Trending AI projects

Files and templates extracted from a 72-hour live-streamed build where three xAI engineers launched a product using coding agents — including agent rules, 40 documented failures, 69 role definitions, 9 role playbooks, and a PR template that forces proof-before-merge.

A Windows-based desktop AI companion featuring a Live2D character that provides interactive support for chatting, work, coding, and learning through a robust agentic engine.

A set of SKILL-formatted prompts and auditing tools designed to remove robotic patterns from AI-generated Chinese text and mimic human writing styles.

A declarative infographic visualization engine that uses AI-friendly syntax to transform text descriptions into high-quality, professional infographics.

UniFace is a Python library that bundles many state‑of‑the‑art face‑analysis models (detection, recognition, landmarks, gaze, demographics, emotion, anti‑spoofing, etc.) behind a unified API. Install via pip, run a single `FaceAnalyzer` call and get rich face objects; models download automatically and work on CPU, Apple Silicon or CUDA.

A multi-agent orchestration tool that turns a single sentence into a coordinated team of specialized AI roles, running parallel DAG workflows with quality checks, supporting many LLM providers (including key-free reuse of existing coding CLIs), and offering both CLI and visual web interfaces.

Ouroboros is an open‑source Agent‑OS that turns vague AI coding requests into reproducible, verified code by running a Socratic interview, locking the resulting specification (Seed) in an immutable ledger, and automatically generating, testing, and iterating with any supported LLM.

A real-time audio translation tool for Windows that captures system audio and microphone input to provide live translated subtitles via an LLM-powered overlay.

LongLive is NVIDIA’s open‑source suite for very long video generation and long‑context robot‑action modeling. It contains four self‑contained projects—Long‑WAM (world‑action models with streaming visual memory), LongLive‑Plug (once‑for‑all LoRA distillation), LongLive 2.0 (NVFP4‑based high‑performance training/inference infrastructure), and LongLive 1.0 (real‑time interactive text‑to‑video). The repo provides code, training recipes, inference pipelines, and pretrained checkpoints (hosted on HuggingFace) for each component, all under Apache‑2.0.

A real-time, streamable diffusion framework for generating infinite-length expressive portrait animations from a single reference image and a driving video.

A curated list of production-ready Claude Code plugins — the key one, connect-apps, lets Claude take real actions (email, Slack, GitHub, etc.) across 1,000+ apps via an MCP Gateway. Includes setup instructions, a categorized catalog of 40+ plugins, and contribution guidelines.

GoModel is an open‑source Go‑written AI gateway that provides a single OpenAI‑compatible endpoint for dozens of LLM providers. It adds caching, cost‑tracking, budgets, rate‑limits, fail‑over, observability (Prometheus/OpenTelemetry), a web dashboard, and extensible plugins, helping teams control spend and reliability while keeping existing SDKs unchanged.

...

RoboDojo is an open‑source benchmark that evaluates robot‑manipulation policies in both Isaac Sim and on real robots. It provides 42 simulated and 18 real tasks, parallel simulation, and a one‑command workflow that talks to policies hosted via XPolicyLab, producing scores for a public leaderboard.

Repo2RLEnv automatically converts real code repositories into verifiable practice tasks for AI coding agents. It mines existing material (merged pull requests, commits, security advisories, library functions, terminal recordings) and turns each into a standard Harbor task folder with a private verifier and reference solution. Quality is enforced through controls (oracle scores 1, no-op scores 0) and a review-and-repair loop. It offers 23 generators across five task categories, plus Tasksmith — an agent that adaptively builds environments from individual pull requests. The project publishes 21 datasets with 1,930 tasks on the Hugging Face Hub.

AgentConnect is an open‑source, self‑hosted platform that lets teams run many AI agents (Claude Code, OpenAI Codex, Grok Build, etc.) inside Slack, Discord, Telegram, Google Chat, GitHub, GitLab, Linear and other tools. Agents can call each other, keep memory, be routed by Jev, and act on webhooks or schedules. The system consists of Daemons (run agents), an optional Relay, and a Control‑Plane UI for config, auth and observability. It ships via Docker‑Compose for local use and a Helm chart for Kubernetes, all under Apache‑2.0.

Threejs‑Awesome‑Graphics‑Agent‑Skills is an npm package that bundles ready‑to‑use, high‑quality Three.js code snippets (called “skills”) for advanced visual effects—procedural materials, realistic oceans, volumetric clouds, ray‑marched space scenes, and more. It installs these snippets into the directory layout expected by AI coding assistants (Codex, Claude Code, Cursor, etc.), letting the agents copy‑paste proven implementations when a user asks for “awesome graphics”. The repo includes an example gallery, validation scripts, and a deterministic capture pipeline, making it a practical resource for developers and AI‑augmented workflows.

ONNX is an open-source ecosystem that provides a standard format for representing AI models, covering both deep learning and traditional ML, using an extensible computation graph with built-in operators and data types. It focuses on inference and enables interoperability across many frameworks, tools, and hardware.

AI News Radar is an automated, self-hosted news aggregator that filters, merges, and prioritizes AI-related updates from multiple sources into a clean, daily digest.

海鸥 2.0 是一个为 Claude Code 安装“暴躁安全研究员”人格的配置包，包含大量对话示例和术语映射，支持跨平台安装/卸载，自动备份原配置。

fx0 is an open‑source Vercel‑hosted app that turns a textual page description into a single‑file HTML design using the libfx LLM‑agent framework. It features durable sessions via Vercel Queues, optional free‑trial prompts, a set of design tools (web fetch, search, edit, sub‑agents), and versioned design outputs displayed in a sandboxed preview.

A secure, enterprise-grade infrastructure framework for deploying AI agent applications on Microsoft Azure using Infrastructure-as-Code.

A GitHub‑only tool that daily crawls selected arXiv categories, uses a Large Language Model (default DeepSeek) to generate Chinese summaries, and publishes the results as a static site on GitHub Pages. It requires only a fork, a few secrets/variables, and runs via GitHub Actions—no servers needed.

A bimanual robotic manipulation platform that provides a scalable data generator, a large-scale trajectory dataset, and a benchmark for evaluating robust dual-arm robot policies.

A video cut detection and analysis tool that automatically identifies scene changes and can split videos into individual scenes using content-aware algorithms.

TheRock is the open-source build and release system for HIP and ROCm, providing a modular workflow to integrate and distribute ROCm components, including PyTorch and JAX.

An AI-powered tool that automatically generates and maintains hierarchical, dependency-aware documentation for large codebases, supporting incremental updates and artifact documentation.

A Pi extension that enables compatibility with Anthropic Claude Pro/Max OAuth subscriptions, allowing users to leverage their existing subscriptions within the Pi agent framework.

A prompt compression toolkit that removes non-essential tokens from LLM prompts to reduce costs, bypass token limits, and improve long-context performance.

SkillNet is an open‑source platform that lets you discover, generate, evaluate, analyse and route reusable AI‑agent skills. It offers a public catalogue (500 K+ GitHub‑hosted skills), a Python SDK/CLI with six operations (search, download, create, evaluate, analyze, route), a local web UI, and integration points for agents via an MCP server. The system builds a graph of skill relationships and uses LLM‑powered routing (Claude/Codex) to pick the best skills for a task, providing safety/maintainability scores and reasoning.

Mnemon is a cross‑platform CLI that adds persistent, graph‑based memory to LLM agents. The host LLM supervises what to remember, link, or forget, while Mnemon handles deterministic storage, decay, deduplication, and optional vector search. It integrates with many AI coding tools (Claude Code, Codex, Cursor, etc.) via one‑click setup, requiring only a single local binary and no extra API keys.

OpenCluely is an open-source desktop assistant that provides real-time AI help during technical interviews via a stealth overlay that is invisible to screen-sharing software.

A community directory of Claude Code skills — ready-made capabilities for tasks like testing, debugging, PDF extraction, and security that you can install and use with Claude.

A centralized framework for materials science and quantum chemistry that provides high-speed, machine-learned interatomic potentials for atomistic simulations.

An open, local runtime and hardware-agnostic control framework for industrial robotics that integrates AI perception, motion planning, and real-time execution.

GitReverse is a web app that turns any public GitHub repository into a short, conversational prompt for code‑generation models. It fetches repo metadata, a shallow file tree, and the README, then uses a configurable LLM (Grok, Gemini, Azure, etc.) to synthesize a single prompt. Built with Next.js, React, TypeScript, and Tailwind, it supports optional Supabase caching and Stripe sponsorship. The app runs locally with pnpm and is MIT‑licensed.

Bot Crossing is a local, visual dashboard that maps your active AI coding agent threads into a single 3D colony, allowing you to monitor and manage sessions across different tools without modifying your original data.

An autonomous offensive security framework that automates bug bounty hunting and vulnerability research using an AI agent or MCP server integration.

mu is a command‑line (and optional desktop) coding assistant that pairs a large LLM with a tiny “judge” model (Jev or local Laya). The judge answers ~38 small yes/no/choice questions per interaction turn, handling context management, safety checks, tool‑output filtering, and multi‑agent (hive) coordination. This keeps the big model’s token usage low while preserving safety and relevance. Install via `npm i -g mu-agent` (CLI) or download the desktop installer; configure a model key with `mu setup`. The system logs every judgment for auditability and includes measured reductions of prompt size (≈51 % log folding) and goal‑aware output trimming (≈40 % cut with no required lines lost).

An open-source generative UI toolkit for React that allows AI agents to dynamically render and update UI components based on Zod schemas.

Fusion is an open‑source AI‑powered software factory that turns plain‑language task descriptions into fully reviewed, merge‑ready code. It coordinates a fleet of LLM‑driven agents, runs each task in its own Git worktree, and presents the whole process on a visual dashboard with configurable workflows and human oversight.

A high-performance 3D-printed desktop robotic arm designed to emulate industrial robots in mechanical design and control software.

Hycli is a cross‑platform native tool that automatically turns any logged‑in website into a typed command‑line interface and an MCP server, letting AI agents safely read and (with explicit one‑time permission) write to the site. It prepares actions by analysing docs, OpenAPI/GraphQL specs, and live browser traffic, stores credentials locally, and offers a dashboard plus 20‑language UI. Ideal for building autonomous agents or scripting personal web workflows.

An end-to-end audio understanding model that jointly performs speech transcription and speaker diarization for long-form multi-speaker audio in 50+ languages.

BG0 is an open-source, browser-based background remover that uses WebGPU and WASM to process images locally for maximum privacy.

XGBoost is an optimized distributed gradient boosting library that provides fast and accurate parallel tree boosting for large-scale data science problems.

Sentence‑Transformers is a Python library that simplifies using and fine‑tuning dense, sparse, cross‑encoder, and multi‑vector transformer models for text embeddings, semantic search, and reranking. It offers a one‑line API, thousands of pre‑trained models on Hugging Face, and extensive training utilities.

ZLUDA is a drop-in replacement for CUDA that enables unmodified CUDA applications to run on non-NVIDIA GPUs with near-native performance.