AI open-source projects worth tracking and trying
ByteDance's DeerFlow 2.0 continues to see frequent updates as of June 19, boasting over 71k stars on GitHub. It bets on sub-agents, memory, sandboxing, and skill orchestration.
EverOS 1.0.0 focuses on a local-first memory layer, using Markdown as the source of truth, while syncing with SQLite and LanceDB for retrieval.
Headroom remains actively updated as of June 19, focusing on compressing tool outputs, logs, RAG chunks, and files. Its README claims a 60-95% reduction in token usage.
MemOS 2.0 remains actively maintained, focusing on a unified memory API, multimodal memory, feedback correction, and cross-task skill reuse, with a recent push on June 18.
Solace Agent Mesh remains actively updated as of June 19. It organizes multi-agent systems using event-driven architecture, A2A, and ADK, making it ideal for exploring production-ready collaboration patterns.
The `humanlayer/12-factor-agents` repository has 21,500 stars, making it one of the most watched engineering methodology repos in the AI Agent space. It attempts to answer a single question: What principles should be followed when building production-grade AI Agents? But there's an awkward fact—the project's last push was eight months ago.
AgentMemory gained nearly 8,000 GitHub stars in a week, with a clear positioning: solving the "amnesia" problem of AI coding agents. Every time you launch Claude Code or Codex, they act like a brand new person—unaware of yesterday's changes, past project decisions, or your preferences. AgentMemory aims to solve this with a persistent memory layer.
CodeGraph tackles a real pain point with a pre-indexed code knowledge graph: AI coding agents waste massive amounts of tokens and tool calls by re-reading the entire codebase every time a project is opened. Gaining 6,700 stars in a week, it's not a gimmick—it's a tool that genuinely saves money.
OpenHuman surged by 19k stars in a week, claiming the top spot on GitHub this week. It promises a fully private, personal AI super-assistant—data stays on-device, supports multi-model integration, and enables mobile deployment. But its ambitions go far beyond that.
Supertone's open-source Supertonic surged into the top 5 on GitHub Trending this week, reaching 9,099 stars. It leverages ONNX Runtime to deliver on-device multilingual text-to-speech with latency low enough for real-time conversations. The newly released v3 version now supports a full suite of language SDKs, including C++, Go, Rust, Java, Swift, and Flutter.
9Router routes all your AI coding tools to 40+ providers and 100+ models. RTK technology compresses tool_result to save 20-40% tokens. Three-tier auto-fallback ensures coding never stops. 11K stars, 58 versions.
yikart/AiToEarn hit 14.2k stars in two weeks, promoting "use AI to earn." 2,599 commits, but working cases need careful filtering.
bolt.diy is an open-source alternative to Bolt.new, supporting local deployment and custom models.
CloakBrowser doesn't rely on config patches or JS injection — it applies 49 patches at the Chromium C++ source level. All 30+ anti-detection sites passed, including Cloudflare Turnstile and FingerprintJS.
LangGraph is LangChain agent graph orchestration framework, supporting loops, state management, and conditional branching.
TypeScript educator Matt Pocock open-sourced his .claude directory. 18,795 stars in a week, 87,649 total. Not a new concept, but it breaks down common AI coding agent failure modes into reusable skill packs.
romainsimon/paperasse turns French bureaucratic roles into Claude Agent Skills. 1.8k stars. The vertical agent approach is worth studying.
PraisonAI deploys autonomous agent teams in 5 lines of code, with built-in memory, RAG, and support for 100+ models. 7.8k stars, 3,722 commits.
The team behind million (react-fast-refresh) launches React Doctor: one command scans your codebase for a 0-100 health score across state/effects, performance, architecture, security, and a11y.
Meetily is an open-source AI meeting assistant using Parakeet/Whisper for real-time transcription and Ollama for meeting summaries — all processed locally, no cloud upload. Rust-built, 12K stars.
qiaomu-anything-to-notebooklm is a Claude Skill that processes multi-source content (WeChat articles, web pages, YouTube, PDF, Markdown) into NotebookLM outputs: podcasts, PPTs, mind maps, quizzes. +465 stars today.
Tencent open-source WeKnora is an LLM knowledge platform that transforms raw documents into queryable RAG systems, autonomous reasoning agents, and self-maintaining wikis. 15K stars, 1,661 commits.
new-api is a unified AI model management gateway supporting cross-format conversion of various LLMs to OpenAI/Claude/Gemini compatible interfaces. 33,573 stars, +2,000/week.
cc-connect bridges local AI coding agents (Claude Code, Cursor, Gemini CLI, Codex) to messaging platforms including Feishu, DingTalk, Slack, Telegram, and Discord. No public IP needed. 9.4K stars, +1,400/week.
LMSYS announces SGLang and Miles achieved Day-0 support for DeepSeek-V4 inference and RL training, becoming the first open-source stack to serve and train a new flagship model on launch day.
Astro co-founder Fred K. Schott open-sources flue, a framework providing sandbox execution and skill scheduling for AI agents, with version 0.4.1 released just 9 hours ago.
LMSYS team introduces RDMA-based P2P weight update mechanism in SGLang as a supplement to traditional NCCL broadcast. Weight synchronization for trillion-parameter RL training compressed from minutes to seconds, compatible with all major open-source models.
Open-sourced by Google engineer Addy Osmani, agent-skills gained 10K stars in a week, reaching 38.9K. It packages senior engineering habits like spec-first development, small-step tasks, and test-driven validation into skills that AI Agents can automatically invoke.
OpenAI quietly open-sourced Symphony—a specification plus reference implementation that enables teams to manage work rather than supervise agents. With only 12 commits and an Apache 2.0 license, its directional signal matters far more than the volume of code.
thesysdev/openui proposes an open standard protocol for generative UI. 4.5K stars, 705 stars/week. AI-generated UI is no longer driven by prompt guessing but by structured protocols. The frontend development workflow is being rewritten.
Ruflo gained 10.7K stars in a week, reaching 48.6K, with 6,387 commits and 32 plugins. It isn't building a new framework; instead, it equips Claude Code with an enterprise-grade nervous system that coordinates swarm memory federation.
xingkongliang/skills-manager uses a desktop app to centrally manage AI Agent Skills across 15+ coding tools including Claude Code, Codex, Cursor, and Copilot. 1,286 stars. As AI coding tools fragment, Skills management is the next pain point.
omlx is an LLM inference server designed for Apple Silicon, supporting continuous batching and SSD caching, managed from the macOS menu bar. 13K stars, Apache 2.0 open source.
Bun's experimental Rust rewrite reaches 99.8% test compatibility on Linux x64 glibc. As the underlying runtime for many AI developer tools, Bun's core reconstruction could reshape the performance ceiling and development experience of AI applications.
AiToEarn has grown to nearly 10k stars in just over a year, offering automated publishing across 12+ platforms including Douyin/Xiaohongshu/TikTok/YouTube. It features four core agents — Monetize, Publish, Engage, and Create — along with a content marketplace and MCP protocol support.
AIComicBuilder has grown to 1.4k Stars in less than two months, automating the entire pipeline from script import and character four-view generation to intelligent storyboarding and keyframe-interpolated video output. Supports OpenAI, Gemini, Tongyi Wanxiang, and multiple other model providers.
This week GitHub Trending is dominated by AI Agent projects: Anthropic financial services repo gained 8,800 stars in a week, multi-agent orchestration platforms, persistent memory tools, and coding agent harnesses emerge densely.
HiDream open-sources HiDream-O1-Image, an 8B parameter image generation model using a pixel-level Unified Transformer architecture with no VAE or separate text encoder. It ranks 8th in the Artificial Analysis Image Generation Arena, the leading open-weights text-to-image model.
TanStack team releases AI SDK alpha version, providing a unified AI interface across frameworks, models, and languages. 2.6K stars, active development. A signal from frontend giants entering the AI toolchain.
AstrBot, developed by a BUPT student, surpasses 31K stars, integrating multiple IM platforms and LLMs with 1000+ plugin ecosystem, positioning itself as an OpenClaw alternative. Another path for Chinese agent frameworks.
Rowboat is an open-source AI coworker project with persistent memory capabilities, 1.6k commits, 13.7k stars. Unlike one-off chatbots, it remembers historical interactions and long-term project context.
Agentmemory provides a persistent memory layer for AI coding agents, built on Cloudflare Workers. Supports Claude Code, Codex, OpenClaw, Hermes Agent. 3.3k stars, 518 daily increase.
Google open-sources A2UI (Agent-to-UI) framework, enabling AI agents to output renderable frontend components instead of plain text. Supports React and Angular, 14.6k stars, solving the last-mile UI rendering challenge for agent deployment.
datawhalechina hello-agents project surpassed 45K stars, growing at 667 stars per day. It is the most active Agent principles and practice tutorial in the Chinese community with 773 commits.
ByteDance open-source UI-TARS-desktop surpassed 31K stars, positioning itself as an open-source multimodal AI Agent stack connecting cutting-edge AI models with agent infrastructure.
YC-backed Browserbase open-sourced skills repository gaining 1,621 stars in one week, providing native web browsing capabilities for Claude Agent SDK, enabling AI agents to browse, interact with, and extract information from web pages.
GenericAgent grows a skill tree from 3.3K lines of seed code, achieving full system control with 6x less token consumption. Gained 5,700+ stars in a week on GitHub with extremely high activity.
cc-switch is a cross-platform desktop tool for fast switching between Claude Code, Codex, OpenCode, openclaw, and Gemini CLI. 64.8K GitHub stars, 1658 commits, extremely active community.
CloakBrowser is a stealth Playwright alternative with source-level fingerprint patching, passing 30/30 bot detection tests. Useful for AI agents needing web data scraping, but legal risks are real.
Google Cloud open-sourced Agents CLI, packaging the full AI agent lifecycle from local prototyping to production deployment into a command-line tool that integrates seamlessly with Claude Code, Gemini CLI, Cursor, and other coding agents.
Qwen open-sources WebWorld-14B/32B, a web agent world model based on Qwen3 architecture under Apache 2.0. Instead of traditional browser automation, the model learns to 'imagine' the results of web interactions.
The official chrome-devtools-mcp project from Chrome team surpassed 38.6K stars, with v0.25.0 just released. Coding agents can now connect to DevTools via MCP protocol, no more guessing DOM structures from screenshots.
Google open-sources ADK Go edition with 7.8K stars and 407 commits. Agent logic defined in Go code rather than prompts, with native A2A protocol and VertexAI MemoryBank support.
hugohe3/ppt-master trends on GitHub weekly chart with 13.5k stars — AI generates truly editable PPTX files from documents, with real PowerPoint shapes and native animations, not screenshot pastes.
Sakana AI and NVIDIA jointly publish an ICML2026 paper introducing TwELL sparse packing format and fused CUDA kernels, leveraging the natural 95% zero-activation sparsity in LLM feedforward layers for 20%+ speedup in inference and training.
9router is a local AI gateway routing requests from Claude Code, Cursor, Cline and other coding tools to 40+ free/cheap AI backends, with auto-fallback and RTK token compression. 5.8K GitHub stars with 1,000+ daily growth.
AI-Trader, an open-source project by the HKU Data Science Lab, has hit GitHub Trending with 14.4K stars. Agents only need to read a single SKILL.md file to automatically register and participate in trading discussions, copy trades, and synchronize signals, supporting multiple markets including stocks, crypto, and forex.
fal.ai launched genmedia CLI to generate images, video, 3D, and audio directly from the terminal. Native terminal workflow, no Dashboard needed, directly integrable into Claude Code, Cursor and other AI Agent scripts and pipelines. 33K+ views and 190 bookmarks on day one.
NVIDIA Research proposes Guess-Verify-Refine algorithm, leveraging temporal pattern reuse across decoding steps, achieving 1.88x Top-K attention speedup on Blackwell + TensorRT-LLM. Hardware-aware inference optimization is reshaping the speed competition.
Redis founder Salvatore Sanfilippo (antirez) wrote a DeepSeek 4 Flash local inference engine ds4 in C from scratch, targeting Apple Silicon Metal GPU. Released 13 hours ago, now at 846 stars, already trending on Hacker News.
Mistral AI open-sources Voxtral TTS with 90ms first-sound latency, directly challenging ElevenLabs. The open-source speech AI suite completes its final puzzle piece.
Cursor team releases /orchestrate skill, built on Cursor SDK for recursive agent spawning to complete complex tasks. Internal tests show 20% token reduction and 80% backend cold start time reduction. This is not simple multi-agent concurrency, but a layered architecture from root planner to scoped executor.
FreeMoCap replaces traditional motion capture rigs with pure computer vision, using only USB cameras for multi-person 3D skeleton tracking. Exports to Blender and Unity. 8.1k stars on GitHub.
Hugging Face released nanowhale, a 110M parameter model replicating DeepSeek-V4 core architectures including MLA, MoE, and MTP. Benchmarks are not the point; its value is enabling low-cost architecture research.
strukto-ai/mirage is a unified virtual filesystem for AI Agents, mounting S3, Google Drive, Slack, Gmail, Redis and other backends into a single file tree. Agents only need ls, cat, grep and other bash commands to operate all data sources. Just released v0.0.1-alpha.1.
MLX-VLM, built on Apple MLX framework, enables Mac users to run and fine-tune vision language models locally. 51 new stars today, an important tool for edge multimodal inference.
Open-LLM-VTuber connects Live2D avatars with any LLM, supporting hands-free voice chat and voice interruption. Runs locally across platforms. 7.5k stars on GitHub.
OpenAI open-sources Codex plugin repository, first batch covers Figma, Notion, Slack, Linear and more. Plugins bundle MCP servers, workflows, and app integrations into distributable packages.
OpenSearch-VL is an open-source framework training multimodal deep search agents with reinforcement learning, introducing fatal-aware GRPO to handle cascading failures. Achieves 10+ capability improvements across search tasks.
Rapid-MLX, built on Apple native MLX framework, delivers 2-4x faster inference than Ollama on Apple Silicon, hitting 160 tok/s on 4B models. Day-0 support for Qwen3.6 and DeepSeek V4 Flash, with OpenAI-compatible API. But ecosystem and model library remain weak points.
iamzhihuix/skills-manage is a cross-platform desktop app that统一管理 Agent skills across Claude Code, Cursor, Codex, and 20+ platforms. 1.7k stars, 194 commits, uses symlinks to sync the same Skill to different AI coding tools, avoiding duplicate maintenance.
Anthropic donates open-source alignment evaluation tool Petri to Meridian Labs for independent operation, releasing v3.0. UK AI Security Institute already uses Petri to test every Claude model.
ReactorWorld open-sources real-time world models, achieving sub-50ms latency for prompt-driven scene generation. Type a prompt, modify the scene, control generation in real time.
Anthropic has an open source financial services agent reference repository on GitHub, which includes ready-to-use agents for investment banking, equity research, private equity and other scenarios. It supports two deployment methods: Claude Cowork plug-in and Managed Agents.
Kronos is a foundation model for financial markets that treats price, volume, news and other financial data as a 'language' to understand. 11,751 stars/month, 23,385 total — one of GitHub's fastest-growing financial AI projects this month.
multica is an open-source managed agents platform supporting task assignment, progress tracking, and skill accumulation for coding agents. 23,259 stars/month, 25,782 total — one of GitHub's fastest-growing agent platforms this month.
ouroboros calls itself an 'Agent OS' with a core thesis of shifting from prompts to specifications. Define agent behavior declaratively rather than debugging through natural language. 332 stars/week, new on GitHub weekly trending.
tokscale is a Rust-built CLI tool that automatically tracks token consumption across 14 platforms including OpenClaw, Claude Code, Codex, Gemini CLI, and Cursor. Features global leaderboard and 2D/3D contribution graphs.
WebBrain is a local browser agent powered by the quantized Qwen3.5-9B int4 version, running on just 8GB VRAM (16GB MacBook, RTX 4060/3060 etc.), completely offline with zero cloud dependency. Tested 22 vision-language models to select the optimal solution.
DFlash achieves up to 6x inference speedup through Block Diffusion-based speculative decoding, reaching 400+ tokens/s on Qwen3.5 and other latest models. Outperforms EAGLE-3 by up to 2.5x with zero accuracy loss. RTX 4000 Ada consumer GPU delivers 161 tok/s.
Hermes Agent GitHub Stars surpass 140,000, becoming the largest open-source project in the AI Agent space. This Saturday, a Seoul tech talk will demonstrate how to use Hermes as a "practical CTO" — covering the full workflow of building, testing, and monitoring.
Prime Intellect Lab exits beta, offering an end-to-end RL training platform for building self-improving AI agents that continuously learn from real experience and feedback.
OfficeCLI is an AI-native office suite designed for agents, supporting creation, editing, and automation of Word, Excel, and PowerPoint from terminal commands. It solves the awkward gap where agents can code but can't handle real office documents.
InsForge is rising on GitHub Trending, gaining 230+ stars daily. It is a Postgres-based full-stack backend platform designed specifically for AI coding agents — with auth, storage, compute, hosting, and AI gateway all built-in, solving the problem of agents being able to write frontends but struggling with backends.
Ollama announces native support for DeepSeek-V4-Pro, directly callable in Claude Code, OpenClaw, CodeX and other agent frameworks with zero additional configuration. 1 million token context window plus MoE architecture significantly lowers the barrier for local deployment.
Qwen3.6-27B model has been successfully deployed on NVIDIA GB10 edge computing modules with Q6 quantization reaching usable performance. This marks 27B-class open-source models entering the consumer-grade edge inference era, dramatically lowering the barrier for local AI research.
An open-source project aggregates all freely callable mainstream LLM APIs — GPT-5, Claude, Gemini 3 Flash, DeepSeek R1, Llama 4, Qwen3 — accessible via Groq, Cerebras, OpenRouter, Google AI Studio, and more, enabling zero-cost integration.
OpenHands open-sourced an autonomous coding agent platform that runs thousands of parallel coding tasks across any LLM, any CI/CD pipeline, and any codebase. Supporting code review, test generation, code modernization, and more, enterprise code automation enters a new phase.
ServiceNow launched Project Arc at Knowledge 2026 conference - an autonomous desktop Agent capable of thinking, writing code, and executing tasks. Partnership with NVIDIA ensures full-stack observability and governance from desktop to data center. Action Fabric system also opened, allowing any AI Agent to connect to enterprise workflows.
Community developer Kyle Hessling releases Qwopus3.6-35B-A3B-v1, an open-source model distilled and optimized from Qwen3.6 architecture, now available on HuggingFace with GGUF quantized versions. Clement Delangue (HuggingFace CEO) personally followed the project, signaling platform-level recognition for community distillation models.
Google Cloud AI team open-sources Agent Skills library on GitHub, focusing on engineering standards for AI coding. This toolkit provides reusable best-practice templates for Agent development, marking the evolution from "craft workshop" to "standardized pipeline" in Agent engineering.
Genesis AI officially exits stealth mode, releasing GENE-26.5 — the first robotic AI targeting human-level physical manipulation. Paired tactile sensing glove costs 100x less than existing solutions with 5x precision improvement, demonstrated completing 20+ subtasks in a real kitchen environment.
Tsinghua OpenBMB team releases VoxCPM2 open-source voice model, adopting a radical architecture that completely eliminates the traditional tokenizer, modeling directly in raw audio space. Compared to conventional TTS solutions, VoxCPM2 achieves significant breakthroughs in voice cloning quality and multilingual support, marking a new phase in open-source voice technology.
AionUi announces multi-agent team collaboration mode, automatically recognizing and integrating Hermes Agent, OpenClaw, Claude Code, and Codex into unified teams. Leader Agent handles task decomposition while Teammate Agents execute in parallel, achieving true heterogeneous agent framework synergy.
Meta AI released NeuralBench, a unified open-source framework for evaluating NeuroAI models. EEG foundation models finally have comparability, cross-model evaluation is no longer a pipe dream.
Vercel Labs open-sourced open-agents, a deployable cloud coding agent template. With 962 commits and 263 branches, Vercel is building reusable Agent infrastructure, not just a demo.
ViMax is an open-source all-in-one video generation tool that completes script writing, image generation, voice synthesis, and video editing with a single prompt, claiming to replace the combination of Runway, ChatGPT, Midjourney, and HeyGen, saving $114/month.
OpenGeoAgent is an open-source multimodal AI agent supporting automated geospatial analysis and visualization through QGIS, Jupyter Notebook, and Python scripting, rapidly gaining 831 stars, 133 retweets, and 887 bookmarks after launch.
QwenPaw is an open-source personal AI assistant built on the Qwen model ecosystem, supporting local or cloud deployment, integration with multiple chat platforms, and an extensible skills system, enabling individual users to enjoy LLM capabilities in a private environment.
Nanobrowser, a fully open-source AI web automation tool running as a Chrome extension with multi-agent collaboration capabilities, zero API costs, and flexible LLM selection, is emerging as the strongest open-source alternative to OpenAI Operator.
OpenClaw 2026.5.5 release fixes Feishu, LINE, Telegram, and Discord messaging platform issues, optimizes control UI/TUI responsiveness, maintains SDK link stability during plugin updates, and enhances Gateway status monitoring and restart diagnostics.
Agent Zero v1.13 officially released with core updates including deep LibreOffice integration (ODT/ODS/ODP as primary formats) and enhanced browser skills. This is a fully open-source Agent + office technology stack, focused on open formats and autonomous office capabilities.
Cocoindex is an incremental engine designed for long-running AI Agents. Through incremental state management and checkpoint-based resumption, it solves the problem of state loss and redundant computation during extended Agent runs. Growing 438+ stars daily on GitHub.
HuggingFace community developer Jackrong released Qwen3.6 35B A3B distilled model, using Claude Opus reasoning data for knowledge distillation. Model weights are 71.9GB, with GGUF version coming soon. This demonstrates how the community leverages closed-source model reasoning to enhance open models.
WLFI ecosystem launches WorldRouter, integrating 300+ models including Claude, GPT, Gemini, and Qwen into a unified account with API prices 30% below official rates. Supports smart routing and fallback - developers get one Key for all models.
Sulphur-2 is now open source on Hugging Face, supporting text-to-video (t2v) and image-to-video (i2v) generation without content moderation. As the first truly usable open-source video generation model, it directly challenges commercial closed-source alternatives like SeedDance, Kling, and Veo in creative freedom.
IndexTTS Community Edition V26 update: supports 8-speaker simultaneous dialogue dubbing, permanent voice library storage, emotion control, and 10x inference speed improvement. The GPT-based industrial-grade zero-shot TTS system (20.3k stars) is becoming the de facto standard for open-source speech synthesis.
Google Research publishes TurboQuant at ICLR 2026, compressing KV Cache memory by 4-6x through vector rotation and polarized quantization, achieving 8x attention speedup on H100. Qdrant and llama.cpp integrate it first, ushering in a new era for local LLM inference.
The CyberSentry open-source project is gaining traction — built on Llama 3.1 70B (NVIDIA NIM) + ReAct mode, integrating 8 real security tools including Nmap, directory fuzzing, and SSL checkers. It offers full autonomous penetration testing with target analysis, tool selection, result interpretation, and strategy adjustment.
A research team releases PageIndex, a RAG method that completely subverts traditional approaches — no vector database, no embedding, no chunking, no similarity search. Instead, it builds a tree index letting LLMs read documents like humans. Scores 98.7% on FinanceBench.
Ruflo grows at 2,432 stars per day, total stars exceeding 43,986. This Claude-oriented multi-agent orchestration platform offers enterprise-grade architecture, self-learning swarm intelligence, RAG integration, and native Claude Code/Codex support.
Unsloth releases a complete guide proving that with GGUF-quantized Gemma 4 and Qwen3.6, you can run full agentic coding on a 24GB RAM machine with self-healing tool calls. The barrier to low-cost local AI programming is pushed even lower.
Anthropic publishes a 33-page Claude Skills building guide. Open-source community already has 35+ agent frameworks supporting it. Prompts upgrade to Skills, system prompts evolve into SKILL.md, AI development paradigm is undergoing a fundamental shift.
Adaptive web scraping framework Scrapling gained 5,650 stars this week on GitHub, surpassing 44K total. Its core selling point is automatic handling of anti-scraping mechanisms, dynamic pages, and structural changes. This article analyzes Scrapling's technical advantages, competitor comparison, and use cases.
Google engineer addyosmani open-sourced agent-skills repository surpassed 29K stars, gaining 629 stars daily, collecting production-grade AI coding agent skill files supporting Claude Code, Cursor, Windsurf and more, becoming the de facto standard for agent skills.
Google Gemma 4 series officially supports MTP (Multi-Token Prediction) technology, achieving 2-3x inference speedup through speculative decoding with zero quality loss. SGLang provides Day 0 support covering all 4 Gemma 4 sizes. This is a major breakthrough for local AI deployment.
virattt/dexter has become a financial AI hit on GitHub trending with 23.5K stars. This open-source autonomous agent completes deep financial research like a human analyst: fetching SEC filings, analyzing financial reports, generating valuation reports, supporting OpenAI/Claude/Gemini/Grok/Ollama.
NVIDIA officially releases the Nemotron 3 Nano Omni open-source multimodal model on April 29, deeply optimized for Hopper and Blackwell FP8 inference, while also compatible with consumer GPUs like RTX 5090 and the Jetson Thor robotics platform, targeting AI Agent application development.
A new GitHub project called awesome-hermes-agent systematically organizes the entire Hermes Agent ecosystem of skills, tutorials, community projects, and best practices. For developers new to or wanting to deepen their Hermes knowledge, this is currently the most complete ecosystem navigation resource.
Dify, as an open-source LLM application development platform, has surpassed 50K GitHub stars, but still lags behind LangChain and CrewAI in agent orchestration and ecosystem depth. This article analyzes Dify's technical positioning, competitive advantages, and key challenges in 2026.
Hermes Agent v0.12 introduces Kanban task board, supporting multi-agent parallel collaboration: agents autonomously claim tasks from the board, execute in parallel, and hand off when blocked. One view replaces multiple terminals, marking Hermes key step from single-agent tool to multi-agent operating system.
NousResearch/hermes-agent added 108.1K stars in the first week of May 2026, topping GitHub monthly growth chart with over 127K total stars. This "self-evolving" agent framework is redefining the development paradigm for AI agents.
HermesOS announces the official opening of its free tier, allowing users to deploy persistent AI agent instances on the cloud without a credit card, supporting persistent memory, browser automation, terminal access, and scheduled tasks, with BYO API Key and no markup.
Alibaba's Qwen team releases Qwen-Scope, a complete sparse autoencoder toolkit enabling model internal feature visualization, output control, and long-tail data synthesis — bringing LLM interpretability from research into engineering.
The OpenClaw ecosystem saw explosive growth in May 2026: medical Skills reached 869, an official free 18-step tutorial launched for building from scratch, and a plugin system went live supporting on-demand installation. The AI Agent competition is shifting from frameworks to vertical domain ecosystems.
Community developers released QwenSeek-2B, distilled from Qwen3.5-2B using DeepSeek-V4 think token thought chains as teacher signals. Apache 2.0 licensed, runs on consumer GPUs.
Open-source AI coding tool OpenClaude releases v0.9.1, officially setting OpenAI Codex as the default provider and GPT-5.5 as the default model. This change stems from a public response by Sam Altman — if OpenAI officially supports it, make Codex the default option for OpenClaude.
Hermes Agent's memory system, built on a three-layer architecture (working memory / semantic memory / episodic memory), fundamentally restructures AI Agent memory mechanisms, correcting OpenClaw's information loss during memory compression. Ollama 0.21's local support makes this capability directly runnable on consumer-grade hardware.
TypeScript educator mattpocock's Claude Skills repo surges 31,091 stars in a week to nearly 60K, topping GitHub Trending. Content sourced directly from his personal .claude directory of battle-tested skills.
At Flower AI Summit 2026, Flower Labs launches Flower Agents and Project Kaya, proposing a three-element framework for good agents: context to understand tasks, access to acquire data, and control to execute operations.
Hermes Agent v0.12.0 released: 1,096 commits, 550 PRs, 213 contributors. Introduces Curator for autonomous skill library management, upgraded self-improvement review loop, default ComfyUI/Tou integration. Open source agent framework competition enters a new "self-evolution" stage.
GitHub Trending top project TradingAgents (67,992 stars) is a multi-agent LLM financial trading framework. Through collaboration of specialized agents — researchers, analysts, risk officers — it achieves fully automated workflows from information gathering to trading decisions.
Google Cloud announces 50+ MCP (Model Context Protocol) servers covering BigQuery, AlloyDB, Vertex AI and more, with built-in governance and observability. Enterprise Agent integration enters a standardized phase.
mattpocock/skills gained 31,091 stars in one week, reaching 59,309 total stars. This collection of practical skills extracted from a .claude directory is becoming the de facto standard for AI-assisted coding. This article analyzes its content structure, why it went viral, and how you can use it.
Ruflo, a multi-agent orchestration platform for Claude, added 4,321 stars this week, breaking 41K total. Its core selling points are self-learning swarm intelligence, native RAG integration, and native Claude Code/Codex integration—becoming a go-to orchestration layer for complex AI workflows.
TypeScript expert mattpocock's skills repository gained 31,000+ new stars on GitHub in one week, surpassing 59,000 total. The project provides .claude skill files from real engineers, covering TypeScript best practices, code review, architecture design, and more, setting a quality benchmark for the Claude Skills ecosystem.
Open Design officially open-sources Kanban Mode, supporting simultaneous orchestration of Hermes Agent, Cursor, DevinAI and other AI development tools in one interface, with real-time progress preview and parallel multi-prototype design. This marks a landmark upgrade in AI development experience from single-tool calling to multi-agent collaboration.
OpenAI launched the official Codex Skills Catalog repository on GitHub, providing a standardized skills directory for Codex CLI and API. Developers can directly reference pre-built skills or contribute custom ones, marking the shift of agent capabilities from "personal configs" to "community sharing."
browserbase open-sources Claude Agent SDK with web browsing tool, hitting 2,070 stars with 320 daily gains. Solves the standardization problem for Agent web interaction, providing out-of-the-box browser control, lowering the barrier for Web Agent development.
Alibaba Qwen releases Qwen3.6-35B-A3B, a 35B MoE model that activates only 3B parameters at inference, routing 8+1 experts from 256, supporting 262K native context extensible to 1M tokens, under Apache 2.0 license.
Equip Codex with a game art skill set — describe character animations and scene maps in natural language, and automatically get transparent PNG sprite sheets and engine-ready maps, ready for direct import into Godot and Unity.
Hermes Agent launches a desktop client supporting multi-agent parallel management, cross-platform session sync, multi-model provider switching, and built-in Telegram and Discord integration. Zero terminal operations, browser-level interface dramatically lowers the barrier to AI Agent management.
GitHub open-source project DeepClaude switches Claude Code's model backend to DeepSeek V4 Pro, preserving full agent loop, file I/O, bash, git capabilities, reducing per-million-token cost from $15 to $0.87, with real users reporting $0.55 for 31M tokens.
Google announced Gemma 4 will support fully on-device execution in React Native apps, without cloud APIs. This means millions of mobile developers can directly embed lightweight AI capabilities in iOS/Android apps — a triple win for privacy, latency, and cost.
OX Security discovered 200,000 MCP servers vulnerable to command execution. The STDIO transport layer lacks input sanitization. LangChain-ChatChat and Windsurf are both affected, with CVE-2026-30617 scoring 8.6.
A community developer has open-sourced 8,706 chain-of-thought records from Claude Opus 4.6 and 4.7, making it possible to train small models to "think before answering." Previously, accumulating enough training data required thousands of dollars in API costs — now the data is directly available.
New GitHub open-source project DeepClaude enables seamless replacement of Claude Code AI backend with DeepSeek V4 Pro, reducing cost from $15 to $0.87 per million tokens — a 17x drop. Just three environment variables to configure.
DigitalOcean releases Knowledge Bases, a fully managed RAG service integrating data ingestion, chunking, embedding, retrieval, and reranking. Includes RAG Playground, advanced reranking, two new open-source models, and MCP integration — enabling production-grade RAG apps in minutes.
Samsung Electronics has restarted its silicon carbide (SiC) foundry business, targeting mass production by 2028. SiC is widely regarded as the core material for next-generation power semiconductors, with broad applications in EVs, charging infrastructure, and industrial power.
A new Show HN project, Agent-desktop, topped today's leaderboard with 88 points — it's a CLI tool that enables AI agents to directly control the local desktop environment, marking the crossing of the last boundary from 'code world' to 'GUI world' for agents.
Flipbook, developed by a Samsung engineer (ex-OpenAI) and team, is gaining attention — enter a search term, and it generates dynamic illustrations pixel by pixel in real-time, with text also pixel-based, and any area clickable for deeper layers. This "infinite visual browser" may represent a new paradigm for AI interaction.
Pipecat is currently the most watched real-time voice AI agent framework on GitHub, featuring sub-200ms latency production-grade voice interaction. The community-compiled AI Agent ecosystem directory lists it as "the underlying framework that most production voice agents you have actually used are built on." This article analyzes its architecture, competitor comparison, and quick-start guide.
June AI releases Models 2026 Ultimate Lineup: open side includes GLM 5.1, DeepSeek V4 Pro/Flash, Kimi K2.6, Qwen3.5 397B, Gemma 4 31B; closed side includes GPT 5.5, Grok 4.1 Fast, Claude Opus 4.7, Gemini 3.1 Pro. Open models now directly compete with closed-source flagships in both numbers and comprehensive capability for the first time.
Inclusion AI has open-sourced LLaDA2.0-Uni, a diffusion LLM-based unified multimodal model that integrates vision understanding and image generation into a single architecture. Powered by a MoE backbone and SigLIP-VQ tokenizer, it generates images in just 8 steps and supports native interleaved reasoning — offering a brand-new inference paradigm for multimodal Agents.
TradingAgents surges to GitHub Trending second with 65k Stars, gaining 3,300+ daily. As multi-agent LLM financial trading framework, covers stocks, crypto sentiment analysis and quantitative trading, providing complete open-source quantitative research solution.
Hermes Agent introduces Kanban-based multi-Agent collaboration mode — Agents automatically claim tasks from the board, work in parallel, and hand off when blocked. Monitor all Agent progress from a single interface,告别 terminal switching.
Just 15 days after Anthropic launched Claude Design on April 17, the open-source alternative open-design is live. Supporting 11 CLIs with auto-detection, 31 Skills × 72 brand-level Design Systems built in, enabling Qwen, Kimi, GLM and other models to achieve Claude Design-level design capabilities.
Mercury Agent officially launches, positioning itself as the perfect union of OpenClaw and Hermes. The project merges OpenClaw's idea-driven approach with Hermes's energy-driven execution, delivering true control. The open-source agent framework landscape gets a new integrator.
The 100% open-source, Apache 2.0-licensed AI Perception Protocol has been released, positioning itself as the foundational "perception layer" protocol for AI Agents. If 2025 was the year of Agent framework explosion, 2026 will be the year of Agent perception infrastructure — this protocol could become the next LangChain-level infrastructure.
OpenSlide (open-slide) elevates slide generation from Markdown to React component level. Built-in agent skills support one-prompt full deck generation, fixed 1920×1080 canvas ensures output consistency, emerging as the open-source PowerPoint alternative in the AI era.
Meta releases Tuna-2, an encoder-free multimodal model using pixel embeddings instead of traditional vision encoders. Surpasses encoder-based models on fine-grained perception benchmarks, unifying understanding and generation. MoE architecture, open-source weights, offering a completely new technical route for multimodal AI.
OpenClaw 2026.5.2 released on May 3rd, featuring xAI Grok 4.3 model integration, improved plugin installation stability, gateway and agent hot path optimization, fixes for four major communication platforms (Discord/Slack/Telegram/WhatsApp), and polished TTS real-time voice calling. This is OpenClaw's third major update in three consecutive weeks.
Alibaba Qwen team officially open-sources Qwen-Scope, a complete sparse autoencoder (SAE) toolkit supporting feature extraction, data classification, and output control. This means developers can directly manipulate internal model features to control outputs without prompt engineering. This article breaks down Qwen-Scope three core capabilities, use cases, and getting-started guide.
Alibaba Tongyi team open-sources Qwen-Scope, a sparse autoencoder (SAE) based tool covering 7 Qwen3/3.5 models. Enables directional control of model output during inference, building classifiers and synthesizing long-tail samples with minimal seed data.
AMD launches Mini PC with Ryzen AI Max+ 395 processor, 128GB unified memory, full ROCm support, priced at $2K-$3K. Can run 200B parameter-level large language models locally, directly competing with NVIDIA DGX Spark. The cost-performance threshold for local AI inference is significantly lowered.
n8n-mcp reaches 19,257 GitHub stars, adding 264 daily. The project bridges Claude Desktop, Claude Code, Windsurf, and Cursor with n8n automation platform via MCP protocol. Users build and edit n8n workflows using natural language, eliminating manual node drag-and-drop.
Claude Opus 4.7 implemented an AlphaZero-style self-play pipeline from scratch on consumer hardware in just 3 hours, achieving 7/8 wins against the Pascal Pons professional solver in Connect Four. Other frontier Coding Agents did not exceed 2/8. This is a significant signal for autonomous agent reasoning.
The trending GitHub project agency-agents hits 9.2k Stars, offering 211 plug-and-play AI expert roles. Professional workflows can be injected into local frameworks like Hermes Agent and OpenClaw, delivering a qualitative leap in automation efficiency.
The combination of Chinese and American large models with open-source Agent frameworks like Hermes Agent and OpenClaw is increasing the efficiency of "bug hunting" and "incident response" by an order of magnitude. Harness Engineering has become the core skill for AI developers in 2026.
Tracer-Cloud/opensre gained 1,199 stars this week, reaching 4.3K total. An open-source AI SRE Agent framework supporting 60+ operations tool integrations, training agents in root cause analysis through synthetic incident scenarios, officially deployed on LangGraph Platform.
OpenAI open-sourced Privacy Filter on HuggingFace—a 1.5B parameter bidirectional token classification model for PII detection and masking. Only 50M active parameters, runs in browser, supports 128K long context.
OpenClaw releases version 2026.5.2 with xAI Grok 4.3 model support, more stable plugin installation and updates, Gateway and Agent core path performance optimization, fixes for Discord, Slack, Telegram, and WhatsApp messaging channels, plus improvements to TTS, real-time voice, and web search.
penMonoAgent is a local coding agent built on .NET 10 and llama.cpp—inference runs entirely locally with zero token fees and code never leaves your machine. Comes with 20 built-in tools and 5 specialized sub-agents, supports one-click Docker deployment.
A GitHub directory project gained 233 Stars and 274 Bookmarks in 24 hours, compiling the entire AI Agent ecosystem — frameworks, tools, real use cases, continuously updated. In the fragmented Agent landscape of 2026, this "panoramic map" is being revalued.
Google launches Gemma 4 Good Challenge with $200K prize pool across five tracks: health, education, global resilience, digital equity, and AI safety. The Gemma 4 family provides four sizes of open-source models as the technical foundation.
Poolside releases Laguna XS.2, a 33B total / 3B activated parameter MoE coding model, achieving 68.2% on SWE-bench Verified. Runs on a 36GB Mac, Apache 2.0 licensed, with one-click Ollama deployment.
Google's Gemma 4 26B A4B uses a 26B total parameter, 4B activated MoE architecture with 256K context window. Small total parameters but extremely small activation parameters mean fast inference speed and low resource consumption, making quality local AI on consumer hardware a reality.
Alibaba International Digital Commerce Group AIDC-AI open-sources Pixelle-Video, an AI fully automated short video engine supporting digital human broadcasting, image-to-video, and motion transfer, with 9.2K stars on GitHub providing end-to-end automation from script to final cut.
New generation Agent Harness frameworks claim 20x memory efficiency improvement, supporting massive parallel AI sessions without crashing. From JCode to Pi terminal agent, Harness is becoming the infrastructure layer for running AI agent clusters.
ByteDance open-sourced DeerFlow 2.0, reaching 50.7k GitHub Stars. The framework packages LangGraph state machine, sandbox security, memory management, and IM integration into an out-of-the-box solution suitable for individual developers, startup teams, and enterprise automation.
Maigret is an open-source OSINT tool that can search a target person's digital footprint across 3000+ websites with just a username. Ranked #2 on GitHub Trending today with 1,065 stars gained daily, total stars exceed 22K. Essential reconnaissance tool for security researchers and anti-fraud practitioners.
OpenADE, an open-source AI coding collaboration tool, completes a major update with GPT-5.5 and Codex integration, faster diff engine, while maintaining Claude compatibility — offering teams an open-source path around commercial editors.
GitNexus surged 5,376 stars this week on GitHub trending, reaching 34.6K total. Simply drag a GitHub repo URL or ZIP file to generate an interactive knowledge graph in your browser, with a built-in Graph RAG Agent for zero-server code exploration.
Ruflo surged to #2 on GitHub Trending with 1,258 stars in a single day. It equips Claude Code with agent swarm orchestration, self-learning memory, federated communication, and enterprise-grade security — turning a single AI assistant into a multi-agent collaboration system.
The complete AI to 3D pipeline in 2026 is now established: GPT Image 2 generates concept art → Hunyuan3D/Tripo converts to 3D mesh → MeshyAI/Blender for cleanup and rigging → Mixamo for auto-animation. A game-ready character in one afternoon, costing around $600.
The lambda community releases the hermes-agent-reasoning-traces dataset, first large-scale public release of agent reasoning trace data, with complete toolchain for parsing, analysis, visualization, and fine-tuning. This marks AI agents moving from "black box operation" to "observable, debuggable, optimizable" new phase.
Warp terminal announces open-source transition to Agentic development environment, gaining 3,400 daily GitHub stars, total surpassing 51K. Built-in AI Agent autonomously executes terminal tasks — terminal evolution from "manual input" to "intent-driven."
browserbase/skills hits GitHub Trending, providing Claude Code with a complete web automation plugin suite — from browser control to site debugging, cookie sync, and adversarial UI testing, elevating Agent web interaction from "can see" to "can operate."
obra/superpowers surpasses 175K GitHub stars, gaining over 1K daily. This 'Agentic Skills Framework' modularizes Claude Agent capabilities, marking a paradigm shift from conversational AI to composable agent workflows.
NVIDIA releases Dynamo project, rebuilding inference infrastructure specifically for Agentic AI. Traditional inference assumes independent requests, but agent coding sessions generate hundreds of API calls with redundant context computation. Dynamo achieves 2.7x performance improvement on same silicon via KV-aware routing and context reuse.
OpenClaw and Hermes Agent, the two most active open-source AI agent frameworks, are accelerating integration with Chinese model ecosystems. The latest versions support Kimi, GLM, DeepSeek, MiniMax and more domestic models, achieving task optimization through model routing and automatic degradation. Agent frameworks are evolving from "tools" into "operating systems."
TauricResearch TradingAgents surpasses 59K+ stars on GitHub, a multi-agent LLM framework for financial trading, gaining 2,100+ stars per day.
Archon pivots from a Python Agent builder to a TypeScript AI coding workflow engine, using YAML to define development pipelines. 17 preset workflows cover the full chain from issue fixing to PR creation, with support for remote triggering via Slack and Telegram.
Five open-source tools combined — DeepSeek V4 Flash for inference, Ollama for model management, OpenClaw for 24/7 daemon, Hermes for intelligent routing, Claude Code for coding — building a completely free local AI stack.
GitHub open-source project llm_wiki surges to 4.6k Stars — realizing Karpathy's vision of knowledge bases: instead of searching raw documents for every query, let LLMs pre-process to generate knowledge indexes. Personal knowledge management enters the native Agent era.
RTK is a Rust-based CLI proxy layer that intercepts and compresses command output from AI coding tools. A 30-minute Claude Code session drops from 118k tokens to 24k, with support for 12 AI tools.
Hermes Agent GitHub stars surpass 127K with over 100K new stars last month. Community reports of Agents running 100+ days, processing 10.8B tokens mark the shift from experimental AI Agent to production infrastructure.
Community developers demonstrate integrating Hermes Agent with Open Web UI via Docker for a free, fully-featured ChatGPT-style agent interface, ideal for personal and small-team self-hosted deployments.
NVIDIA released Nemotron 3 Nano Omni, a 30B hybrid MoE architecture unifying image, audio, video, and text into a single inference loop. Canonical partnership delivers Ubuntu inference snap for one-command deployment without complex integration.
CJackHwang/ds2api is a Go-based DeepSeek-compatible middleware focused on high-concurrency protocol adaptation, converting diverse web protocols into standardized DeepSeek API format. This week it gained 1,726 stars reaching 3,066 total, offering a unified API gateway solution for enterprises integrating multiple model sources.
Hermes Agent May 1 update natively integrates LM Studio, auto-discovering local models, on-demand loading, and intelligent routing inference strategies. Zero manual config needed — full agentic workflows run on consumer hardware.
NVIDIA launched NemoClaw and OpenShell at GTC 2026, injecting enterprise-grade security into the OpenClaw ecosystem. OpenShell provides a controlled sandbox environment, precisely managing what agents can access, share, and output — deployable with a single command.
Alibaba Qwen team releases Qwen-Scope, an open-source interpretability toolkit providing 81k SAE features across 64 layers of Qwen3.5-27B, enabling inference steering and mechanistic analysis.
DeepSeek releases V4 technical report, revealing two core innovations: replacing standard Attention with hybrid compressed attention system, and introducing Muon optimizer to accelerate training and improve stability. Also improved inter-layer connections for complex reasoning.
TradingAgents tops GitHub Trending. v0.2.4 adds native support for Qwen, DeepSeek, and GLM. This multi-agent framework models a real trading firm, lowering the barrier to AI-powered quantitative trading.
Hermes Agent v0.12.0 introduces Dashboard Profiles management page, allowing users to configure multi-agent architectures directly in the web interface. New ACP Server adapter enables standard protocol interaction, and Curator skill management system automatically optimizes skills.
Meta FAIR published a paper proposing to directly embed LLM safety, factuality, and reasoning capabilities during pretraining, rather than patching them in post-training. This innate design approach could change the model training paradigm, improving model quality from the root rather than relying on alignment post-processing.
GitHub project The Agency surpasses 50K Stars in under two weeks, featuring 147 specialized AI Agents across 12 departments. Not a prompt template collection, but a complete AI organizational structure.
MuleRun released a browser extension enabling agents to access users local browser tabs directly from the web app for real-time content extraction. Combined with newly launched e-commerce and creator data sources, MuleRun is building a complete data-to-output agent workflow.
Sim (simstudioai/sim) is an open-source AI agent orchestration platform with 28K stars. Positioned as the "central intelligence layer for AI labor," it supports building, deploying, and orchestrating multi-agent collaborative workflows. In the context of Sam Altman saying he "envies entrepreneurs in 2026," Sim represents the infrastructure for the "one-person company."
AgentKit releases a new version providing on-chain transaction, identity registration, and reputation system capabilities for AI Agents, natively compatible with OpenClaw, Hermes, ClawUp and other mainstream Agent runtimes. AI Agents officially upgrade from "execution tools" to "economic entities" — they can earn, trade, and build on-chain reputation.
dirac is a newly open-sourced AI coding agent built on the core design principle of "only make small changes" — avoiding the chaos caused by traditional AI agents that modify large numbers of files at once. By limiting the scope of each edit and generating more reviewable diffs, it significantly reduces AI programming failure rates and rollback costs.
HuggingFace open-sources ml-intern, an AI agent that reads papers, trains models, and auto-deploys. 7,774 stars in one week with 5,665 stars growth, maintained by HuggingFace core developers.
trycua/cua gained 15,450 stars in a week, providing cross-platform Computer Use Agent infrastructure. Supports macOS, Linux, Windows desktop control sandboxes, SDKs, and benchmarks — the open-source standard for desktop automation agents.
Nous Research Hermes Agent surpasses Anthropic Claude Code in GitHub stars, while launching Creative Suite with TouchDesigner and ComfyUI integration, marking open-source agents entering creative production.
Warp officially open-sourced its core codebase (MIT + AGPL v3), surpassing 49,000 GitHub stars in just 3 days and topping Trending. The project features a built-in AI coding Agent and supports external agents like Claude Code, Codex, and Gemini CLI, with OpenAI as founding sponsor.
Hugging Face open-sources ml-intern, an open-source AI ML engineer agent that reads papers, trains models, and deploys to Hub. Gained 6,388 stars this week, reaching 7,651 total.
TauricResearch/TradingAgents continues trending on GitHub with 58,369 stars, gaining 2,023 stars today. The project builds a multi-agent LLM financial trading framework, decomposing analysis, decision-making, and risk control into independent collaborating agents. Running alongside the Agent Arena S3 live trading competition, this project provides an open-source reference implementation for building autonomous trading agents.
TauricResearch/TradingAgents gained 2,023 stars this week, reaching 57,943 total, dominating GitHub Trending. This multi-agent LLM financial trading framework evolved from academic research to practical use, with v0.24 bringing key architectural and strategy updates.
Xiaomi open-sources MiMo-V2.5-ASR, natively supporting Mandarin, English, and Chinese dialects including Wu, Cantonese, Minnan, and Sichuanese. First open-source ASR model to reach practical dialect recognition levels.
Future AGI open-sources its complete Agent engineering and optimization platform (Apache 2.0), covering tracing, 50+ eval metrics, simulation engine, 18 built-in guardrails, 6 prompt optimization algorithms, supporting 50+ frameworks and 100+ LLM providers.
Hermes Agent v0.12.0 "The Curator Release" earned 930+ likes. But most users utilize less than 8% of its features. We dig into 134 slash commands to find the most useful hidden capabilities, from named Personas to cross-session memory.
trycua/cua is an open-source project with 15K+ stars that provides complete Computer-Use Agent infrastructure: cross-platform sandboxes (macOS/Linux/Windows), SDKs, and evaluation benchmarks. Unlike OpenAI Codex Computer Use, cua is fully open-source, supporting custom training and evaluation.
Mception is an open-source MCP Server audit tool with 46 built-in security rules, detecting tool poisoning, rug pulls, RCE, and supply-chain attacks. Outputs SARIF format reports, requires no API key. As the MCP ecosystem grows exponentially, this is an infrastructure-level project for agent security.
Beads hits GitHub Trending, providing AI coding assistants with a structured task tracking system based on dependency graphs. Using DAGs to manage multi-agent tasks, solving context loss and task chaos — like giving AI Agents a project manager.
Google releases Gemini Embedding 2, the industry's first fully multimodal embedding model based on Gemini architecture, supporting unified encoding of 100+ languages across text, image, and audio into a single vector space. Available via Gemini API and Vertex AI preview. This means semantic precision for text-to-image and image-to-image search receives a generational upgrade, and multimodal fusion for RAG knowledge bases becomes possible.
Moonshot AI open-sources FlashKDA, a CUTLASS-based high-performance kernel for Kimi Delta Attention, achieving 1.72×–2.22× forward pass speedup on H20, with native variable-length batching support and drop-in compatibility as a flash-linear-attention backend.
OpenClaw and Hermes Agent now fully support integrating domestic Chinese AI models including GLM-5.1, Kimi K2.6, and DeepSeek V4 Pro. This guide covers configuration, pricing comparison, and best practices for building cost-effective AI Agent workflows.
OpenClaw released v2026.4.29 with Active-Run Steering for group chat agent experience, memory system overhaul, and expanded model provider support. This is a key step in OpenClaw's evolution from personal assistant to enterprise-grade AI gateway.
OpenRouter debuts anonymous model Owl Alpha with 105K token context, 262K max output, native tool calling support — currently free. Designed for agent workloads, compatible with Claude Code and OpenClaw, but all prompts and outputs are logged by the provider for model improvement.
Hermes Agent launches Labyrinth visualization tool, mapping all Agent runtime states — prompts, tool calls, failures, model switches, memory, and sub-agents — into a traceable graph. 63K views, 203 bookmarks, community demand for Agent observability far exceeds expectations.
PKU DCAI team, together with Kuaishou Kling and other institutions, open-sources OpenWorldLib — a unified, extensible world model inference framework that standardizes interactive video generation, 3D reconstruction, VLA embodied control, and multimodal reasoning through a single interface, ending the fragmentation of disjointed APIs and redundant development.
GitHub open-sources Spec-Kit, promoting a "spec first, code second" development paradigm. Through a Specify → Plan → Tasks → Implement four-step process, it transforms vague requirements into structured documents, significantly improving output quality from AI coding tools like Claude Code, Cursor, and Copilot.
Open source project CutClaw achieves end-to-end AI Agent video editing: input raw footage and music, the Agent plans every cut like a screenwriter, automatically completing material selection, rhythm matching, and editing output.
Qwen Code is Alibaba Tongyi team open-source terminal AI coding agent with 24.1k stars, 5,649 commits, supporting multi-protocol APIs, SubAgent parallelism, and skill plugin system — delivering Claude Code-like experience with full self-deployment capability.
Skill Retrieval Augmentation (SRA) framework goes open source with SRA-Bench containing 26,262 skills, 636 gold skills. It provides an efficient skill retrieval augmentation method for AI agents, enabling them to quickly find the most appropriate existing skills when facing new tasks.
The awesome-codex-skills repository maintained by ComposioHQ topped GitHub Trending today with 1,177 new stars in a single day. The repository systematically curates practical skills for Codex CLI and API, reflecting that Agent skills ecosystem is becoming a new hotspot in developer infrastructure.
Andrej Karpathy's observations on LLM coding pitfalls were packaged into a CLAUDE.md skill file, reaching 102K stars on GitHub in one week. This article analyzes what problem it solves, its core rules, and how to apply it to your workflow.
AgentField project proposes the concept of "Kubernetes for AI Agents", providing a complete Agent control plane. Not another Agent framework, but core infrastructure solving large-scale deployment.
Microsoft Research Asia open-sources Agent Lightning framework, achieving zero-intrusion reinforcement learning optimization for the first time—any AI agent can be trained and optimized without modifying its internal code.
On April 29, NVIDIA released the Nemotron 3 Nano Omni open-source multimodal model, featuring deep FP8 inference optimization, compatibility with RTX 5090 consumer GPUs and Jetson Thor robotics platforms, dramatically lowering the deployment barrier for AI agents.
After Anthropic removed Claude Code from the $20 Pro plan, the community quickly launched OpenClaude—an open-source alternative supporting any model, no subscription limits, running locally—becoming a developer favorite.
GitHub trending project TradingAgents releases v0.2.4 with structured agents, checkpoint recovery, and memory logging. This multi-agent LLM financial trading framework has accumulated 56,534 stars with 386 daily new stars.
The newly appeared Mercury Agent project on GitHub is seen by the community as the ultimate combined upgrade of Hermes and OpenClaw, making significant improvements for local Agent background running. Four core mechanisms address pain points like permission失控 and cost overruns.
OpenClaw releases v2026.4.27, with the highlight being the integration of OpenAI Codex Computer Use, enabling AI agents to directly control user desktops. Desktop control experience upgraded, more communication channels added, and overall startup speed optimized.
Alibaba Qwen team open-sources Qwen-Scope, using sparse autoencoders (SAE) to provide an internal "X-ray" view of Qwen3/3.5 models, decomposing internal activations into human-readable semantic switches, pushing LLMs from black box toward transparency.
Qwen team open-sources FlashQLA, high-performance linear attention kernels built on TileLang, delivering 2-3× forward inference speedup and 2× backward speedup, optimized for Agent AI on personal devices.
OpenAI releases Symphony, an orchestration framework connecting task trackers like GitHub Issues to Codex Agents, enabling every issue to auto-assign an Agent for processing while humans focus on review and direction.
free-claude-code tops GitHub Trending this week with 16,154 new stars (18,791 total). The project enables free Claude Code usage in terminal, VS Code, and Discord, with voice interaction support.
Hermes Agent introduces Curator, an automated skill library governance tool that merges overlapping skills, removes stale entries, and demotes granular skills to templates, solving the core pain point of Agent skill bloat.
ComposioHQ maintains awesome-codex-skills, a curated collection of practical Codex skills covering automation workflows across Codex CLI and API, gaining 1,177 stars daily.
Craft.do team open-sourced Craft Agents, integrating Claude Agent SDK and Pi SDK into a unified desktop agent workbench with multi-session inbox, MCP connections, multi-LLM provider support, and remote server mode under Apache 2.0 license.
ZhuLinsen/daily_stock_analysis uses LLM as its core engine to aggregate multi-source market data, news and sentiment, automatically generating daily A/H/US stock decision dashboards pushed to WeCom, Feishu, Telegram and more—all at zero server cost.
GitNexus is a fully browser-based code knowledge graph generator. Drop in a GitHub repo to get an interactive graph with a built-in Graph RAG Agent. 33.4k stars.
TypeScript developer mattpocock open-sources his .claude directory skills collection, gaining 7,280 stars daily to reach 44.8k total, providing real engineering scenario skill templates for Claude Code users.
Microsoft open-sources VibeVoice on GitHub, integrating speech recognition, synthesis, and voice cloning with vLLM inference plugin and Apple Silicon support, reaching 45.7k stars.
NVIDIA releases Nemotron 3 Nano Omni, an open-source omni-model optimized for Hopper/Blackwell FP8 inference, compatible with RTX 5090 consumer GPUs, achieving 9x efficiency gains—marking the AI Agent competition shift toward efficiency.
OpenClaw 4.24 release sets DeepSeek V4 Flash as the default model for new users, fully打通 real-time voice calling with complete Agent capabilities, and adds pixel-level browser clicking. The update simultaneously reduces costs and expands interaction methods.
Rust-based Warp terminal, an agentic development environment, explodes on GitHub Trending with 12,822 daily stars, reaching 44.2k total — signaling the terminal transformation from command executor to AI collaboration platform.
ComposioHQ/awesome-codex-skills, with 4700 stars and 1180 daily growth, curates practical Codex CLI and API skills covering DevOps, testing, code review, security scenarios, serving as a practical reference library for AI programming automation.
PurpleAILAB/Decepticon is an open-source multi-Agent autonomous red team testing platform that simulates AI hackers executing the full kill chain from reconnaissance to privilege escalation in a Kali sandbox, providing a new tool for AI system security assessment.
The Agency project broke through 50K GitHub stars in two weeks, offering 147 specialized AI Agents across 12 departments, from engineering to marketing, QA to spatial computing, upgrading AI Agents from single-point tools to complete organizational structures.
jcode is a Rust-based Coding Agent Harness with 2811 commits and rapid iteration, positioning itself as a unified orchestration layer between Claude Code and Codex, supporting skill plugins and self-development capabilities.
obra/superpowers has become GitHub hottest AI coding project with 173K stars, providing a unified Skills framework across Claude Code, Codex, Cursor, and Copilot, defining the capability extension standard for AI coding tools.
Anything Analyzer is a TypeScript protocol analysis tool integrating browser packet capture, MITM proxy, fingerprint spoofing, and AI analysis, with MCP Server for direct AI Agent/IDE integration. 2,055 stars since mid-April launch.
FalkorDB released GraphRAG SDK 1.0, ranking first against 8 systems on its GraphRAG-Bench. Graph-structured RAG shows clear advantages over vector retrieval in multi-hop reasoning and cost predictability.
Higgsfield released its MCP server, connecting top multimodal models like Seedance 2.0 and GPT Image 2.0 to AI Agent workflows. Supports OpenClaw, Hermes Agent, and any MCP-compliant client for end-to-end content automation.
Moonshot AI open-sourced Kimi K2.6, a 1T parameter MoE model with 256K context window. It approaches GPT-5.4 and Opus 4.7 on LiveBench and Terminal-Bench at roughly one-sixth the cost.
Graphify gained 37,989 GitHub stars within a week of launch. It converts code, docs, and papers into queryable knowledge graphs, plugging directly into Claude Code, Cursor, and Codex as a skill.
KellerJordan modded-nanogpt is a community-driven LLM training speedrun, compressing GPT-2-level model training from 45 minutes to 90 seconds through dozens of techniques including rotary embeddings, Muon optimizer, and FP8 matmul.
Upstash open-sourced Context7, providing real-time code documentation for LLMs and AI code editors. Solves hallucinated APIs and outdated examples from stale training data. 54K stars, MIT licensed, one-click MCP integration.
Microsoft open-sources Playwright MCP server, exposing browser automation capabilities through Model Context Protocol. Uses accessibility trees instead of screenshots — no vision model needed. 32K stars, Apache-2.0 licensed.
Fission AI open-sourced OpenSpec, a lightweight spec layer for AI coding assistants. Through proposal/specs/design/tasks artifacts, developers and AI align on goals before writing code. 44K stars, supports 25+ platforms.
OpenClaw releases v2026.4.26 featuring Google Live Talk real-time voice, comprehensive Ollama local model refactor, one-click Claude Code config migration, and Matrix E2EE. The release garnered 1,837 likes and 340K views on the official announcement.
AIDC-AI/Pixelle-Video is an open-source AI fully automated short video engine supporting one-stop generation from text script to finished video. 7600 stars, 1200+ forks, providing a locally deployable automation solution for short video creators.
mksglu/context-mode open-sources a context window optimization plugin that reduces AI coding agent context consumption by 98% through sandboxing tool output. 11k stars, supporting 14 coding platforms.
trycua/cua open-sources complete infrastructure for Computer-Use Agents, including sandbox environments, SDK, and multi-platform benchmarks. 15.1k stars, 3200+ commits, supporting macOS/Linux/Windows desktop control.
Zilliz open-sources claude-context, a code search MCP server that uses vector retrieval to make entire codebases the context for Claude Code and other coding agents. 10.2k stars, 3700+ growth this week, providing a scalable approach for Agent-assisted programming on large projects.
HuggingFace releases ml-intern, an open-source AI ML engineer capable of autonomously reading papers, designing experiments, training and publishing ML models. 6400+ new stars this week, marking a new Agent-driven era in automated machine learning.
ace-step-ui provides a professional Web UI for the open-source ACE-Step 1.5 music generation model. Local execution, unlimited generation, zero cost, directly competing with paid services like Suno. +162 stars today, community interest growing.
A roundup of noteworthy AI open-source projects on GitHub, covering fine-tuning, agents, and multimodal applications.
LangChain v1.0 brings a more stable API and a redesigned agent system for production AI applications.
davila7/claude-code-templates provides CLI tools for quick configuration, management, and monitoring of Claude Code runtime environments. +346 stars today, surpassing 26K total, one of the most popular third-party Claude Code tools.
ds2api is a high-performance full-stack middleware converting DeepSeek client protocols to universal API formats. Supports multi-account auto-rotation, compiled binaries, Vercel Serverless and Docker deployment, compatible with Google/Claude/OpenAI API formats. +658 stars today.
TypeScript author Matt Pocock open-sourced his .claude skills directory with AI configs for real engineering scenarios. 7,321 stars on day one, 38K total, topping GitHub Trending.
GitNexus is a client-side code knowledge graph generator that runs entirely in the browser. Drop in a GitHub repo or ZIP file, and get an interactive knowledge graph with built-in Graph RAG Agent. No server deployment needed, 1,607 stars added today, surpassing 32K total stars.
This week on GitHub Trending, multiple AI projects deserve attention: Matt Pocock open-sources a 39K-star skills library, Voice-Pro integrates a full voice cloning stack, pentest-ai-agents provides 28 Claude pentesting sub-agents. Developer tool ecosystem continues heating up.
Voice-Pro integrates ElevenLabs and Descript core features into one open-source tool, supporting zero-shot voice cloning, Whisper transcription, 100+ language dubbing, running locally via Gradio WebUI. 1,550 bookmarks and 55K views on X.
Euphony is an open-source tool that converts Codex session logs and chat data into visual interfaces, supporting translation, filtering, and editing. 4,336 likes and 2,842 bookmarks on X, with over 380K views.
Microsoft open sources VibeVoice speech AI model family, supporting 60-minute audio transcription without chunking, built-in speaker diarization, 50+ languages. GitHub hits 44.7K stars, runs locally at zero cost.
