into skill packs that plug directly into coding agents like Claude Code and Codex. The goal is explicit: stop letting AI “vibe code” and make it work in a controlled, engineer-like manner.

Highlights:

• Alignment tools (grill-me, grill-with-docs) to clarify requirements and domain models
• Built-in feedback loops: TDD red-green-refactor, issue diagnosis, static type checking
• Full workflows: breaking requirements into tickets, implementation, code review, architecture improvement
• Shared vocabulary via CONTEXT.md and ADRs to reduce communication noise
• One-line install, pick skills on demand, developers stay in control

GitHub: https://github.com/mattpocock/skills

2. hallmark

+8,948 stars. A design skill built specifically to fight the “template smell” of AI-generated web pages (AI slop), open-sourced by a developer affiliated with Together AI and compatible with Claude Code, Cursor, and Codex. It goes far beyond swapping color palettes: from macro structure selection and theme choice to anti-pattern checks and self-critique, the entire pipeline forces the AI to produce pages that are distinctive rather than default-template clones. If you have ever complained that vibe-coded landing pages all look the same, this project hits exactly that pain point.

Highlights:

• Chooses layouts from multiple macro structures to guarantee structural variety
• Ships 20+ distinctive themes (Distil, Cold Snap, Bubble, Garden, Riso, and more)
• Automatically runs 57 “slop gates” after generation to intercept common AI anti-patterns
• Self-critique mechanism that refuses to output substandard results
• Verbs for audit (scoring), redesign (rebuild with a new fingerprint), and study (extract design DNA from screenshots/URLs)
• Custom themes supported; generates self-contained HTML + CSS

GitHub: https://github.com/Nutlope/hallmark

3. orca

+5,733 stars. Orca positions itself as an Agent Development Environment (ADE) — a dedicated environment for managing a “fleet of parallel coding agents.” It is not tied to any single agent: run Claude Code, Codex, OpenCode, Pi, or any CLI agent simultaneously using your own subscriptions. Each task runs in an isolated git worktree, so nothing interferes, and a unified dashboard lets you compare results and merge the best solution. Desktop, mobile, and even VPS deployments are supported — ideal for anyone who wants multi-agent collaboration as a daily workflow.

Highlights:

• Runs Claude Code, Codex, OpenCode, Pi, and any CLI agent in parallel
• Each agent gets its own git worktree — compare outcomes and merge the best
• Ghostty-grade terminal (WebGL rendering, unlimited splits, persistent scrollback)
• Mobile companion app for live progress, notifications, and direct takeover
• GitHub/Linear integration — open a worktree straight from an Issue or PR
• SSH remote worktrees, Design Mode (click UI elements to feed the agent), CLI scripting

GitHub: https://github.com/stablyai/orca

4. awesome-llm-apps

+5,385 stars. A hand-written, end-to-end-tested open source collection of 100+ runnable AI agents, agent skills, and RAG applications. It gathers real, working examples covering everything from simple starters to production-grade multi-agent teams, voice agents, MCP tools, and memory apps. Supporting Claude, Gemini, GPT, DeepSeek, Llama, Qwen, and other mainstream models, many projects run right after cloning — perfect for finding inspiration, validating ideas, or turning directly into product prototypes.

Highlights:

• Covers Claude, Gemini, GPT, DeepSeek, Llama, Qwen, and other mainstream and open models
• Well-organized categories: Starter Agents, Advanced Agents, Always-on Agents, Multi-agent Teams, Voice Agents, MCP Agents, RAG tutorials, Memory apps, Chat with X, and more
• Most projects run right after cloning; quick skill install via npx skills add
• Complete production-grade examples: multi-step reasoning, tool calling, memory, generative UI
• Weekly template updates — great for ideas or direct product adaptation

GitHub: https://github.com/Shubhamsaboo/awesome-llm-apps

5. code-review-graph

+4,791 stars. A local-first code intelligence graph tool. It uses Tree-sitter to parse an entire codebase into a persistent knowledge graph (functions, calls, imports, inheritance, test coverage, and more), then feeds precise context to Claude Code, Cursor, Codex, and other AI coding tools via MCP. Instead of re-reading the whole repository every time, the AI walks the dependency and change graph to fetch only what is truly relevant — dramatically cutting token consumption. Especially valuable on large projects, and your code never leaves your machine.

Highlights:

• Tree-sitter-built persistent graph of functions, calls, imports, and inheritance
• Blazing-fast incremental updates (only changed files parsed, usually under 2 seconds)
• 30+ languages supported (Python, TS/JS, Go, Rust, Java, C/C++, PHP/Laravel, Swift, Kotlin, etc.)
• Blast-radius analysis, risk-scored PR review, execution-flow detection, community partitioning
• Roughly 82x average reduction in context tokens
• Serves precise context to Claude Code, Cursor, Codex via MCP
• GitHub Action posts risk-score comments automatically; code stays fully local

GitHub: https://github.com/tirth8205/code-review-graph

6. OfficeCLI

+4,047 stars. An office suite built specifically for AI agents. This single-binary, cross-platform command-line tool lets agents read, modify, and generate Word (.docx), Excel (.xlsx), and PowerPoint (.pptx) files — no Microsoft Office installation required. Path-based element addressing, a built-in formula engine, high-fidelity rendering (HTML/PNG), and MCP Server support let agents complete the full “generate → inspect → fix” loop. Great for office automation, report generation, and batch processing.

Highlights:

• Fully local, single self-contained binary, cross-platform (macOS/Linux/Windows)
• Create, read, modify, and analyze .docx/.xlsx/.pptx
• Path-based element addressing (e.g., /slide[1]/shape[1]) for precise operations
• Built-in engine with 350+ Excel formulas (dynamic arrays, finance, statistics)
• Template merging, JSON round-trip, resident in-memory editing, atomic batch operations
• High-fidelity HTML/PNG rendering for the agent “render → inspect → fix” loop
• MCP Server included for direct integration with Claude, Cursor, and others
• Advanced features: charts, SmartArt, animations, conditional formatting, i18n

GitHub: https://github.com/iOfficeAI/OfficeCLI

7. Vibe-Trading

+3,679 stars. An open source project from HKUDS positioned as “your personal trading research agent workspace.” It turns natural-language questions directly into runnable market analysis, strategy generation, backtesting, and reports. It supports multiple markets — China A-shares, Hong Kong stocks, US equities, crypto, futures, and forex — with built-in multi-agent team collaboration, an alpha factor library, Shadow Account behavior diagnostics, and optional broker integration. Feature-packed, but the authors repeatedly warn: verify data quality, backtest bias, and risk controls yourself before trading real money.

Highlights:

• Natural language drives the full pipeline: market research, strategy generation, backtesting, reporting
• Multi-market support: A-shares, Hong Kong, US, crypto, futures, forex
• Built-in multi-agent teams (investment committee, quant, crypto, risk) working together
• Alpha Zoo (large library of preset factors) and multi-engine backtesting
• Shadow Account parses broker logs for behavior diagnostics
• Cross-session persistent memory, self-evolving skills, MCP tool exposure
• Connects to multiple brokers for paper or restricted live trading (verify risk controls yourself)
• CLI, Web UI, and IM channel entry points

GitHub: https://github.com/HKUDS/Vibe-Trading

8. pi

+3,569 stars. Pi is a lightweight, highly extensible AI agent toolkit (agent harness). It provides a unified multi-model API, an agent runtime (tool calling and state management), a terminal UI library, and an interactive coding agent CLI. Its design principle: “minimal core, workflows defined by you” — no forced processes. Developers keep their own habits and swap only the model, tools, or interaction layer. Extensions, Skills, prompt templates, and Themes can be packaged and shared, making it ideal for anyone who wants a fully customized agent experience.

Highlights:

• Unified multi-provider LLM API (OpenAI, Anthropic, Google, 15+ providers)
• Built-in agent loop, tool calling, and state management
• Interactive coding agent CLI with a differential-rendering TUI
• Highly extensible: Extensions, Skills, prompt templates, Themes — all packageable and shareable
• Tree-structured session history with branching, replay, and export
• Container/sandbox permission controls for safe execution
• “Minimal core, your workflow” philosophy — no mandatory processes

GitHub: https://github.com/earendil-works/pi

9. DeepTutor

+2,908 stars. Also from HKUDS, DeepTutor is a self-hostable, agent-native personalized AI tutoring system. It integrates tutoring, problem solving, quiz generation, research, visualization, and mastery paths into a single agent loop, with knowledge bases, books, notebooks, question banks, personas, and memory shared across modes. It supports multiple RAG engines and a three-tier inspectable memory system, and can plug in external agents as sub-agents or partners. Ideal for private knowledge-base teaching, personal learning assistants, or long-term learning progress tracking.

Highlights:

• One agent loop, six modes: Chat, Quiz, Research, Visualize, Solve, Mastery Path
• Knowledge bases, books, notebooks, question banks, personas, and memory share context across modes
• Multiple RAG engines supported (LlamaIndex, PageIndex, GraphRAG, LightRAG, Obsidian)
• Three-tier inspectable memory (traces, summaries, synthesis) plus a memory graph
• External agents as sub-agents or partners; IM channels supported
• Co-Writer collaborative writing, editable “living books,” skill extensions
• Deploy via PyPI, source, or Docker

GitHub: https://github.com/HKUDS/DeepTutor

10. Codex

+2,445 stars. OpenAI’s open source lightweight terminal coding agent. It runs locally, reads and modifies code directly, executes shell commands and tests, and connects to IDE extensions, a desktop app, and cloud tasks. Recent releases focus on configuration migration from Cursor/Claude, audio input with real-time voice, and more stable multi-agent capabilities. A great fit for developers who live in the terminal and want an AI coding agent embedded in their existing workflow.

Highlights:

• Runs locally in the terminal: read/modify code, run shell commands and tests
• Connects to IDE extensions, desktop app, and cloud tasks
• Recent versions strengthen config migration/import from Cursor and Claude
• Audio input and real-time voice interaction
• Stable multi-agent support: configurable sub-agent models, reasoning levels, concurrency, roles
• Approval modes, sandboxing, MCP tool connections
• Sign in with a ChatGPT subscription or API key; easy cross-platform install

GitHub: https://github.com/openai/codex

The Takeaway

These ten projects added 52,156 stars in a single week. The leaderboard makes one thing clear: the community is now seriously competing on the agent execution layer. No matter how powerful the models get, the work ultimately gets done through skills, orchestration, context management, and scenario-specific tools. These projects are pushing AI agents from “usable” to “reliable, controllable, and token-efficient.”