awesome-ai-pulse-georgia

<p align="center"> <img src="assets/banner.svg" alt="Awesome AI Pulse Georgia" width="100%"/> </p> <h1 align="center">Awesome AI Pulse Georgia</h1> <p align="center"> <b>A curated collection of AI agent frameworks, developer tools, and automation resources</b><br/> Curated by <a href="https://aipulsegeorgia.ge">AI Pulse Georgia</a> </p> <p align="center"> <a href="https://awesome.re"><img src="https://awesome.re/badge.svg" alt="Awesome"/></a> <img src="https://img.shields.io/badge/repos-207-00D0FF?style=flat-square&labelColor=111827&v=34" alt="Repos"/> <img src="https://img.shields.io/badge/categories-9-A949DA?style=flat-square&labelColor=111827" alt="Categories"/> <img src="https://img.shields.io/badge/made_in-Georgia_%F0%9F%87%AC%F0%9F%87%AA-00D0FF?style=flat-square&labelColor=111827" alt="Made in Georgia"/> <a href="https://aipulsegeorgia.ge"><img src="https://img.shields.io/badge/aipulsegeorgia.ge-website-A949DA?style=flat-square&labelColor=111827" alt="Website"/></a> </p> <p align="center"><i>A curated collection of AI agent frameworks, developer tools, and automation resources — by AI Pulse Georgia</i></p> --- ## ⚡ Use Claude Code or Cursor with MCP Server The entire collection is available as an **MCP server** — accessible from within Claude Code, Cursor, Codex, or any MCP-compatible client. Search 207 repos with one command. Instead of opening GitHub and scrolling through READMEs, ask your AI assistant *"What's the best RAG in the AI Pulse collection?"* — and it will return the right repo with a Georgian description. **Install** — in `~/.claude/claude_desktop_config.json` or Cursor MCP settings: ```json { "mcpServers": { "aipulsegeorgia": { "command": "npx", "args": ["-y", "@aipulsegeorgia/mcp-server"] } } } ``` > *Available as an MCP server — query 207 curated AI repos directly from Claude Code, Cursor, or any MCP client.* Full docs: [`mcp/README.md`](./mcp/README.md) · Source: [`mcp/`](./mcp/) ## 🤖 Coding Agents > **Coding Agents & CLI IDEs** — Independent AI-powered coding tools for terminals and editors. If you're a programmer looking for AI assistance with coding, testing, or debugging, this section is for you. | Repository | ⭐ | Description | |---|---|---| | [Claw Code](https://github.com/ultraworkers/claw-code) | 187K | Claw Code is an open-source alternative to Claude Code and other commercial coding agents, written in Rust. Works with Claude, OpenAI, Grok, and any OpenAI-compatible model API. Rust architecture ensures maximum speed and minimal resource consumption. Released on March 31, 2026, and gained 187,000+ stars by April 24. | | [opencode](https://github.com/anomalyco/opencode) | 148K | opencode is a fully open-source coding agent for terminals — a free, self-hosted alternative to Claude Code and Cursor. Written in TypeScript with a beautiful TUI interface. Works with any AI model, including Anthropic, OpenAI, Google Gemini, and Ollama. One of the most popular open-source coding agents since late 2025. | | [Gemini CLI](https://github.com/google-gemini/gemini-cli) | 102K | Gemini CLI is Google's official, open-source AI coding agent for terminals — featuring the Gemini 2.5 Pro model directly in your command line. 1 million token context window means entire large codebases can be loaded at once. Fully free under Apache 2.0 license, supports MCP protocol and Google Workspace integrations. | | [Codex CLI](https://github.com/openai/codex) | 77K | Codex CLI is OpenAI's official, lightweight coding agent for terminals — written in Rust for speed and stability. Runs code in a sandboxed environment with approval-mode flags for precise control. Authorizes directly with ChatGPT accounts (Plus, Pro, Business, Edu, or Enterprise) without separate API keys. | | [OpenHands](https://github.com/All-Hands-AI/OpenHands) | 71K | OpenHands (formerly OpenDevin) is an autonomous AI software engineering tool — creates and edits files, runs shell commands, uses web browsers, and calls APIs without human intervention. Runs in a Docker container with a web UI and CLI. Supports 60+ LLM backends, including Anthropic, OpenAI, Gemini, and Ollama. | | [Cline](https://github.com/cline/cline) | 60K | Cline is an autonomous coding agent embedded in VS Code, requesting permission at every step — cannot change files or run commands without approval. Creates, edits files, runs commands, and uses browsers — all within one IDE. Written in TypeScript, has 5 million+ installations as a VS Code extension, and supports Claude, GPT, Gemini, and 500+ LLMs. | | [omo (ex-oh-my-openagent)](https://github.com/code-yeongyu/oh-my-openagent) | 53K | omo (formerly oh-my-openagent) turns a single AI coding session into a multi-agent system — specialized AIs perform different tasks in parallel within one workflow. Written in TypeScript, works with Claude Code, Codex, opencode, and Cursor. A general agent harness, not specific to Claude Code. | | [Aider](https://github.com/Aider-AI/aider) | 43K | Aider is a terminal AI pair programming tool — one of the pioneers in this field since 2023. Simply ask what you want to change, and Aider opens necessary files, fixes code, and commits to git. Written in Python, fully open-source under MIT license, and supports Claude, GPT, Gemini, DeepSeek, Ollama, and 100+ LLMs. | | [Goose](https://github.com/aaif-goose/goose) | 43K | Goose is an open-source, extensible AI agent. Originally created by Block (Square/Cash App), the project moved to the Agentic AI Foundation (Linux Foundation) in 2026. Written in Rust, supports 15+ AI providers and 70+ MCP extensions. | | [Tabby](https://github.com/TabbyML/tabby) | 33K | Tabby is an AI coding assistant that runs on your own server — a self-hosted, open-source alternative to GitHub Copilot. Offers code auto-completion and chat in VS Code, JetBrains, and other IDEs. Written in Rust with OpenAPI interface, LDAP/SSO, and personalized fine-tuning for corporate use. | | [Continue.dev](https://github.com/continuedev/continue) | 32K | Continue is an open-source AI coding extension for VS Code and JetBrains — supports any AI model, including Claude, GPT, Gemini, Mistral, and local models. Offers chat, autocomplete, edit, and agent modes, with CI-integrated AI checks. Written in TypeScript, fully customizable, and has 30,000+ stars. | | [Void](https://github.com/voideditor/void) | 28K | Void is an open-source fork of VS Code that puts AI at the core of the IDE — a potential alternative to Cursor. Offers tab completion, agent mode, and inline edit. Supports self-hosted LLMs (Ollama, vLLM) for data privacy. **Important**: development paused in 2026 to explore new coding ideas. | | [GitNexus](https://github.com/abhigyanpatwari/GitNexus) | 28K | GitNexus is a browser-based code intelligence tool that works without a server. Loads GitHub repositories or ZIP files and displays interactive knowledge graphs of files, functions, and dependencies. Includes a Graph RAG Agent for natural language question-answering. Written in TypeScript, fully client-side. | | [Taskmaster](https://github.com/eyaltoledano/claude-task-master) | 26K | Taskmaster is an AI-powered project management system — breaks down large tasks into manageable subtasks, prioritizes, tracks progress, and provides context. Integrates with Cursor, Windsurf, Lovable, Roo, and Claude Code via MCP server or CLI. Written in JavaScript. | | [DeepSeek-TUI](https://github.com/Hmbown/DeepSeek-TUI) | 25K | DeepSeek-TUI is an open-source coding agent for terminals, specialized in DeepSeek models — written in Rust for speed and TUI interface. Supports DeepSeek's V3, R1, and Coder models directly in the command line. Offers a more affordable alternative to OpenAI and Anthropic's paid models. | | [OpenClaude](https://github.com/Gitlawb/openclaude) | 23K | OpenClaude is an open-source terminal coding agent that makes 200+ AI models accessible from one CLI interface — including OpenAI, Gemini, DeepSeek, Ollama, Codex, and GitHub Models. Written in TypeScript, compatible with OpenAI API standards. | | [Qwen Code](https://github.com/QwenLM/qwen-code) | 23K | Qwen Code is Alibaba's official, open-source terminal coding agent — written in TypeScript and compatible with OpenAI API standards. Supports Qwen 3 models, particularly effective in code navigation and automation. Competes with Claude Code and Gemini CLI in the Alibaba ecosystem. | | [Roo Code](https://github.com/RooCodeInc/Roo-Code) | 23K | Roo Code is a VS Code extension that gives you a whole team of AI coding agents in your editor. It's a fork of Cline, but with much more features — custom agent modes, parallel multi-agent work, wide MCP support. 500+ LLM providers, 1.5M+ installs. TypeScript, Apache 2.0 license, very active development. | | [Kilocode](https://github.com/kilo-org/kilocode) | 18K | Kilocode is a VS Code AI coding agent, the open-source heart of kilo.ai's agentic engineering platform. Supports 500+ AI models, writes, tests, and deploys code with natural language commands. TypeScript, 1.5M+ installs. A leading coding agent in OpenRouter's list — a direct competitor to Roo Code and Cline. | | [Codex Plugin](https://github.com/openai/codex-plugin-cc) | 15K | Codex Plugin is OpenAI's official Claude Code plugin — allows you to invoke Codex CLI's agent from within Claude Code. For code review and task delegation — combines OpenAI's Codex and Anthropic's Claude Code in one workflow. JavaScript, available since March 2026. A separate repo from openai/codex CLI (which is a terminal agent — this is a Claude Code plugin). | | [cmux](https://github.com/manaflow-ai/cmux) | 15K | cmux is a macOS terminal emulator built on top of Ghostty, specifically optimized for AI coding agents. Vertical tabs, push notifications, and easy management of parallel sessions — if you run multiple AI agents at once (Claude Code + Codex + opencode), this terminal simplifies the process. Written in Swift, macOS-only. | | [free-claude-code](https://github.com/Alishahryar1/free-claude-code) | 11K | free-claude-code is an open-source tool for using Claude Code for free across three channels — in the terminal, as a VS Code extension, or as a Discord bot. Three interfaces in one package: locally writing code with CLI/IDE, team collaboration over Discord. Written in Python, MIT license, 11,000+ stars since January 2026. | | [ml-intern](https://github.com/huggingface/ml-intern) | 4.2K | ml-intern is an **official Hugging Face** open-source AI-ML engineer — an autonomous agent that reads research papers (arXiv), trains ML models, and ships them to Hugging Face Hub. Works in Python, with OpenAI/Anthropic API or local models. **GitHub Trending #1 on April 24, 2026** — +720 stars in one day. For researchers interested in automating ML research — a perpetual ML intern. | | [Open Agents](https://github.com/vercel-labs/open-agents) | 4.1K | Open Agents is an open-source reference template by Vercel Labs for building cloud AI agents. Agents run in a VM sandbox, includes GitHub integration, workflow runtime, and a built-in web interface. Written in TypeScript, easily deployable on Vercel. A reference implementation for background coding agents without a local computer — not a production-ready platform. | | [Pi (pi-mono)](https://github.com/badlogic/pi-mono) | 57K | Pi is a complete coding agent toolkit in one monorepo — interactive CLI agent, unified LLM API (Anthropic, OpenAI, Google, Groq in one interface), agent runtime, and TUI/web UI libraries, including Slack bot and vLLM deployment. Interchangeable components allow you to build your own coding agent from pre-built pieces. TypeScript (93.5%), MIT, v0.77.0 (May 28, 2026), 57,000+ stars. A more "Lego-like", toolkit-oriented alternative to **opencode** and **Cline** — a rapidly growing project by badlogic (Mario Zechner). | ## ⚡ Claude Code Plugins & Skills > **Claude Code Plugins & Skills** — Extensions for Claude Code that add new capabilities to the AI assistant. If you're using Claude Code, these plugins will enhance your workflow. | Repository | ⭐ | Description | |---|---|---| | [Superpowers](https://github.com/obra/superpowers) | 165K | Superpowers is an agentic skills framework for AI coding agents — a software development methodology that teaches AI to "think before you write." Before code is written, a specification, design document, and test plan are created. Supports Claude Code, Codex, Cursor, and Gemini CLI. Shell scripts, MIT license, 165,000+ stars. The ideological predecessor of Spec Kit (on GitHub) and GSD. | | [Everything Claude Code](https://github.com/affaan-m/everything-claude-code) | 165K | Everything Claude Code is a comprehensive performance optimization system for Claude Code, Codex, opencode, and Cursor — skills, memory, security scanning, and research-first workflow in one package. JavaScript, ecc.tools — the most popular Claude Code enhancement collection with 165,000+ stars. The "instincts" system of AI distinguishes it from other skill-packs — the agent adapts to situations with predefined behaviors. | | [Spec Kit](https://github.com/github/spec-kit) | 90K | Spec Kit is GitHub's official Spec-Driven Development toolkit — teaches AI to "plan before code." With `/specify`, `/plan`, `/tasks`, `/implement` slash-commands, AI creates a full specification before moving on to coding. Supports Claude Code, Codex CLI, Gemini CLI, Cursor, Copilot, Windsurf, and Qwen Code. GitHub's official response to Superpowers and GSD-like methodologies — 90,000+ stars. | | [Karpathy Skills](https://github.com/forrestchang/andrej-karpathy-skills) | 80K | Karpathy Skills is a single CLAUDE.md file based on Andrej Karpathy's (former head of OpenAI and Tesla AI) notes on LLM coding pitfalls. It combats AI hallucinations, excessive refactoring, and poor debugging with predefined rules. Created by Forrest Chang in January 2026 — 80,000+ stars, #1 on GitHub Trending. Named after Andrej Karpathy but not officially created by him. | | [UI UX Pro Max](https://github.com/nextlevelbuilder/ui-ux-pro-max-skill) | 69K | UI UX Pro Max adds professional UI/UX design intelligence to AI coding agents — 50 design styles, 161 color palettes, and 25 diagram types in one skill. Supports React, Vue, Flutter, SwiftUI, and other frameworks. Python, 69,000+ stars — significantly improves the visual output of Claude Code, Cursor, and other agents. Build beautiful UIs without a designer. | | [Claude Mem](https://github.com/thedotmack/claude-mem) | 66K | Claude Mem is a memory plugin for Claude Code — automatically records what you do in a Claude coding session, compresses it with the Claude agent SDK, and returns relevant context in the next session. TypeScript, claude-mem.ai. No need to explain again — Claude "remembers" what you were working on. 66,000+ stars. | | [GSD (Get Shit Done)](https://github.com/gsd-build/get-shit-done) | 56K | GSD (Get Shit Done) is a meta-prompting and context engineering system for Claude Code by TÂCHES. It stores AI context structured with special specification files — even between sessions, solving the problem of quality degradation. Supports Claude Code, Codex, Cursor, and Gemini CLI. JavaScript, 56,000+ stars — a practical implementation of spec-driven development, an alternative to Superpowers and Spec Kit. | | [Matt Pocock Skills](https://github.com/mattpocock/skills) | 55K | Matt Pocock's (author of Total TypeScript, one of the most well-known TypeScript instructors in the world) personal `.claude/skills/` directory shared — slogan: *"Skills for Real Engineers. Straight from my .claude directory"*. Shell, MIT license, published in February 2026, already 54,800+ stars. A real reflection of Pocock's daily workflow — TypeScript, React, and modern web development-specific skills he uses in a professional context. If you work seriously with TypeScript or React with Claude Code, this collection is a must-consider. | | [Caveman](https://github.com/JuliusBrussee/caveman) | 44K | Caveman reduces token consumption of Claude Code (and other agents) by 65% — teaches AI to "speak like a caveman," omitting unnecessary words. Quality does not decrease, only verbosity decreases. Lite, Full, and Ultra intensity levels. Python, by Julius Brussee. Supports 40+ coding agents (Claude Code, Codex, Gemini CLI, Cursor). Appeared in April 2026, 44,000+ stars — the most popular skill for reducing AI costs. | | [Career-Ops](https://github.com/santifer/career-ops) | 38K | Career-Ops is an AI job search system built on Claude Code — 14 skill modes, a dashboard written in Go, PDF CV generation, and batch processing for automatic vacancy assessment. Find vacancies, assess with A-F scale, match CV to position. JavaScript, career-ops.org. 38,000+ stars — one of the most complete open-source implementations of using AI for job search. | | [Open Design](https://github.com/nexu-io/open-design) | 36K | Open Design is a fully open-source alternative to Anthropic's paid **Claude Design**, released on April 28, 2026, and gained 36,000+ stars in just 2 weeks (4,000+ fork — organic signal). Contains **19 skills** and **71 brand-grade design systems**: from prompt to HTML/CSS prototype, slides, infographic, carousel, animation, and MP4 export. **BYOK model** — runs with your own Claude/GPT/Gemini/Kimi/GLM/Ollama key, without lock-in. Preview in Sandbox, HTML/PDF/PPTX/MP4 export. TypeScript, Apache 2.0. Works in Claude Code, Codex, Cursor, OpenCode, Qwen Code, Hermes, and Kimi CLI — agent-agnostic. A natural addition to **Huashu Design** and **UI UX Pro Max** — one of the main pillars of the 2026 design-skill stack, one of the fastest-growing entries in the collection. | | [Graphify](https://github.com/safishamsi/graphify) | 33K | Graphify transforms code, documentation, PDFs, images, and videos into a queryable knowledge graph for AI coding agents. Supports Claude Code, Codex, opencode, Cursor, Gemini CLI, and others. Python, graphifylabs.ai. Graph-based questioning method reduces AI tokens by 71 times compared to direct file questioning. 33,000+ stars — rapidly grew in April 2026. | | [RuFlo (ex-Claude Flow)](https://github.com/ruvnet/claude-flow) | 33K | RuFlo v3.5 (formerly **Claude Flow**; repo URL still retains the old name) is an Enterprise AI Orchestration Platform — with a "distributed swarm intelligence" approach, multiple AI agents perform one workflow in parallel. RAG integration, native Claude Code + Codex support, enterprise-level architecture. TypeScript, part of the `ruvnet` (ruv.io) platform. 33,000+ stars — one of the most functional multi-agent orchestration tools in the Claude Code ecosystem. | | [Oh My ClaudeCode](https://github.com/yeachan-heo/oh-my-claudecode) | 30K | Oh My ClaudeCode turns Claude Code into a multi-agent orchestration system — works autonomously in autopilot mode with a 5-phase pipeline (Expansion, Planning, Execution, QA, Validation), and in team mode, N agents perform tasks in parallel. TypeScript, 19 specialized agents, 36 skills. Distinguishes from oh-my-openagent: this is Claude Code-specific orchestration, while the other is a general agent harness. 30,000+ stars. | | [Impeccable](https://github.com/pbakaus/impeccable) | 27.6K | Impeccable is a design-language skill for AI coding agents by **Paul Bakaus** (ex-Google DevRel, Disney Interactive, jQuery UI core contributor — one of the most well-known front-end engineers on the web in the 2010s). Slogan: *"The design language that makes your AI harness better at design"*. Works in Claude Code, Codex, Cursor, OpenCode, and other MCP-compatible clients. JavaScript, **Apache 2.0** license, impeccable.style. Released in November 2025, gained 27,600+ stars (1,500+ fork — organic signal). A natural relative of **Open Design**, **UI UX Pro Max**, and **Huashu Design** — together they form a stacked design-skill configuration that elevates the quality of AI visual output to a professional level. | | [Obsidian Skills](https://github.com/kepano/obsidian-skills) | 26K | Obsidian Skills is a package of agent skills created by Obsidian's CEO, Steph Ango (kepano). It teaches AI coding agents Obsidian's Markdown format, Bases data structure, JSON Canvas, and CLI. Claude Code, Codex, or any MCP-compatible agent can create, search, and edit files in an Obsidian vault. Are you using Obsidian as a knowledge base? This skill turns AI into an expert in working with this base. | | [last30days-skill](https://github.com/mvanhorn/last30days-skill) | 23.8K | last30days-skill is a Claude Code/Codex skill that conducts a thorough study of any topic in one command — **Reddit, X/Twitter, YouTube, Hacker News, Polymarket**, and web are processed in parallel and a **grounded summary** is obtained with sources. In Python, it has accumulated over 23.8K stars since January 2026. **Top-trending Claude Code skills in April 2026** — ideal for competitive analysis, trend research, and product research. | | [Agent Skills](https://github.com/addyosmani/agent-skills) | 21K | Agent Skills is a collection of production-grade engineering skills for AI coding agents by Addy Osmani, head of Chrome DevRel at Google. It teaches 20+ Markdown workflows — spec-driven development, TDD, code review, shipping, and more — to agents, making them senior engineers. In Shell, it has over 21,000 stars. | | [Taste-Skill](https://github.com/leonxlnx/taste-skill) | 17.3K | Taste-Skill is a skill that gives AI coding agents **"taste"** — the slogan is *“stops the AI from generating boring, generic slop”*. It fights against AI's default monotonous, sterile, and "generic AI slop" output with predefined taste curation rules. In Shell, **MIT** licensed, it has over 17,200 stars since February 2026. | | [Claude Plugins Official](https://github.com/anthropics/claude-plugins-official) | 17K | Claude Plugins Official is the official catalog of Anthropic-managed Claude Code plugins — only verified, high-quality plugins. In Python, link to code.claude.com/docs. GitHub org `anthropics` (note: Anthropic's main org is usually called `anthropic`). It has over 17,000 stars. | | [Agent Skills Spec](https://github.com/agentskills/agentskills) | 16K | Agent Skills Spec is a collection of open standards and specifications for giving skills to AI agents — write-once, use-everywhere. It includes a complete specification document, reference SDK, and example skills. In Python, agentskills.io — an independent, community-driven project. | | [Claude Code Game Studios](https://github.com/Donchitos/Claude-Code-Game-Studios) | 16K | Claude Code Game Studios turns Claude Code into a full-fledged game studio — **49 AI agents and 72 workflow skills**, mirroring a real game development studio hierarchy. Agents specialize in roles like Producer, Designer, Artist, Programmer, and QA Tester to create video games together. In Shell, it has accumulated 16K stars since February 2026. | | [Huashu Design](https://github.com/alchaincyf/huashu-design) | 11.7K | Huashu Design (画术 Design) is an HTML-native design skill for Claude Code — high-fidelity prototypes, presentations/slides, animations, and MP4 export in one skill. With **20 design philosophies** and **5-dimensional design review**, it adds professional UI/UX thinking to agents. Agent-agnostic — works with Claude Code, Codex, and other MCP-compatible clients. In HTML, it was released in April 2026 and has over 11,700 stars. | | [n8n Skills](https://github.com/czlonkowski/n8n-skills) | 4.6K | n8n Skills turns Claude Code into an n8n automation expert — 7 complementary skills, 525+ n8n node knowledge, and 2,653+ configuration examples. In Shell, n8n-skills.com. Used with n8n-MCP (also in this collection), it builds even more powerful workflows. | | [Architecture Diagram Generator](https://github.com/Cocoon-AI/architecture-diagram-generator) | 4.3K | Architecture Diagram Generator is a Claude AI skill that creates standalone HTML/SVG diagrams from textual system descriptions — with dark theme, semantic color coding (cyan: frontend, emerald: backend, violet: databases, amber: cloud, rose: security). In HTML, by Cocoon-AI. No design knowledge required — just describe the system in natural language. It has over 4,000 stars. | | [Remotion Skills](https://github.com/remotion-dev/skills) | 2.9K | Remotion Skills are the official agent skills for Remotion (a React video programming framework). It adds Remotion expertise to AI coding agents — composition, rendering, timing, and animation. In TypeScript. If you create data visualizations, reels, or animations from React components, this skill teaches AI all nuances — by the official Remotion team. | | [Headroom](https://github.com/chopratejas/headroom) | 1.5K | Headroom is a context optimization layer for LLM applications — reducing data (files, logs, RAG results) by 70-95% without sacrificing response quality. In Python, headroom-docs.vercel.app. It runs in one command in Claude Code, Codex, and Aider. For significantly lower AI costs and longer work sessions — over 1,500 stars. | | [Claude Code Setup](https://github.com/tornikebolokadze1-cyber/claude-code-setup) | 9 | Claude Code Setup is a production-grade configuration for Claude Code — 17 rules, 7 hooks, 7 templates, and a /setup command for new projects. One installation includes security rules, automated testing, CI/CD integration, and secret protection. In TypeScript, created in Georgia. 9 stars — a new, local project in development. | | [Georgian Payments Skills](https://github.com/erekle1/georgian-payments-skills) | 7 | Georgian Payments Skills adds knowledge of Georgian bank APIs to AI assistants — integrations with TBC Bank (Checkout, TPay) and Bank of Georgia (iPay, Installments, Open Banking). In Python, created in Georgia. For developers working with Georgian payment systems — expert-level local knowledge that doesn't exist elsewhere. 7 stars — a new local resource. | | [CodeGraph](https://github.com/colbymchenry/codegraph) | 33K | CodeGraph builds a pre-indexed semantic knowledge graph for AI coding agents — functions, classes, imports, and call relationships in a local SQLite database for 20+ languages. Instead of reading files from scratch, Claude Code or Cursor query this index directly — resulting in ~22% cheaper and ~50% fewer tool calls. In TypeScript, MIT, v0.9.7 (May 28, 2026), fully local, no external API. GitHub Trending #1 on May 21, 2026, over 33,000 stars — a natural relative in code-intelligence direction. | | [Academic Research Skills](https://github.com/Imbad0202/academic-research-skills) | 24K | Academic Research Skills is a Claude Code skill suite covering the full cycle of scientific research — literature search, paper writing, peer review, and revision. Four interconnected skills, 10-step pipeline, 13+ agent literature exploration, and 12 agents for writing; built-in anti-hallucination and citation-fabrication detection gates. In Python, CC BY-NC 4.0 (non-commercial license), v3.9.2 (May 18, 2026), exports Markdown/DOCX/APA-7 LaTeX. GitHub Trending #3 in 2026, over 23,800 stars — the first specialized skill collection for academic and research workflows. | | [Claude Skills (alirezarezvani)](https://github.com/alirezarezvani/claude-skills) | 17K | Claude Skills is a 338 production-ready Claude Code skill and plugin library for 13 AI coding tools (Claude Code, Codex, Gemini CLI, Cursor, Aider, etc.). Covers beyond engineering to marketing, product, compliance, finance, and C-level advisory — 533 dependency-free Python CLI tool, 25 “POWERFUL-tier” skills, personas (e.g. Startup CTO) and security auditor skills to test. Python, MIT, v2.9.0 (May 28, 2026), 16,500+ stars. **Difference from Matt Pocock Skills and Agent Skills** — focuses on business and multi-domain breadth, not just engineering. | ## 🔌 MCP Integrations > **Model Context Protocol Servers** — Standard "bridges" that connect AI assistants (Claude, ChatGPT, others) to external services — GitHub, Notion, Slack, Figma, Stripe, databases. Install once, works everywhere. | Repository | ⭐ | Description | |---|---|---| | [MCP Servers](https://github.com/modelcontextprotocol/servers) | 84K | MCP Servers is the official mono-repo of Model Context Protocol servers — the foundation of the MCP ecosystem and reference implementation. File system, Git, Slack, memory, browser automation — dozens of built-in servers in one place. 10 language SDKs (TypeScript, Python, Go, Rust, Java, and more). Maintained by Anthropic, Apache 2.0 license. 84,000+ stars — any MCP project starts here. | | [Context7](https://github.com/upstash/context7) | 53K | Context7 is Upstash's MCP server that provides AI coding agents with always-up-to-date, version-specific API documentation. Typically, AI has issues with outdated or fictional API recommendations — Context7 automatically checks which version you're using and inserts accurate documentation into the context. Written in TypeScript, available at context7.com. Works with Claude Code, Cursor, Copilot, and any MCP-compatible agent. 53,000+ stars. | | [GitHub MCP](https://github.com/github/github-mcp-server) | 29K | GitHub MCP is GitHub's official MCP server — making GitHub's full API available to AI assistants: creating issues, reviewing pull requests, monitoring CI/CD, searching repositories, and analyzing code. Written in Go. Automate GitHub operations with natural language commands using Claude, Copilot, Cursor, or any MCP client. 29,000+ stars. | | [FastMCP](https://github.com/jlowin/fastmcp) | 24K | FastMCP is the fastest, Pythonic way to create MCP servers and clients in Python — "FastAPI for MCP". Write a full MCP server with minimal code using decorators. Created by Jeremiah Lowin (Prefect's CEO), heavily influences Anthropic's MCP Python SDK. gofastmcp.com, 24,000+ stars. The go-to choice for starting your own MCP server in Python. | | [Serena](https://github.com/oraios/serena) | 23K | Serena exposes "IDE" features for coding agents as an MCP. Uses Language Server Protocol (LSP) to understand code semantics — symbols, references, definitions directly. AI truly "understands" code, not just text grep. Written in Python, 23,000+ stars. An open alternative to Cursor's paid features, works with Claude Code, Cursor, or any MCP client. | | [n8n-MCP](https://github.com/czlonkowski/n8n-mcp) | 18K | n8n-MCP is an MCP server that turns AI assistants into n8n automation experts — 1,396 n8n component documentation, 2,646 configuration examples, and 2,709 workflow templates. Written in TypeScript, available at n8n-mcp.com. Works with Claude Desktop, Claude Code, Cursor, Windsurf. Together with n8n Skills (in this collection), AI builds any n8n workflow for you. 18,000+ stars. | | [Figma Context MCP](https://github.com/GLips/Figma-Context-MCP) | 14K | Figma Context MCP provides AI coding agents with precise data from Figma design files — colors, sizes, spacing, component hierarchy — structured data, not screenshots. Written in TypeScript, by Framelink.ai. Works with Cursor, Claude Code, or other MCP clients: Figma design → pixel-perfect code on the first try. 14,000+ stars. | | [Code Review Graph](https://github.com/tirth8205/code-review-graph) | 12K | Code Review Graph creates a persistent knowledge graph of your codebase for Claude Code — AI only reads files relevant to the change, not the entire repo. 6.8x less tokens for code reviews, up to 49x for daily coding tasks — significantly changing costs and speed. Written in Python, 22 MCP tools, 19 programming languages. Works with Claude Code, Cursor, Windsurf, and Continue. | | [Claude Context](https://github.com/zilliztech/claude-context) | 10.4K | Claude Context is Zilliz's MCP plugin that gives Claude Code and other coding agents semantic code search across your entire codebase. Instead of loading files fully, it stores code as embeddings in a vector database (Milvus/Zilliz Cloud) and returns only relevant fragments in context on demand — reducing token costs and discovery rounds on large monorepos. Written in TypeScript, MIT license, installable via `@zilliz/claude-context-mcp` npm package. Requires OpenAI API key and Milvus/Zilliz Cloud endpoint/token to work. 10,400+ stars. | | [Exa MCP](https://github.com/exa-labs/exa-mcp-server) | 4.3K | Exa MCP is Exa's (exa.ai) official MCP server — enabling AI assistants with semantic web search and crawling capabilities. Written in TypeScript. Search code examples, research companies, find real-time web information — all in one interface. One-click install in Cursor and VS Code. An alternative to Perplexity MCP and Tavily MCP with Exa's neural search power. 4,000+ stars. | | [Notion MCP](https://github.com/makenotion/notion-mcp-server) | 4.3K | Notion MCP is Notion's official MCP server — giving AI assistants full access to your Notion workspace: reading, creating, updating pages, querying databases. Written in TypeScript, by Notion's makenotion org. Works with Cursor, Claude Code, or any MCP client. Use Notion for project management or as a knowledge base — AI has direct access. 4,000+ stars. | | [Draw.io MCP](https://github.com/jgraph/drawio-mcp) | 3.2K | Draw.io MCP is draw.io's (JGraph Ltd) official MCP server — enabling AI assistants to automatically create network diagrams, UML, flowcharts, and other diagrams in diagrams.net. Written in JavaScript. Natural language to system architecture — draw.io file → open in browser. 4 integration methods, 3,000+ stars. | | [Supabase MCP](https://github.com/supabase-community/supabase-mcp) | 2.6K | Supabase MCP connects AI assistants to Supabase's (open-source Firebase alternative) database, Auth, and Storage. Written in TypeScript, available at supabase.com/mcp. Works with Cursor, Claude Code, and Windsurf: ask in natural language to create tables, query data, or modify your schema — without the Supabase dashboard. 2,600+ stars. | | [Markdownify MCP](https://github.com/zcaceres/markdownify-mcp) | 2.6K | Markdownify MCP converts any type of file to Markdown for AI assistants — PDFs, images, Word, Excel, PowerPoint, YouTube video (transcript), web pages, audio. Written in TypeScript. Feed AI from any format — one server for all conversions. For RAG pipelines and document analysis workflows. 2,600+ stars. | | [Perplexity MCP](https://github.com/perplexityai/modelcontextprotocol) | 2.1K | Perplexity MCP is Perplexity AI's official MCP server — enabling AI assistants with real-time web search, deep research, and reasoning capabilities. Written in TypeScript/Node. **3 tools**: `sonar-pro` (general search+chat), `sonar-deep-research` (comprehensive research), `sonar-reasoning-pro` (complex analysis). One-click install in Cursor, VS Code, Kiro. Perplexity's citation-based answers reduce hallucinations. 2,000+ stars. | | [Claude Peers MCP](https://github.com/louislva/claude-peers-mcp) | 1.9K | Claude Peers MCP is a local MCP server that exposes different Claude Code sessions to each other and enables real-time messaging. Written in TypeScript. If you work on multiple projects in parallel with different Claude Code instances, they coordinate with each other. Technically an MCP server — included in the Plugins section but belongs to MCP integrations. | | [Tavily MCP](https://github.com/tavily-ai/tavily-mcp) | 1.8K | Tavily MCP is a production-ready MCP server for real-time web search, content extraction, sitemap, and crawling — all operations in one server. Written in JavaScript. A hosted version is also available (mcp.tavily.com) without local installation. For AI research agents and automated web monitoring. 1,800+ stars. | | [LinkedIn MCP](https://github.com/stickerdaniel/linkedin-mcp-server) | 1.7K | LinkedIn MCP is an open-source MCP server for LinkedIn — enabling Claude or any MCP-compatible AI assistant to access LinkedIn profiles, companies, jobs, and messages. Written in Python. For job seekers (automated job analysis), recruiters (candidate research), or sales teams (prospecting). 1,600+ stars — unofficial integration, no official LinkedIn API SDK. | | [Stripe MCP](https://github.com/stripe/agent-toolkit) | 1.5K | Stripe Agent Toolkit is Stripe's official toolkit for building AI-powered financial products — including an MCP server, LangChain integration, CrewAI integration, and Vercel AI SDK support in one package. Written in TypeScript, available at docs.stripe.com/agents. AI assistants can search customers, create payments, manage subscriptions. Stripe's official, 1,500 stars — a toolkit focused on MCP. | | [Apify MCP](https://github.com/apify/apify-mcp-server) | 1.1K | Apify MCP is Apify's official MCP server — making 1,000+ ready-to-use web scrapers and automation tools from Apify Store available to AI assistants. Structured data from any website: Instagram, TikTok, Amazon, Google Maps, LinkedIn. In TypeScript, at mcp.apify.com. For social media monitoring, market research, or lead generation — 1,100+ stars. | | [Power BI Modeling MCP](https://github.com/microsoft/powerbi-modeling-mcp) | 677 | Power BI Modeling MCP is Microsoft's official MCP server for semantic modeling in Power BI. AI assistants can run DAX queries, create and edit data models, set up row-level security, and enable multi-language translations — in natural language. Works with Power BI Desktop, Fabric workspace, and Power BI Project files. Official Microsoft release, 677 stars — new but essential for enterprise BI. | | [Sentry MCP](https://github.com/getsentry/sentry-mcp) | 666 | Sentry MCP is Sentry's official MCP server for integrating AI with error tracking. AI assistants have access to Sentry's error reports, stack traces, issue frequency trends, and performance data. In TypeScript, at mcp.sentry.dev. 666 stars — for AI-assisted diagnostics of production bugs. (Note: 1.8K stars mentioned in README is significantly outdated.) | | [E2B MCP](https://github.com/e2b-dev/mcp-server) | 393 | E2B MCP was E2B's MCP server for running code in a cloud sandbox — Python, JavaScript, Bash in isolated environments — for AI agents. In JavaScript. **Note: This repo is archived as of 2026** — no active support. Check e2b.dev's official documentation for new E2B MCP integrations. 393 stars. | | [Metricool MCP](https://github.com/metricool/mcp-metricool) | 36 | Metricool MCP is Metricool's official MCP server for social media management. AI assistants can analyze Instagram, Facebook, Twitter/X, TikTok, and LinkedIn analytics and schedule posts. In Python. 36 stars — new but enables AI automation for social media management for Metricool users. | ## 🕷️ Web Scraping and Browser Automation > **Web Scraping & Browser Automation** — Automatically extract data from websites and control browsers with AI. If you want your AI to "travel" on websites, collect data, or fill out forms, these tools are for you. | Repository | ⭐ | Description | |---|---|---| | [Firecrawl](https://github.com/mendableai/firecrawl) | 111K | Firecrawl is Mendable AI's (YC S23) web scraping API, which transforms any URL into optimized Markdown or JSON for AI. It solves one of the major problems of LLMs - converting "dirty" HTML of web pages into clean, structured formats. Scrape, crawl, search, and extract operations are unified in one API; JavaScript rendering and anti-bot protection page evasion are built-in. Separate Python and TypeScript SDKs are available, with self-hosted or cloud deployment options. With 111K+ stars, it's the most popular tool in this category for providing web data to AI pipelines. | | [Browser Use](https://github.com/browser-use/browser-use) | 89K | Browser Use is a Python library that enables AI agents to use a real browser - navigate websites, fill out forms, click buttons, and extract information just like a human. It solves the problem that LLMs are designed for text-based interfaces, while most of the web is visual. Built on Playwright, with multi-tab support, automatic DOM structure analysis, and integration with OpenAI, Anthropic, and Gemini models. With 89K+ stars and a rapidly growing ecosystem, it becomes the de facto standard for browser automation in AI-native projects. | | [Crawl4AI](https://github.com/unclecode/crawl4ai) | 64K | Crawl4AI is an open-source Python web crawler optimized specifically for LLMs and RAG pipelines. Its main advantage is that, unlike Firecrawl, it's completely free, can be run locally, and has no API limits. JavaScript rendering, semantic chunking, structured data extraction (based on CSS/XPath/LLM), and async crawling are built-in. It's 6 times faster than Scrapy based on benchmarks; built on Python AsyncIO. For RAG systems where local web data collection is needed, this is the most effective choice. | | [Playwright MCP](https://github.com/microsoft/playwright-mcp) | 31K | Playwright MCP is Microsoft's official MCP server that gives Claude, Cursor, and other AI assistants full control over a web browser. The AI reads the browser's accessibility tree - the structure of visual elements - without screenshots, making the work much more accurate and fast. Navigation, click, fill, screenshot, PDF export - all operations are available via the MCP protocol. Written in TypeScript, published on npm. | | [Agent Browser](https://github.com/vercel-labs/agent-browser) | 30K | Agent Browser is Vercel Labs' (creators of Next.js) CLI tool that automates browser interactions for AI agents. Written in Rust, it provides a combination of high speed and low memory consumption. The tool offers a command-line interface for integrating with AI workflows, rather than a visual API. Part of the Vercel ecosystem, released in January 2026. | | [Lightpanda](https://github.com/lightpanda-io/browser) | 29K | Lightpanda is a headless browser written from scratch in Zig, specifically for AI workflows and automation. It's 11 times faster than Chrome and consumes 9 times less RAM, making it an economical alternative for scaling in server environments. Compatible with Puppeteer and Playwright scripts, allowing for easy migration. Zig's low-level control combined with full JavaScript interpretation is used for web scraping tools on loaded servers. | | [Stagehand](https://github.com/browserbase/stagehand) | 22K | Stagehand is Browserbase's (YC W24) TypeScript SDK that combines Playwright with AI's natural language. Instead of fixed selectors, you can give commands in natural language: `page.act("click the login button")` or `page.extract("get the product price")`. This hybrid approach - code + natural language - is more stable than Playwright's pure automation, as LLMs adapt to changes in the web page structure. Written in TypeScript, with a Python SDK also available. | | [Agent-Reach](https://github.com/Panniantong/Agent-Reach) | 18K | Agent-Reach is a Python CLI tool that enables AI agents to read over 10 social and information platforms with a single interface. Twitter/X, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu, and more - all with one standard API, without official API keys (zero API fees). Written in Python, it has accumulated 18K stars since February 2026 - a pace that speaks to its speed. Can be used for AI research, monitoring, and data collection workflows. | | [opencli](https://github.com/jackwener/opencli) | 17K | opencli is a JavaScript universal CLI runtime that transforms any website, Electron app, or local program into a standard terminal command. It uses the AGENT.md format for AI-native integration, which gives AI context about using the tool. Over 70 tools are already supported, and new tools can be added without external APIs. Released in March 2026, it's quickly gaining popularity in the AI CLI ecosystem. | | [Playwright CLI](https://github.com/microsoft/playwright-cli) | 9.2K | Playwright CLI is Microsoft's TypeScript tool for managing Playwright browser tests and scripts from the terminal. It records any action in the browser and automatically generates Playwright code - in JavaScript, TypeScript, Python, or C#. Selector inspection, screenshots, and tracing are built-in. Works without MCP server overhead, requiring only the command line. | | [Firecrawl MCP Server](https://github.com/firecrawl/firecrawl-mcp-server) | 6.1K | Firecrawl MCP Server is Firecrawl's official MCP integration, which adds full web scraping capabilities to Claude Desktop, Cursor, and other MCP-compatible AI assistants. Scrape, crawl, search, deep research - 12+ tools in one server, with automatic retry and rate limiting. Written in JavaScript, requires a Firecrawl API key. After installation in Cursor or Claude, AI can scrape any web page directly from conversation. | | [Browser Harness](https://github.com/browser-use/browser-harness) | 5.9K | Browser Harness is a Python framework from the Browser Use team that enables LLMs to perform any task in a browser with a self-healing mechanism. "Self-healing" means that if the web page structure or CSS selector changes, the system automatically adapts and the workflow is not disrupted. Released in April 2026, it's rapidly evolving within the Browser Use ecosystem. Written in Python, it enhances the robustness of LLM browser automation. | | [Cloudflare MCP](https://github.com/cloudflare/mcp-server-cloudflare) | 3.7K | mcp-server-cloudflare is Cloudflare's official collection of MCP servers that enables AI assistants to fully control the Cloudflare platform. Workers, KV, R2, D1, Browser Rendering, DNS analytics, and more services - over 13 MCP servers in one repository. Written in TypeScript. Cloudflare infrastructure can be managed by Claude or Cursor directly from conversation, without a CLI. | | [Bright Data MCP](https://github.com/brightdata/brightdata-mcp) | 2.3K | Bright Data MCP is Bright Data's official MCP server that enables AI assistants to perform web scraping while evading anti-bot protection. Bright Data's commercial proxy network - with millions of IPs - is the backend of this server, making it possible to access blocked web pages. Written in JavaScript, requires a Bright Data API key (paid service). For organizations where web data is of commercial importance and standard scrapers are blocked, this is a premium solution. | | [rdrr](https://github.com/fkonovalov/rdrr) | 138 | rdrr is a TypeScript CLI tool that transforms any URL into clean Markdown for AI without a headless browser. Web pages, YouTube videos, GitHub issues/PRs, PDFs, and X/Twitter posts - all with one command. Over 20 site-specific extractors are built-in, providing much cleaner results than regular HTML parsing. Can be used as a CLI or library. Released in March 2026, it's a new project with 138 stars so far. | | [CloakBrowser](https://github.com/CloakHQ/CloakBrowser) | 22K | CloakBrowser is a stealth Chromium that works as a **drop-in replacement for Playwright** — it alters fingerprints at the C++ source level before compile (58 patches: canvas, WebGL, audio, GPU, automation signals) instead of using JS patches. It passes reCAPTCHA v3, Cloudflare Turnstile, FingerprintJS, and 30+ detection services; with support for residential proxy, timezone/locale auto-detection, and human-like interaction. Python/JS, MIT (wrapper), latest updated on May 21, 2026, 22,400+ stars, GitHub Trending #7. A **dual-use tool** — the README is authorized for automation and AI agent integration; a natural addition to **stagehand** and **browser-use** in anti-bot contexts. | ## 🧬 AI Agent Frameworks > **Agent Frameworks & Infrastructure** — libraries and tools for building your own AI agents. If you're a programmer and want to build your own AI agent or multi-agent system — these frameworks can help. | Repository | ⭐ | Description | |---|---|---| | [LangChain](https://github.com/langchain-ai/langchain) | 134K | LangChain is a Python and JavaScript framework for building LLM-based applications and AI agents, created by Harrison Chase in October 2022. It offers 700+ integrations, chains, agents, memory, document loaders, and tools primitives in one ecosystem. LangGraph (graph-based orchestration) and LangSmith (monitoring) are part of the LangChain ecosystem. With 134K stars, it's the most widely used AI framework; many other frameworks follow its conventions, although not directly dependent on LangChain. | | [MetaGPT](https://github.com/geekan/MetaGPT) | 67K | MetaGPT is a Python multi-agent framework where AI agents play the roles of Product Manager, Architect, Engineer, and QA engineer to jointly create software products. From one natural language request to PRD, design doc, code, and tests — the entire pipeline is automatically generated. Created by Chinese company DeepWisdom, released in 2023, and quickly gained 67K stars. It represents an alternative approach to CrewAI and AutoGen — based on the role-playing paradigm. | | [AutoGen](https://github.com/microsoft/autogen) | 57K | AutoGen is a Microsoft Research Python framework where multiple AI agents converse and jointly solve complex problems. Conversation patterns, tool use, human-in-the-loop, code execution — all abstractions are built-in. Announced in April 2026 as part of Microsoft Agent Framework 1.0, alongside Semantic Kernel. Licensed under MIT in Python. Deeply integrated with Microsoft Azure for enterprise AI deployments. | | [AI Hedge Fund](https://github.com/virattt/ai-hedge-fund) | 57K | AI Hedge Fund is a Python simulation project where AI agents mimic the investment styles of Warren Buffett, Charlie Munger, and Michael Burry to jointly analyze stock market data. Built on top of LangGraph's multi-agent orchestration — different "analysts" (fundamental, technical, sentiment, risk) process data in parallel and combine results. Not a real trading system — for educational purposes only. One of the best real-world examples of LangGraph's production-ready usage. | | [CrewAI](https://github.com/crewaiinc/crewai) | 49K | CrewAI is a Python framework where AI agents are assigned different "roles" (researcher, writer, reviewer) and work together as a team to accomplish tasks. The role-playing paradigm distinguishes it from other frameworks — agents are configured with backstory, goals, and specific tools. Licensed in Python, used by 100,000+ developers in production, with 12+ million tasks completed daily on CrewAI. The easiest way to quickly assemble multi-agent workflows. | | [nanobot](https://github.com/HKUDS/nanobot) | 40K | nanobot is an ultra-lightweight Python AI agent from HKUDS (Hong Kong University of Science and Technology), with a total codebase not exceeding 4,000 lines. Unlike complex frameworks, nanobot's core is transparent — easy to understand, modify, and deploy. Works on Telegram, Discord, Slack, WhatsApp, and 5+ other platforms with minimal configuration. Part of the HKUDS OpenSpace ecosystem. Quickly gained 40K stars due to its simplicity and efficiency combination. | | [agno](https://github.com/agno-agi/agno) | 39K | agno (formerly phidata) is a lightweight Python production framework for AI agents, rebranded in 2025. It unites agents, multi-agent teams, workflows, memory, tools, knowledge bases, and monitoring in one minimal dependencies package. Unlike other frameworks, agno's core is simple and understandable — no excessive abstractions are needed for production deployment. 39K stars; an effective framework for small teams creating production-ready agents without LangChain's complexity. | | [DSPy](https://github.com/stanfordnlp/dspy) | 34K | DSPy is a Stanford NLP Python framework that changes the paradigm of working with LLMs: instead of manually writing prompts, it writes programmatic structures (signatures, modules, optimizers). DSPy automatically optimizes prompts and few-shot examples according to a given metric — a compiler for prompts. Created under Omar Khattab's leadership, supported by NeurIPS and EMNLP papers. A key instrument for those writing LLM pipelines and replacing manual prompt engineering work with minimal code. | | [LangGraph](https://github.com/langchain-ai/langgraph) | 30K | LangGraph is a LangChain AI Python framework where AI agents' logic is laid out in a directed graph (nodes + edges). This allows for complex scenarios — cycles, branches, intermediate state storage — beyond LangChain's linear chains. Built-in checkpointing, human-in-the-loop, and streaming are included. Used in production by Uber, LinkedIn, Klarna. LangGraph has become the production standard for long-running, stateful AI workflows — customer support, code review, research. | | [Semantic Kernel](https://github.com/microsoft/semantic-kernel) | 27K | Semantic Kernel is a Microsoft SDK for integrating AI into enterprise applications, in C#, Python, and Java. The plugin-based architecture allows for chained AI function calls; deeply integrated with Azure OpenAI, Azure AI Foundry, and Microsoft Copilot. Part of Microsoft Agent Framework 1.0 (April 2026) alongside AutoGen. The most natural way for enterprise teams working in the Microsoft ecosystem (.NET, Azure, Office 365) to integrate AI. | | [smolagents](https://github.com/huggingface/smolagents) | 26K | smolagents is a Hugging Face minimalist Python library for AI agents, based on a "code-first" approach — agents write Python code instead of JSON to perform actions. The library's core is under 2,000 lines of code, making it particularly understandable and modifiable. Over 100 LLM models are supported — from Hugging Face, OpenAI, Anthropic, Ollama, and others. ToolCallingAgent (JSON tool calls) and CodeAgent (Python code) are the two main interfaces. For those who want simple, transparent agent logic, smolagents is an optimal starting point. | | [OpenAI Agents SDK](https://github.com/openai/openai-agents-python) | 24K | OpenAI Agents SDK is an official OpenAI Python framework for multi-agent workflows. Agents, handoffs, guardrails, tracing, human-in-the-loop, voice agents — all abstractions are built-in with minimal code growth. A production-ready successor to Swarm (OpenAI's earlier educational framework). Over 100 LLM models are supported, not just OpenAI models. Rapid development from March 2025 to v0.14; a direct competitor to CrewAI and AutoGen from OpenAI. | | [Vercel AI SDK](https://github.com/vercel/ai) | 23K | Vercel AI SDK is a Vercel (Next.js creator) TypeScript toolkit for quickly building full-stack AI applications. Streaming responses, tool calling, structured outputs, multi-step agents — all abstractions are built-in, optimized for React/Next.js. OpenAI, Anthropic, Google, Mistral, Ollama, and other providers are accessible via a unified API. Over 20 million downloads per month on npm. A de facto standard for building AI web applications with Next.js and React. | | [Mastra](https://github.com/mastra-ai/mastra) | 23K | Mastra is a TypeScript AI framework from the Gatsby team, created in YC W25 batch, and received $13M in funding. Agents, workflows, RAG, evaluations, and integrations are united in one toolkit. In TypeScript — following the Zod-based typing philosophy of Vercel AI SDK. Distinguishes itself from other JS/TS frameworks with a built-in evaluation suite (measuring agent quality) and workflow engine. A good alternative for Node.js/TypeScript teams who want to bring AI into production. | | [Swarm](https://github.com/openai/swarm) | 21K | Swarm is an OpenAI Solution Team educational Python framework that reflects the basic concepts of multi-agent orchestration — agents and handoffs — with minimal code. Not intended for production (official GitHub status indicates this), but rather an educational resource to understand the concepts behind OpenAI Agents SDK (openai-agents-python). The simplicity philosophy of Swarm laid the foundation for OpenAI's new official Agents SDK. For those who want to quickly grasp the logic of multi-agent systems — this is the shortest educational path. | | [Archon](https://github.com/coleam00/Archon) | 19K | Archon is an open-source tool in TypeScript that makes AI coding workflows deterministic and reproducible — with YAML harness definitions. The idea is simple: yes to AI in coding, but following fixed phases (planning, implementation, validation, review), where each step is auditable and rollbackable. Especially effective for integration with CI/CD pipelines — where AI code is brought into production — it is a practical implementation of the concept of "Dockerfile for AI". | | [Pydantic AI](https://github.com/pydantic/pydantic-ai) | 16K | Pydantic AI is a Python framework for AI agents from the Pydantic team, placing type safety and validation at the center of the structure. Many Pydantic models (OpenAI SDK, Anthropic SDK, LangChain) are already built on Pydantic — Pydantic AI is a natural continuation of this ecosystem. Dependency injection, structured outputs, multi-turn conversations, built-in monitoring (Pydantic Logfire) are embedded. Using Python's type system fully simplifies agent validation and debugging. | | [OpenSandbox](https://github.com/alibaba/OpenSandbox) | 10K | OpenSandbox is a Python sandbox runtime from Alibaba, serving as a secure environment for AI agents to execute code. "Sandbox" means that AI code runs in an isolated container — access to the host system, network abuse, or unwanted file changes are excluded. Python, Java, JavaScript, C# — four major languages are supported. Open source, with MIT license; it is an open-source variant of Alibaba Cloud's commercial sandbox service. | | [OpenSpace](https://github.com/HKUDS/OpenSpace) | 5.7K | OpenSpace is a Python framework from HKUDS for self-development of AI agents — with AUTO-FIX, AUTO-IMPROVE, AUTO-LEARN mechanisms. Agents share "skills" learned from new situations with each other through a distributed skill library. According to benchmarks, it shows a 4.2 times higher result than the baseline with 46% lower costs. It is part of the HKUDS nanobot and DeepTutor ecosystem. Since March 2026, 5.6K stars — the growth rate in research AI directions is impressive. | | [Microsoft Agent Framework](https://github.com/microsoft/agent-framework) | 11K | Microsoft Agent Framework is Microsoft's official multi-language framework for production-grade AI agents and multi-agent workflows — a unified successor to **AutoGen** and **Semantic Kernel** (both in this collection). Python + .NET/C#, MIT, python-1.7.0 (May 28, 2026). Sequential/concurrent/handoff/group orchestration patterns, built-in OpenTelemetry observability, YAML-declarative agents, middleware, and interactive DevUI. 10,900+ stars — for those who used AutoGen or Semantic Kernel, this enterprise multi-agent stack is the official migration target. | ## 💼 Ready-Made AI Agents for Business > **Ready-Made AI Agents & Platforms** — Pre-built AI platforms that work without writing code. Are you a business owner, marketer, or manager? These tools are for you — visually assemble AI bots, automation, and search systems. | Repository | ⭐ | Description | |---|---|---| | [OpenClaw](https://github.com/openclaw/openclaw) | 362K | OpenClaw is a personal AI assistant in TypeScript that works on any operating system and platform. Supports 20+ channels, including WhatsApp, Telegram, Slack, Discord, and iMessage, with a single AI backend. Features voice interface, Canvas (visual workspace), and real-time internet search. Suitable for individual users or small teams to manage all communication channels with one AI. | | [n8n](https://github.com/n8n-io/n8n) | 185K | n8n is a fair-code automation platform in TypeScript with 400+ integrations, featuring a visual no-code builder that works with custom JavaScript/Python code. Supports Gmail, Slack, Airtable, and major cloud services and AI providers as nodes. Offers self-hosted or cloud deployment with a fair-code license allowing commercial use. A popular alternative to Zapier with 185K stars. | | [Langflow](https://github.com/langflow-ai/langflow) | 147K | Langflow is a Python visual platform to build AI agents and workflows with a drag-and-drop interface — no code required. LangChain components are available as graphical blocks: LLMs, vector stores, agents, and tools are connected. Supports MCP, 100+ integrations. Acquired by DataStax, used by 20,000+ organizations in production. Ideal for those who prefer the Python stack and need a visual AI builder. | | [Dify](https://github.com/langgenius/dify) | 138K | Dify is a production-ready platform to build AI agents, chatbots, and RAG-based systems with a visual interface. Supports chatbots, workflows, knowledge bases, and text generators — four primary application types configured from the GUI. Features Google Search, DALL-E, and 100+ tools; offers self-hosted or Dify Cloud deployment. A strong solution for companies to bring AI products into production without significant code investment. | | [Open WebUI](https://github.com/open-webui/open-webui) | 133K | Open WebUI is a ChatGPT-style AI interface that runs on your server or computer — all data is local, no cloud usage. Supports Ollama, OpenAI API, Claude, Gemini, and dozens of other providers. Features RAG (reading your documents), voice calls, agent builder, and multi-user auth — all in one control panel. | | [Hermes-agent](https://github.com/nousresearch/hermes-agent) | 119K | Hermes Agent is a Python AI agent that learns from experience and retains memory between sessions — "grows" with the user. Works on Telegram, Discord, Slack, WhatsApp, and automated scheduling. One of the fastest-growing projects in this collection with 119K+ stars. | | [gpt4all](https://github.com/nomic-ai/gpt4all) | 77K | GPT4All is a C++ tool by Nomic AI to run local LLMs on any device — Mac, Windows, Linux, without GPU. Offers desktop GUI, Python SDK, and REST API in one package. **⚠️ Active development paused**: last GitHub commit in May 2025, Nomic AI shifted focus to other products. Still suitable for non-technical users needing a GUI. | | [LobeChat](https://github.com/lobehub/lobe-chat) | 75K | LobeChat is an open-source ChatGPT alternative in TypeScript/Next.js, allowing any LLM server — OpenAI, Claude, Gemini, Ollama, Mistral — in one interface. Features multi-agent collaboration, plugin marketplace, voice chat, document analysis, and knowledge base — all in self-hosted or Lobe Cloud. | | [AnythingLLM](https://github.com/Mintplex-Labs/anything-llm) | 58K | AnythingLLM is a JavaScript all-in-one AI platform — RAG, agents, chat, multi-user auth, and 20+ LLM providers in one desktop or server application. Notably, the desktop version requires no complex configuration and keeps all data local. | | [Paperclip](https://github.com/paperclipai/paperclip) | 58K | Paperclip is a TypeScript open-source AI orchestration platform for automating business processes — aiming for "zero-human companies". Assembles teams of AI agents, features budget limits, audit log, and progress tracking with a management panel. | | [MiroFish](https://github.com/666ghj/MiroFish) | 57K | MiroFish is a Python "swarm intelligence" simulation engine where thousands of AI agents with individual personalities and memories simulate markets or systems for prediction. Each agent has a different behavior model, and collective decisions emerge without a central "brain". | | [Flowise](https://github.com/FlowiseAI/Flowise) | 52K | Flowise is a TypeScript drag-and-drop visual AI builder connecting LLMs, agents, vector stores, and tools as graphical blocks. Built on LangChain.js components — JavaScript/Node.js stack. A parallel to Langflow (Python) with 100+ integrations. | | [Jan](https://github.com/janhq/jan) | 42K | Jan is a TypeScript open-source desktop AI assistant for Mac/Windows/Linux, working 100% offline — no internet or cloud required. Features Llama, Mistral, Phi, Qwen models downloadable with one click. Cortex backend for speed and resource efficiency. | | [LibreChat](https://github.com/danny-avila/LibreChat) | 35K | LibreChat is a TypeScript self-hosted ChatGPT clone that integrates all frontier LLMs — Claude, GPT-5, o3, DeepSeek, Gemini, Mistral, Groq, Azure OpenAI, Vertex AI, OpenRouter — in one interface. Features DALL-E-3, Code Interpreter, message search, MCP, OpenAPI Actions, and multi-user auth. | | [Khoj](https://github.com/khoj-ai/khoj) | 34K | Khoj is a Python self-hostable "AI second brain" — a personal AI assistant that searches the web, your documents, and saved information simultaneously. Features custom agents, scheduled automations, deep research mode, and Notion, iMessage integration. | | [Vane (ex-Perplexica)](https://github.com/ItzCrazyKns/Perplexica) | 33K | Vane (formerly **Perplexica**; repo URL still retains the old name, while the repo's current description already reads „Vane is an AI-powered answering engine") — TypeScript open-source AI search engine, a self-hosted alternative to Perplexity AI. Searches the internet, processes results, and provides answers with sources, typically much more contextually than regular browser search. Works with Ollama local models. 33K stars; for web research without leaving your data — the closest self-hosted analog to Perplexity. | | [AstrBot](https://github.com/AstrBotDevs/AstrBot) | 30K | AstrBot — Python AI Agent Assistant that positions itself as **an open alternative to OpenClaw**. Native integration with many IM platforms (Telegram, Discord, QQ, WeChat, Lark, Slack), plugin system, extensive LLM provider support — a single instance unifies the entire communication landscape. AGPL-3.0 license, active since 2022, 30K+ stars. Moved from the Chinese world of AI agent ecosystems and is now one of the most feature-complete OpenClaw-style tools. | | [openhuman](https://github.com/tinyhumansai/openhuman) | 24.4K | openhuman — tinyhumans.ai's Rust-written personal AI „super intelligence" that works entirely locally — slogan: *„Your Personal AI super intelligence. Private, Simple and extremely powerful."* Rust's low-level control ensures maximum speed and minimal resource consumption — compared to other local-AI tools (gpt4all, Jan, AnythingLLM, LobeChat all in JS/TS/Python), this difference is noticeable on desktop-class devices. **GPL-3.0** license, tinyhumans.ai/openhuman, released in February 2026 and gained 24,000+ stars in just 3 months — one of the fastest-growing new entrants in the local personal-AI category. A natural alternative to Jan, gpt4all, and AnythingLLM for those who prioritize Rust's speed and GPL's open solidity over JavaScript-stack convenience. | | [Activepieces](https://github.com/activepieces/activepieces) | 21K | Activepieces is a TypeScript AI-first workflow automation platform — an open alternative to n8n, with 400+ MCP Server native support. AI agents, MCP integrations, and visual workflow builder are combined in one package. Self-hosted or cloud, MIT license. Differs from n8n with an MCP-first approach — if you want AI agents & MCP ecosystem with visual automation, Activepieces is a specific solution for this combination. | | [DeepTutor](https://github.com/HKUDS/DeepTutor) | 21K | DeepTutor is a Python AI teaching assistant by HKUDS, specifically designed for students and researchers. Chat, Deep Solve, Quiz Generation, Deep Research, and Math Animator — 5 modules in one platform. RAG (include your own teaching materials), web search, scientific paper search, and code execution are built-in. Create TutorBots with sustainable memory and „personality" — tailored learning experience for each student. Part of the HKUDS ecosystem (nanobot, OpenSpace). | | [Multica](https://github.com/multica-ai/multica) | 20K | Multica is a TypeScript open-source managed agents platform that turns AI coding agents into team members. Assign tasks, monitor progress, accumulate experience (compounding skills) — with each cycle, AI performs better on specific tasks for this team. Released in January 2026, quickly grew to 20K stars. For technical teams looking for an instrument to manage Claude Code, Codex, or other coding agents. | | [NemoClaw](https://github.com/NVIDIA/NemoClaw) | 20K | NemoClaw — **NVIDIA's official** open-source tool that runs OpenClaw more securely in NVIDIA OpenShell environment with managed inference. Corporate policy, audit log, sandbox, and GPU-managed access — all OpenClaw capabilities with enterprise-grade security. TypeScript, Apache 2.0 license, released in March 2026. For organizations that want the value of OpenClaw without deploying it on employees' machines — in a datacenter, under full IT control. | | [Claude for Financial Services](https://github.com/anthropics/financial-services) | 18.6K | Claude for Financial Services — **Anthropic's official** reference repo for financial services workflows: investment banking, equity research, private equity, wealth management. Ready-made Claude agents, skills, slash commands, and MCP data connectors — all in self-contained Claude plugins that deploy with **Claude Cowork** or **Claude Managed Agents API**. Practical use cases: preparing pitch decks, valuation models, earnings reviews, compliance tasks — all outputs staged for human review (human-in-the-loop by design). Python, Apache 2.0 license, released in February 2026 and gained 18,600+ stars. For financial industry IT administrators and analysts — one of the first official Anthropic vertical-specific repos. **Fincept Terminal** (also in this collection — an open-source Bloomberg alternative) together form a complete financial AI workflow stack. | | [QwenPaw](https://github.com/agentscope-ai/QwenPaw) | 16K | QwenPaw — Alibaba's AgentScope team's Python personal AI assistant that runs locally or in the cloud and is accessible through multiple chat applications. Easily extensible capabilities — a wide ecosystem of plugins and combining with Qwen models. Apache 2.0 license, released in February 2026 and gained 16K stars. A Chinese ecosystem's response to OpenClaw, especially for those working with Qwen models. | | [Fincept Terminal](https://github.com/Fincept-Corporation/FinceptTerminal) | 13K | Fincept Terminal is an open-source financial terminal in Python, a Bloomberg alternative. 100+ data connectors (Yahoo Finance, FRED, IMF, World Bank, Polygon) and 20+ AI investor-personas across multiple LLM providers (OpenAI, Anthropic, Gemini, Groq, Ollama) are integrated. DCF models, portfolio optimization, VaR/Sharpe metrics, crypto WebSocket, and algo-trading are built-in. Python (README mistakenly mentioned C++20/Qt6). For individual investors and small investment teams that need institutional-grade tools at a non-institutional price. | | [Rowboat](https://github.com/rowboatlabs/rowboat) | 13K | Rowboat is a TypeScript open-source AI „colleague" with sustainable memory — AI that retains work context across sessions. Built on OpenAI API, self-hosted, with optional cloud. 13K stars since January 2025; a new entrant in this category. For small teams that need a long-term AI assistant that remembers their work context, solutions, and tasks. | | [Feynman](https://github.com/getcompanion-ai/feynman) | 5.8K | Feynman is an open-source AI research agent with 4 specialized sub-agents (Researcher, Reviewer, Writer, Verifier) working in parallel throughout the scientific investigation process. `/deepresearch`, `/lit`, `/audit`, `/replicate`, `/compare` slash commands are available for each workflow. Each conclusion directly points to the source — paper, doc, or repo — to minimize hallucination risk. TypeScript, works with local models (Ollama, LM Studio) or cloud; macOS/Linux/Windows installer available. | | [Mercury Agent](https://github.com/cosmicstack-labs/mercury-agent) | 1.5K | Mercury Agent — TypeScript open-source personal AI agent that runs 24/7 and interacts via CLI or Telegram. Designed with „soul-driven" principles — has a refined permission-hardened tool system and token budget (i.e., cost control). Accessible from multiple channels simultaneously (CLI + Telegram bot on one instance). MIT license, released on April 20, 2026, and quickly grew. For those seeking a personal AI assistant that proactively notifies you of significant events on Telegram instead of constantly switching to a browser. | | [agentic-inbox](https://github.com/cloudflare/agentic-inbox) | 1.3K | agentic-inbox is **Cloudflare's official** self-hosted email client with a built-in AI agent that runs entirely on Cloudflare Workers (edge network). AI automatically sorts emails, writes responses, and prioritizes them — a „self-managing" version of your inbox. TypeScript, released on April 10, 2026. A privacy-first open alternative to Gmail's AI features — your domain, your data, Cloudflare's speed. For businesses that already store domain email on Cloudflare. | | [TradingAgents](https://github.com/TauricResearch/TradingAgents) | 81K | TradingAgents is a multi-agent trading framework that mimics the structure of a real trading firm — fundamental analyst, sentiment expert, technical analyst, trader, and risk manager agents discuss market conditions and make decisions together. It supports multiple LLM providers (OpenAI, Google, Anthropic, DeepSeek), has an interactive CLI, backtesting, paper trading, Docker, and checkpoint recovery. Python, Apache-2.0, v0.2.5 (May 11, 2026), 80,000+ stars — Based on TauricResearch's arXiv paper (research tool, not investment advice). Related to **ai-hedge-fund** and **FinceptTerminal**, but the largest open-world financial AI agent project. | ## 🧠 Memory and RAG > **Memory & RAG** — systems that give AI "memory" and the ability to understand documents. RAG (Retrieval-Augmented Generation) means that AI searches your documents before answering — providing accurate answers, not invented ones. | Repository | ⭐ | Description | |---|---|---| | [RAGFlow](https://github.com/infiniflow/ragflow) | 78K | RAGFlow is an open-source RAG engine that enables AI to find accurate information in your documents. It solves the problem common to all AI chatbots: models often invent answers — RAGFlow searches your files and then answers. Written in Python, supports PDF, Word, Excel, PowerPoint, HTML formats, and includes a built-in agent workflow. Foundation of InfiniFlow's commercial platform — companies use it to build knowledge bases and AI-powered support systems. | | [mem0](https://github.com/mem0ai/mem0) | 53K | mem0 is a universal memory layer for AI agents and chatbots. It solves the fundamental problem: standard LLMs start each conversation from scratch — mem0 stores user preferences, past interactions, and context. Python SDK, REST API, and managed cloud service available. Integrates with LangChain, LlamaIndex, and OpenAI Agents. Used by Perplexity, Pika, Wordware, and other products. | | [MemPalace](https://github.com/milla-jovovich/mempalace) | 49K | MemPalace is an open-source AI memory system that receives the highest ratings on benchmarks compared to similar free solutions. Written in Python, for storing long-term memory and context for AI agents. Completely free and self-hosted, no API keys required. Appeared in April 2026 and quickly gained traction in the AI agent builders community. | | [LlamaIndex](https://github.com/run-llama/llama_index) | 48K | LlamaIndex is a Python framework that prepares AI to work with any data source — files, databases, APIs, web pages. Primarily used for building RAG pipelines: from document indexing to search and response generation in one framework. Over 300 integrations — OpenAI, Anthropic, Hugging Face, Pinecone, Weaviate, PostgreSQL. Developed into an OCR and document agent platform since 2024. Written in Python, Apache 2.0 license, ready for production use. | | [Milvus](https://github.com/milvus-io/milvus) | 43K | Milvus is a cloud-native vector database optimized for storing and searching billions of embeddings. Standard choice for vector store in RAG systems, recommendation engines, and semantic search applications. Written in Go and C++, with a distributed architecture, deployable on Kubernetes. LF AI Foundation project, supported by Zilliz. Used by NVIDIA, Roblox, Shopee, and Salesforce in production environments. | | [LightRAG](https://github.com/hkuds/lightrag) | 34K | LightRAG is a fast RAG system that combines traditional vector similarity search with knowledge graph analysis. Standard RAG searches for words — LightRAG analyzes connections between concepts, providing more accurate answers to complex questions. Written in Python, developed by Hong Kong University's Data Science lab, backed by a scientific paper published at EMNLP 2025. Works with Neo4j, NetworkX, or a built-in graph backend. | | [Qdrant](https://github.com/qdrant/qdrant) | 30K | Qdrant is a high-performance vector database written in Rust, designed for next-generation semantic search and RAG systems. Distinguished by speed, small memory footprint, and thread safety due to Rust. Features payload filtering, quantization, sparse vector support, and horizontal scaling. Easily deployable locally with Docker, managed Qdrant Cloud also available. Python, TypeScript, Go, Rust SDKs. | | [Chroma](https://github.com/chroma-core/chroma) | 27K | Chroma is an open-source embedding database for AI, commonly used for building semantic search and RAG pipelines. Default vector store for LangChain and LlamaIndex — making it a great introduction to working with embeddings. Runs in embedded mode on SQLite, as a standalone server, or on Chroma Cloud. A 5-line Python RAG system can be set up. Written with a Rust core, YC W23 company, Series A funded. | | [Graphiti](https://github.com/getzep/graphiti) | 25K | Graphiti is a real-time knowledge graph engine for AI agents, developed by Zep AI (YC W24). Unlike standard vector-based RAG, Graphiti stores factual context over time — knowing "when" information changed. Bi-temporal model stores event time and ingestion time separately, and LLM extracts entities and relations from text automatically. Runs on Neo4j or FalkorDB backend, with native MCP integration in Claude Desktop, Cursor, and Cline. | | [Letta](https://github.com/letta-ai/letta) | 22K | Letta is a platform for stateful AI agents, built on the former MemGPT project. Unlike standard AI chatbots, Letta agents have long-term memory: working memory (current context) and archival memory (long-term storage) — weeks, months. Founded by a team of UC Berkeley researchers. OpenAI API-compatible REST API, self-hosted deploy. For long-running agents where memory is vital. | | [RAG-Anything](https://github.com/HKUDS/RAG-Anything) | 18K | RAG-Anything is an all-in-one multimodal RAG framework developed by Hong Kong University's Data Science lab, creators of LightRAG. Standard RAG systems only handle text — RAG-Anything combines text, images, tables, diagrams, and other formats in one pipeline. Written in Python, implementation of the scientific paper "RAG-Anything: All-in-One RAG Framework". Extends LightRAG's graph-based reasoning to multimodal input. | | [gbrain](https://github.com/garrytan/gbrain) | 10K | gbrain is a TypeScript-written AI agent's personal "brain" that aggregates and stores email, calendar, meetings, and ideas in one place. Hybrid search (vector + keyword) ensures fast and accurate retrieval. MCP server for Claude and other LLMs is built-in. Runs locally with PGLite or Supabase. Appeared in April 2026, quickly gained traction in AI personal assistant builders. | | [Hindsight](https://github.com/vectorize-io/hindsight) | 10K | Hindsight is an AI agent memory system from Vectorize that learns and improves over time. Unlike standard RAG systems, Hindsight identifies patterns in past interactions and makes subsequent responses more relevant. Written in Python, backed by a scientific paper published on arXiv. For long-running agents where memory quality needs to improve over time. | | [Crawl4AI RAG](https://github.com/coleam00/mcp-crawl4ai-rag) | 2.1K | mcp-crawl4ai-rag is an MCP server that combines web crawling and RAG in one pipeline for AI coding assistants and agents. AI collects information from the internet with Crawl4AI, stores it in Supabase's vector database, and answers questions based on this context. Written in Python, used with Claude, Cursor, Cline MCP. Small, focused tool for rapid prototyping of new AI Agent workflows. | | [Prism MCP](https://github.com/dcostenco/prism-mcp) | 129 | Prism MCP is a HIPAA-standard cognitive architecture for AI agents using the MCP protocol. Features on-device LLM (prism-coder:7b), Hebbian learning, and ACT-R spreading activation for memory organization similar to human cognition. No API keys required — fully local operation. Multi-agent Hivemind synchronization, adversarial evaluation, and visual dashboard built-in. Written in TypeScript, designed for confidentiality, suitable for doctors and regulated industries. | | [agentmemory](https://github.com/rohitg00/agentmemory) | 20K | agentmemory provides persistent memory for AI coding agents with lifecycle hooks, silently recording agent activity across 12+ tools (Claude Code, Copilot CLI, Cursor, Gemini CLI), compressing observations into searchable structured memory (keyword + vector + knowledge graph hybrid, 95.2% retrieval accuracy R@5), and returning relevant context in new sessions. Works on SQLite without an external database and runs as a 53-tool MCP server — reducing tokens by ~92%. TypeScript, Apache-2.0, v0.9.24 (May 29, 2026), 19,600+ stars, GitHub Trending #5 — **Claude Mem** and **mem0**'s MCP-native competitor. | ## ⚙️ AI Infrastructure & Tools > **AI Infrastructure & Tools** — Run, serve, train, and create with AI models locally, including voice AI, video generation, and productivity tools. | Repository | ⭐ | Description | |---|---|---| | [Ollama](https://github.com/ollama/ollama) | 169K | Ollama - The easiest way to run LLMs locally. Run models like Llama, Mistral, Qwen, DeepSeek, Gemma, Phi, and more with a single command. Written in Go, works on macOS, Windows, Linux, and supports GPUs. Includes a REST API for easy integration with OpenAI SDKs. | | [Transformers](https://github.com/huggingface/transformers) | 159K | Hugging Face Transformers - A standard framework for state-of-the-art machine learning models in Python. Supports models like Llama, Mistral, Qwen, DeepSeek, Gemma, Phi, and more. Offers a unified API for text, audio, visual, and multimodal models. | | [ComfyUI](https://github.com/comfy-org/ComfyUI) | 109K | ComfyUI - A flexible platform for generating images and videos with diffusion models, featuring a node-based visual interface. Written in Python, includes a REST API, and supports over 1,000 community extensions. | | [llama.cpp](https://github.com/ggml-org/llama.cpp) | 106K | llama.cpp - A C/C++ library for LLM inference, optimized for CPU and minimal hardware. Introduces the GGUF format for quantized models. Used by Ollama, LM Studio, and LocalAI. | | [vLLM](https://github.com/vllm-project/vllm) | 77K | vLLM - A high-throughput inference and serving engine for LLMs in Python. Utilizes PagedAttention technology for efficient GPU memory management. Includes an OpenAI-compatible REST API for production deployment. | | [Stable Diffusion](https://github.com/CompVis/stable-diffusion) | 72K | Stable Diffusion - A latent diffusion model for text-to-image generation, made widely accessible in 2022. A collaborative research project by CompVis, Stability AI, and Runway. | | [Unsloth](https://github.com/unslothai/unsloth) | 62K | Unsloth - A tool for fine-tuning and running LLMs locally, featuring a web interface. Supports models like Gemma, Qwen, DeepSeek, and Llama. Offers faster fine-tuning and reduced GPU memory usage. | | [whisper.cpp](https://github.com/ggerganov/whisper.cpp) | 48K | whisper.cpp - A C/C++ port of OpenAI's Whisper speech-to-text model. Supports transcription in over 100 languages, including real-time streaming and speaker diarization. | | [LocalAI](https://github.com/mudler/LocalAI) | 45K | LocalAI - An open-source AI engine, serving as a local replacement for the OpenAI API. Supports LLMs, vision, TTS, STT, image generation, and video, all in one server, without requiring a GPU. | | [Coqui TTS](https://github.com/coqui-ai/TTS) | 45K | Coqui TTS - An open-source deep learning toolkit for text-to-speech in Python. Supports voice cloning and various TTS architectures. | | [Bark](https://github.com/suno-ai/bark) | 39K | Bark - A text-prompted generative audio model by Suno AI. Generates speech, songs, sound effects, and non-verbal sounds. | | [Diffusers](https://github.com/huggingface/diffusers) | 33K | Hugging Face Diffusers - A Python library for state-of-the-art diffusion models for image, video, and audio generation. Supports models like Stable Diffusion, FLUX, and SDXL. | | [SGLang](https://github.com/sgl-project/sglang) | 26K | SGLang - A high-performance serving framework for LLMs and multimodal models in Python. Utilizes RadixAttention technology for efficient batch workloads. | | [FLUX](https://github.com/black-forest-labs/flux) | 25K | FLUX - The official inference repository for FLUX.1 image generation models by Black Forest Labs. Offers Apache 2.0 and non-commercial licenses. | | [Google Workspace CLI](https://github.com/googleworkspace/cli) | 25K | Google Workspace CLI - The official command-line tool for Google Workspace services, including Gmail, Drive, Calendar, Sheets, Docs, Chat, and Admin. | | [toon](https://github.com/toon-format/toon) | 23K | TOON (Token-Oriented Object Notation) - An alternative serialization format to JSON, optimized for LLM prompts. Offers compact, human-readable, and schema-aware data representation. | # Tool List | [Kronos](https://github.com/shiyu-coder/Kronos) | 20K | Kronos is a Python pre-trained foundation model specifically designed for the "language" of financial markets — K-line (candlestick) series and financial indicators — trained on financial text corpora instead of general text corpora. An academic paper presented at AAAI 2026 supports it. This is not an agent framework but a domain-specific LLM research sample for financial ML researchers and trading system back-end developers. | | [VoxCPM](https://github.com/OpenBMB/VoxCPM) | 15K | VoxCPM2 is a tokenizer-free multilingual text-to-speech system from OpenBMB (Tsinghua University research group). Works on 30 languages without language tags, voice cloning from 3-second samples, 48kHz studio quality. 2 billion parameters, runs on 8GB VRAM, Apache 2.0 license. 85.4% voice similarity on Minimax benchmark. An open alternative to ElevenLabs — studio-quality TTS without commercial service. | | [HunyuanVideo](https://github.com/Tencent-Hunyuan/HunyuanVideo) | 12K | HunyuanVideo is a large-scale video generation framework from Tencent. Generate videos from text, image, or audio prompts. Written in Python, available on Hugging Face. 13 billion parameters, competitive with other open video gen models like Wan. Requires powerful GPU for local deployment (80GB VRAM recommended). For broad video generative research and ComfyUI workflows. | | [NotebookLM Python](https://github.com/teng-lin/notebooklm-py) | 11K | notebooklm-py is an unofficial Python API and agentic skill for Google NotebookLM. Not an official Google library but a reverse-engineered Python wrapper that unlocks functions inaccessible through the web interface. AI podcast generation, research automation, CLI support. Used as an agentic skill in Claude Code, Codex, OpenClaw. For local automation in Python. Stability depends on Google API changes. | | [Hyperframes](https://github.com/heygen-com/hyperframes) | 11K | Hyperframes is a TypeScript framework from HeyGen (a leading AI avatar video company) that converts HTML/CSS into video. Specifically designed for AI agents: agents write HTML, Hyperframes renders it as video — a code-driven video pipeline. Includes a [Claude-specific design guide](https://github.com/heygen-com/hyperframes/blob/main/docs/guides/claude-design-hyperframes.md) — create a first draft of video templates in Claude.ai, download as ZIP, then Claude Code finishes animations. A competitor to Remotion with a declarative agent-first approach — Plain HTML + GSAP, no React. Released in March 2026, Apache 2.0. | | [LM Studio](https://github.com/lmstudio-ai/lms) | 4.7K | LM Studio CLI is the command-line interface for LM Studio's local AI platform, written in TypeScript. Complements LM Studio's graphical interface: start/stop servers, load models with GPU acceleration, model management, process monitoring from the terminal. Useful in CI/CD pipelines, automation scripts where GUI is inaccessible. An alternative to Ollama for those who prefer LM Studio's visual interface. | | [RealtimeTTS](https://github.com/KoljaB/RealtimeTTS) | 3.9K | RealtimeTTS is a Python library that converts text to speech in real-time, optimized for streaming input. Receives LLM output as a stream and renders the first words within 2 seconds — without waiting for the full response. Supports multiple TTS backends (Coqui, ElevenLabs, Azure, Piper, Kokoro) with a single API. Voice cloning with Coqui backend. Independent project by Kolja Beigel (KoljaB), MIT license. For building full voice AI pipelines. | | [video-use](https://github.com/browser-use/video-use) | 3.8K | video-use is a tool for video editing through AI coding agents by the Browser Use team. Takes natural language instructions (merging fragments, adding subtitles, replacing audio, changing speed) and automatically writes and executes AI FFmpeg scripts. Written in Python, released in April 2026. An early demo of agentic media processing — a new direction for non-linear editing (NLE) UI-less, software video workflows. | | [ViMax](https://github.com/HKUDS/ViMax) | 8.3K | ViMax is an agentic video generation framework from HKUDS (the same lab that created LightRAG, RAG-Anything, nanobot, DeepTutor, and OpenSpace — all in this collection). Specialized agents in Director, Screenwriter, Producer, and Video Generator roles turn ideas, scripts, or even novels into complete videos — Idea2Video, Script2Video, Novel2Video, and AutoCameo (user photo as character). Script generation, storyboard, shot planning, reference image selection, and consistency validation are automated. Python, MIT, 8,300+ stars, GitHub Trending #8 — differs with **HunyuanVideo** and **HunyuanVideo** with an agentic, role-based approach. | | [Supertonic](https://github.com/supertone-inc/supertonic) | 11K | Supertonic is an ultra-fast, on-device multilingual TTS with ONNX — a compact 99M-parameter model, 31 languages, 44.1kHz audio, cloud-free on desktop/mobile/browser/edge. 10 expression tags for natural intonation and ready-to-use implementations in 11 languages (Python, Node, Java, C++, Swift, Rust, Flutter…). MIT (code) + OpenRAIL-M (model), 11,000+ stars, GitHub Trending #6 on May 21, 2026 — one of the lightest on-device TTS options alongside **VoxCPM**, **RealtimeTTS**, and **bark**. | | [Pixelle-Video](https://github.com/AIDC-AI/Pixelle-Video) | 21K | Pixelle-Video is a fully automated short-video engine from Alibaba's AIDC-AI team — give a topic and the system assembles the video: writes scripts with LLMs (GPT, Qwen, DeepSeek), generates images/videos with ComfyUI integration, synthesizes voiceovers (with voice cloning), adds music, and composes the final video. Digital human narration and motion transfer are built-in, with a web UI for configuration. Python, Apache-2.0, 20,500+ stars — a leader in the Chinese ecosystem for end-to-end short-video pipelines alongside **HunyuanVideo** and **ViMax**. | ## 📚 Learning Resources > **Learning Resources** — Learning resources, awesome lists, free AI APIs, and references. New to AI or looking to dive deeper? Find everything here, from templates to access to free AI. | Repository | ⭐ | Description | |---|---|---| | [Build Your Own X](https://github.com/codecrafters-io/build-your-own-x) | 493K | Build Your Own X — One of the most starred educational repositories on GitHub by CodeCrafters: a catalog of tutorials where you build technologies from scratch. Git, Docker, database, operating system, programming language, LLM, neural network, web browser, game engine — thousands of step-by-step guides by category. Written in Markdown, language-neutral. A companion list to CodeCrafters' interactive platform. Ideal for gaining in-depth programming knowledge — the best resource for acquiring system-level knowledge. | | [Public APIs](https://github.com/public-apis/public-apis) | 426K | Public APIs — One of the most starred repositories on GitHub, a collective list of hundreds of categorized free APIs. Weather, music, finance, maps, sports, government data, AI services — with authentication, CORS, and license information. Written in Python with validation scripts, community-maintained. The first stop when assembling AI agent workflows and searching for APIs. | | [System Prompts Collection](https://github.com/x1xhlol/system-prompts-and-models-of-ai-tools) | 135K | System Prompts Collection — A collection of leaked and public system prompts from over 25 AI tools: Claude Code, Cursor, Devin AI, v0, Lovable, Replit, Windsurf, Manus, Perplexity, Kiro, Warp, Xcode, and more. Essential learning material for prompt engineering — how frontier companies write system prompts in terms of structure, language, and logic. For building AI agents, product prompt engineering, or competitive analysis. Updated by the community. | | [Excalidraw](https://github.com/excalidraw/excalidraw) | 122K | Excalidraw — A virtual whiteboard for hand-drawn style diagrams, written in TypeScript, with an MIT license. Has become a **de facto standard** for AI engineers to visualize agent architectures, system designs, RAG pipelines, and LLM workflows — frequently featured in engineering blogs and technical documentation from Anthropic, OpenAI, and other AI companies. Excalidraw+ commercial version includes a text-to-diagram AI feature (with OpenAI integration) that generates diagrams from natural language, plus native support for Mermaid — enabling a unified workflow with Claude/GPT to generate Mermaid syntax and refine it in Excalidraw. A primary tool for AI architecture documentation, presentations, or visualizing Claude Code output. | | [Mermaid](https://github.com/mermaid-js/mermaid) | 88K | Mermaid — The unconditional open-source standard for text-to-diagram. From Markdown-style syntax to flowcharts, sequence diagrams, ER, Gantt, state machines, class diagrams, mindmaps, C4 architecture, and more — all in one language. Written in TypeScript, MIT license, 87,900+ stars. **The de-facto format in AI workflows** — Claude, ChatGPT, Cursor, Codex, and nearly all LLMs natively understand Mermaid syntax and generate diagrams directly in Markdown responses. GitHub, GitLab, Notion, Obsidian, Confluence render Mermaid natively — diagrams appear directly in pull requests, code reviews, and version control. A mandatory part of ARCHITECTURE.md, RFCs, and technical documentation in modern engineering teams. A key part of the diagramming workflow, often paired with **Excalidraw** (visual) and **D2** (code-based — check separately). | | [tldraw](https://github.com/tldraw/tldraw) | 47K | tldraw — A modern infinite canvas SDK and whiteboard application, written in TypeScript, with real-time collaboration. **The most serious competitor to Excalidraw**: known for its bright, contrasting, professional visual style, as opposed to Excalidraw's hand-drawn aesthetic. Embeddable React component — import `<Tldraw />` into your application to get a whiteboard without needing to build one from scratch. **AI feature** [tldraw/make-real](https://github.com/tldraw/make-real) transforms sketches into working HTML/React components with one button (using OpenAI/Claude backend). Funded by Lightspeed Venture Partners, 46,900+ stars. Ideal for those needing a whiteboard feature within an app (an alternative to Figma or Miro) or a stricter visual alternative to Excalidraw. | | [LLMs from Scratch](https://github.com/rasbt/LLMs-from-scratch) | 91K | LLMs from Scratch — A collection of Jupyter Notebooks from Sebastian Raschka's (Lightning AI researcher) book "Build a Large Language Model (From Scratch)". Building a ChatGPT-like LLM from scratch in PyTorch: attention mechanism, tokenization, pre-training, fine-tuning, RLHF — all stages with code. A bestseller from Manning Publications, with open Jupyter notebooks. The most comprehensive and practical course for understanding AI internal mechanisms — essential for both research and engineering. | | [AutoResearch](https://github.com/karpathy/autoresearch) | 76K | AutoResearch is an experimental Python project by Andrej Karpathy (former research lead at OpenAI and Tesla AI lead) where AI agents train on nanochats with one GPU, automating research. The project demonstrates the concept of "automated ML research" — agents conduct experiments, evaluate results, and repeat the cycle without human intervention. Due to Karpathy's authority, it quickly gathered 76K stars, but it's a research/demo project, not a production-ready framework. A noteworthy direction for those interested in automating ML research. | | [DESIGN.md](https://github.com/google-labs-code/design.md) | 6.3K | DESIGN.md — Google Labs' official format specification for describing a brand's visual identity (colors, typography, spacing, component behavior, interaction patterns) in a structured way for AI coding agents. Place a DESIGN.md file in your project's root, and Claude Code, Cursor, or Codex will have a consistent understanding of your design system — generating UI in a consistent style for each new feature, without stylistic drift or hallucinations. Written in TypeScript, Apache 2.0 license. Published on April 10, 2026, and gathered 6,300+ stars in 14 days (about 450 ⭐ per day) — viral growth similar to Anthropic's Agent Skills Spec, but in the design systems domain. The companion `awesome-design-md` provides over 55 examples built with this specification — a standard foundation. | | [Awesome DESIGN.md](https://github.com/VoltAgent/awesome-design-md) | 64K | Awesome DESIGN.md — A collection of over 55 Markdown files with popular brand design systems (Stripe, Google, Apple, Linear) created in the DESIGN.md style. Place a DESIGN.md file in your project, and tell Claude Code, Cursor, or other coding agents — UI will be generated automatically in your brand's style. By VoltAgent, since March 2026. The fastest path to brand-consistent UI in visual code generation workflows without manual coding. | | [Anthropic Cookbook](https://github.com/anthropics/anthropic-cookbook) | 41K | Anthropic Cookbook — An official Jupyter Notebook collection by Anthropic with practical examples of using the Claude API: tool use, prompt caching, multi-agent workflows, vision, PDF processing, computer use, extended thinking, batch API — all core Claude capabilities with code examples. Maintained directly by Anthropic's team — updated with new feature releases. The official starting point for first-time Claude API users. | | [Awesome Claude Code](https://github.com/hesreallyhim/awesome-claude-code) | 40K | Awesome Claude Code — A curated list of the Claude Code (Anthropic) ecosystem: skills, hooks, slash-commands, agent orchestrators, applications, plugins. For maximum efficiency with Claude Code — community-curated workflows, automation scripts, and integrations in one place. Written in Python, community-maintained. A quick start for new Claude Code users — learning usage patterns and advanced workflows. | | [Claude × HyperFrames Guide](https://github.com/heygen-com/hyperframes/blob/main/docs/guides/claude-design-hyperframes.md) | Guide | Claude × HyperFrames Guide — An official guide by HeyGen on team workflows using Claude.ai and Claude Code for creating video templates. Claude.ai builds the first draft (brand, layout, scenes) with HTML/CSS/GSAP — then users download it as a ZIP and Claude Code polishes it (timing, easing, mid-scene activity). Includes a banned fonts/pairings list for anti-monoculture, skeletons for video types (social reel 9:16, launch teaser 16:9, product explainer, cinematic title), and structural lint rules. One of the first official workflow documents for template-first AI design. | | [LLM-Maintained Wiki (Karpathy)](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) | Gist | Andrej Karpathy's (former AI lead at OpenAI and Tesla) conceptual gist — AI automatically creates and updates a personal Markdown wiki based on new information. An alternative approach to RAG: information isn't stored in a vector DB, but rather the LLM itself rewrites, updates, and adds cross-references to wiki pages. Not code — a thought-provoking paper that sparked a wide discussion on LLM-maintained knowledge bases. | | [Claude How-To](https://github.com/luongnv89/claude-howto) | 28K | Claude How-To — a visual, example-based guide for Claude Code: from basic concepts to building advanced agents, with copy-paste templates. In Python, created in 2025. Unlike Anthropic Cookbook, less official and more community-style — but with more ready-to-use templates. For new Claude Code users who need a quick, practical guide. | | [Free LLM API Resources](https://github.com/cheahjs/free-llm-api-resources) | 19K | Free LLM API Resources — a curated list of over 26 free LLM inference services with API access: OpenRouter, Google AI Studio, NVIDIA NIM, Mistral, Groq, Cerebras, Together AI, and more. For each service — available models, rate limits, registration requirements. In Python, by cheahjs, with community updates. For prototypes and learning projects where you don't want to pay — the most comprehensive list. | | [12-Factor Agents](https://github.com/humanlayer/12-factor-agents) | 23K | 12-Factor Agents — humanlayer's 12 engineering principles for building production-grade LLM software — an AI analog of the "12 Factor Apps" methodology. Argues that the best way is not to rely heavily on a framework, but to embed small, modular agent concepts into existing products — prompt ownership, context window management, tool design, state unification, human coordination. Apache-2.0 (code) + CC BY-SA 4.0 (content), 22,800+ stars, GitHub Trending #9 — a foundational learning resource alongside **awesome-claude-code** and **Anthropic Cookbook**. | --- ## About Us <p align="center"> <a href="https://aipulsegeorgia.ge"><img src="https://img.shields.io/badge/AI_Pulse_Georgia-2026-00D0FF?style=for-the-badge&labelColor=111827" alt="AI Pulse Georgia 2026"/></a> </p> This list is maintained by **[AI Pulse Georgia](https://aipulsegeorgia.ge)** — a community focused on AI agents, automation, and the future of autonomous systems. > *"Exploring Georgia's AI Future"* If this list helps you, give it a star and share it with others who build with AI agents. ## Contributing Find a great repository that fits this list? Open an issue or send a pull request. ## License [](https://creativecommons.org/publicdomain/zero/1.0/)