Trending AI projects

A serving stack that runs GLM-5.3-Flash (125 GB) on a single RTX 3090 24 GB by splitting its 12,096 routed experts across VRAM, RAM, and NVMe, delivering up to ~28 tok/s decode with near-exact output.

ACRYL is an early-stage (v0.2.5, MIT-licensed) "Agentic Development Environment" that keeps your project, context, and work persistent while letting any coding agent (Claude Code, Codex, etc.) plug in as a replaceable worker. It features a canvas-based UI, a catalog of 33 agents, agent-driven app control with per-action approval, a plugin system built on Cordis, shareable "Blend" environments, and three interchangeable surfaces (desktop, local web, terminal). It builds on a pinned, unmodified DeepSeek Harness checkout as its capability substrate.

Awesome REA Cases is an unofficial, actively maintained collection of 817 reverse-engineering projects, experiments, and ideas where AI tools like Claude and REA help revive old devices, rebuild apps, preserve games, and understand hidden behavior. Each case links to public evidence and notes limits, updated six times daily.

Muesli is a native macOS app for on‑device dictation and meeting transcription, with optional cloud LLM summarisation and an agent‑friendly CLI.

A desktop application for agent builders to prototype, trace, debug, and evaluate AI agents with local-first data management.

NoteGen is a local‑first Markdown note‑taking app that captures raw thoughts (text, voice, screenshots, links, etc.) and later uses LLMs to organise them into structured notes, with optional vector‑based knowledge‑base search, canvas diagrams, and extensible plugins.

A comprehensive library for post-training foundation models using techniques like SFT, DPO, and GRPO to align them with human preferences.

DeepEP is a high‑performance communication library for Mixture‑of‑Experts training and inference. It provides JIT‑compiled all‑to‑all dispatch/combine kernels, a unified EPBuffer API, support for expert replication (weight prefetch & gradient reduction), and experimental pipeline/remote‑memory primitives. Designed for NVIDIA Hopper (and Ascend) GPUs with NVLink/RDMA, it lets large MoE models scale efficiently across multiple GPUs.

A Claude‑compatible “skill” that provides a markdown guide for writing correct Seedance 2.0 video‑generation prompts (ByteDance’s multimodal model). Install by copying the markdown into `~/.claude/skills` or via `npx skills add`. The guide covers constraints, reference syntax, camera language, prompt structures, and templates for ads, dramas, MVs, education, etc., and is sourced from ByteDance’s official docs. MIT‑licensed.

A high-efficiency pretraining framework that allows users to pretrain 1B parameter foundation models from scratch with significantly reduced compute and data requirements.

The project is a curated collection of prompts for Google's Nano Banana Pro AI image model, with a GitHub repository and a web gallery. It contains thousands of prompts organized by categories, featured examples with images, and supports Raycast dynamic arguments. The README provides statistics, language versions, and links to a more visual gallery.

FileCodeBox 是一个自托管的轻量文件分享工具，核心流程是“上传→获得口令→凭码取件”，无需注册。支持按时间/次数/永久三种失效策略，存储可对接本地、S3、OneDrive、WebDAV 等后端，并提供管理后台。技术栈为 FastAPI + Vue 3 + SQLite，Docker 一条命令即可部署，同时支持 MCP 协议供 AI 客户端调用。开源协议为 LGPL-3.0。

A personal, technically deep blog series (in English and Chinese) about building and optimizing the software infrastructure for modern AI — especially reinforcement-learning training systems and inference engines — written by an active SGLang community contributor who cares about making research conclusions trustworthy by making the underlying infrastructure correct.

A stealth browser automation server for MCP-compatible AI agents that bypasses anti-bot protections like Cloudflare while providing deep browser control and network inspection.

A secure, Rust-based Python sandbox designed for the high-speed execution of AI-generated code without the latency of containers.

Dockerized FastAPI server that wraps the open‑source Kokoro‑82M TTS model, offering an OpenAI‑compatible speech endpoint, streaming, multilingual voices, voice‑mixing, inline SSML tags, custom voice‑tuning and a web UI. Runs on CPU, NVIDIA/AMD GPUs or Apple Silicon via Docker or uv.

A library of over 230 standardized, portable skill files that extend the capabilities of AI coding agents like Claude Code and Cursor for complex engineering, design, and business tasks.

A curated library of AI-agent instruction packs that teach Claude Code the submission conventions of 744 academic journals across 11 disciplines, with deep packs for 12 flagship journals and breadth packs for hundreds more, maintained by Stanford REAP's CoPaper.AI.

A library for efficient similarity search and clustering of dense vectors, capable of handling billion-scale datasets using CPU and GPU acceleration.

A token-efficient MCP server that enables AI agents to interact with the Cloudflare API using a code-execution pattern to avoid prompt bloat.

EchoBird is a unified management dashboard for AI coding tools and LLMs that allows users to configure model settings once and apply them across multiple CLI and desktop applications, while providing smart routing and multi-account switching.

ExploitGym is a large-scale benchmark of real-world vulnerabilities used to evaluate the ability of AI agents to develop functional security exploits.

Power Platform Skills is a Microsoft‑maintained plugin marketplace that adds AI‑driven “skills” for GitHub Copilot CLI and Claude Code. The plugins let an LLM agent create, edit, and deploy Power Platform artefacts—Power Pages sites, model‑driven apps, Canvas apps, Power Automate flows, process‑mining analytics, and more—by invoking the Power Platform CLI, Node scripts, or MCP servers. Install via a one‑line script, then use natural‑language commands like `/app-builder` or `/analytics` to have the agent generate tables, forms, code apps, mobile apps, or run analytics. The repo includes auto‑update, permission‑control guides, telemetry opt‑out, and full development instructions. It is a genuine AI‑enabled tooling project for Power Platform development.

js‑reverse‑mcp is a Node.js MCP server that lets LLM agents debug and reverse‑engineer JavaScript in Chrome. It wraps DevTools data (scripts, breakpoints, network, WebSocket, storage) into 24 concise tools designed for AI‑driven reasoning, and includes optional stealth layers (Patchright fork and CloakBrowser) for anti‑bot sites.

Muteki is an open‑source, Python‑based framework that orchestrates a swarm of LLM‑driven agents (Claude, Codex, Cursor, etc.) to automatically solve CTF and penetration‑testing challenges. It provides a web UI for configuring agents, running tasks locally or in Docker, tracking evidence, and generating write‑ups.

A cloud-native data pipeline orchestrator for managing the development and maintenance of data assets like machine learning models and tables with integrated lineage and observability.

Percho is an Electron‑based desktop GUI that wraps the official Pi coding‑agent SDK, offering multi‑session chat, tool‑call approvals, UI plugins, custom themes, and support for many LLM providers.

An open-source AI tool that automatically converts long videos and live stream replays into viral vertical short-form videos using local transcription and highlight detection.

A set of modular AI agent skills that implement a professional engineering workflow, moving project state from chat sessions into repository files for persistence and team collaboration.

YouDub WebUI is an open‑source, plugin‑based web application that localizes videos: it transcribes speech with Whisper, translates via an OpenAI‑compatible API, and generates sentence‑by‑sentence dubbed audio with VoxCPM2, optionally keeping background music. It can also burn translated subtitles onto the video. Built on the Cordis plugin framework, every component (ASR, translation, TTS, alignment, UI) is a replaceable plugin, allowing custom models or services. The repo provides scripts to install Node and Python dependencies, auto‑download default models, and run a host backend plus a Next.js front‑end. Supports MP4/MOV/MKV/WEBM up to 4 GiB, 10 min, and runs on CPU, CUDA, Apple MPS, or AMD ROCm. Licensed under Apache 2.0, with a QQ community for support.

A collection of agent skills and structured instructions that extend AI agent capabilities for specialized scientific research tasks in genomics, biology, and chemistry.

An agent-powered vulnerability scanner for large-scale repositories that uses AI to find and verify deep-seated security issues.

Vendor‑neutral Claude Fable 5 system prompt adapted for use across multiple frontier LLMs.

A DSH plugin that drives DeepSeek agents using your authenticated web session instead of an API key, allowing free access to advanced model features.

Humming is a lightweight, high-performance JIT-compiled library for quantized matrix multiplication, enabling efficient inference across a wide range of low-bit precision formats on NVIDIA GPUs.

An AI-powered toolset for Spine 2D that automates character rigging, asset positioning, and animation generation from raw images.

Kong is a cloud-native API and AI gateway that centralizes routing, security, and orchestration for microservices, LLMs, and Model Context Protocol (MCP) traffic.

Zero is a Go‑based terminal AI coding agent that lets you pick any LLM (OpenAI, Anthropic, local Ollama, etc.) to inspect, edit, and run code in your repository. It offers a rich TUI, a scriptable `zero exec` mode, sandboxed permissions, persistent local sessions, and extensibility via plugins and MCP servers. Install via npm, a one‑line script, or by building from source, and keep full control over what files or commands the agent can access.

DeepTeam is an open-source LLM red teaming framework that simulates adversarial attacks like jailbreaking and prompt injection to uncover vulnerabilities such as bias, data leakage, and SQL injection in AI agents, RAG pipelines, and chatbots. It runs locally, offers 50+ pre-built vulnerabilities, 20+ attack methods, and production guardrails.

A library of modular skills for AI agents that enables financial analysis, market data retrieval, and social research using an open standard for agent interoperability.

An open-source pipeline for rewriting AI-generated text into natural human prose using either a structure-first training pipeline or a multi-hop translation chain.

OpenEvolve is an open‑source Python library that combines large‑language‑model prompting with a MAP‑Elites evolutionary algorithm (island parallelism) to automatically rewrite, optimize, or invent code across many languages. Users supply an initial program and an evaluator; the system mutates the code via LLMs, evaluates candidates, and maintains a diverse Pareto‑optimal archive. It supports multiple LLM back‑ends (OpenAI, Gemini, Claude, local Ollama, etc.), offers deterministic reproducibility, extensible hooks, and can be run via CLI, Docker, or as a Python module. Example use‑cases include GPU kernel tuning, novel algorithm discovery (e.g., adaptive sorting, circle packing), scientific computing, and competitive‑programming solution generation.

A native Android companion and desktop CLI that allows you to access and control your self-hosted Hermes agent across your devices with voice, chat, and system-level tools.

A cross-platform desktop assistant that uses AI to provide real-time answers by analyzing screen captures or listening to live audio conversations during interviews and exams.

A unified command-line tool and TUI for managing configurations, API providers, MCP servers, and session history across multiple AI coding assistants like Claude Code, Codex, and Gemini.

A set of 36 Claude‑Skill packages that turn Anthropic’s Claude LLM into a domain‑expert for major GRC frameworks (ISO 27001, SOC 2, FedRAMP, GDPR, HIPAA, etc.). Install the `.skill` files, mention a relevant term, and Claude automatically provides gap analyses, policy drafts, control mappings and technical guidance, all with up‑to‑date regulatory citations.

A multimodal AI-powered CAPTCHA solver that automatically detects and solves various web challenges via a browser extension or API.

A comprehensive framework for training and deploying humanoid robot whole-body controllers, featuring the GEAR-SONIC foundation model for natural motion and real-time teleoperation.