📦 cristianolivera1/aura

Aura: The Gradient Tool That Makes Backgrounds Feel Like Light

Aura is a web app that turns simple color layers into backgrounds that actually feel like glowing light. It ships over two hundred preset gradients built with CSS blend modes, so instead of flat fills you get soft, atmospheric washes that look like light bleeding through glass. The standout feature is the in-browser generator: stack layers, tweak blur, opacity, and blend modes in real time, add a grain overlay for a filmic finish, and even generate a prompt to rebuild the exact gradient in AI tools.

🆔 @hackernewsgithubprojects

📦 tgo-app-dev/vpipe

Vpipe Runs Massive AI Video Models on a 16GB MacBook

Generate video with sound locally on a standard sixteen-gigabyte MacBook Air. This is the surprising part. Vpipe is a local runtime built specifically for Apple Silicon that runs massive AI models, like the thirty-three billion parameter MiniMax H3, without needing a high-end graphics card or cloud server. It achieves this through custom Metal inference kernels and clever weight streaming, allowing you to create videos and their matching soundtracks entirely on your laptop. The entire runtime is incredibly compact, weighing in under thirty megabytes, which means it fits easily on your machine while handling heavy multimodal tasks like image editing, real-time chat, and video generation.

📰 https://news.ycombinator.com/item?id=49397410

🆔 @hackernewsgithubprojects

📦 cayu-dev/cayu

Cayu: The Runtime for Real AI Agents

Cayu is the production-grade runtime that actually lets AI agents survive long, complex jobs without falling apart. Most agent tools handle the quick chat fine, but they crumble when your agent runs for hours, loses connection, or needs to pause for a human approval. Cayu fixes this by making every step of your agent’s work durable and replayable. It treats your agent’s memory and actions like a real transaction, so if a server crashes, the agent picks up exactly where it left off. It also strictly controls what tools your agent can use and how it spends money, giving you a clear audit trail of every move.

🆔 @hackernewsgithubprojects

📦 liguobao/deepseek-harness-remote

DeepSeek Harness Remote: Control Your AI From Anywhere

Your work computer keeps running your AI coding agent, even when you are on a train, and this tool lets you check in from your phone. DeepSeek Harness Remote creates a secure bridge so you can follow active sessions, send new instructions with images, and answer permission prompts from your Android phone or another browser without moving your files. It works by turning your desktop into a host that connects outbound over an encrypted channel, meaning no public ports are opened and your data stays private. The project solves the problem of being stuck away from your workstation while your agent is still processing complex tasks.

🆔 @hackernewsgithubprojects

📦 r44vc0rp/pgrok

pgrok: Own Your Tunnel

pgrok is the self-hosted tunneling tool that gives you instant HTTPS on your own VPS. Forget renting from a third party to share your local dev server with the world. With one command, it spins up a secure SSH tunnel and Caddy handles the SSL certificates automatically, so any subdomain you pick works instantly. You even get a live terminal dashboard to watch requests flow through in real time. It is the perfect way to test webhooks or share a staging environment without exposing your home network or paying for a service you do not control.

🆔 @hackernewsgithubprojects

📦 only-cli/oc

Stop Wasting Tokens on Web Browsing

Turn any website into a compact command line interface for AI agents, cutting browsing costs from tens of thousands of tokens to just hundreds. This tool fetches a page and returns a tidy, numbered view, allowing agents to read, follow links, and search without the massive overhead of raw HTML. It even slips past basic blocks by mimicking a real browser. The single most impressive thing is that it lets an agent browse a live site using less data than a single tweet, making complex web navigation incredibly cheap. If you are building AI tools, this is the fix for your token bill. Check it out.

📰 https://news.ycombinator.com/item?id=49367419

🆔 @hackernewsgithubprojects

📦 raiyanyahya/llmaker

llmaker: The One-Command LLM Stack

A single terminal command can spin up a private, production-ready AI assistant, complete with vector databases and caching, without writing a single line of configuration glue. This project, called llmaker, solves the painful problem of manually wiring together different AI tools by provisioning the entire fleet on a private network from one simple file. It handles the networking, service discovery, and GPU management so you can focus on building applications instead of managing containers. Because everything runs locally, your sensitive data stays on your own hardware, and you can talk to the model using standard interfaces.

📰 https://news.ycombinator.com/item?id=48726917

🆔 @hackernewsgithubprojects

📦 techomancer/iris

IRIS: Emulating a 1990s SGI Workstation

Boot a full 1990s SGI Indy workstation on your modern computer using Iris, a new emulator written in Rust. It does not just mimic the screen; it simulates the actual MIPS processor, letting you run the original IRIX operating system with a working desktop, shell, and network access. The project is particularly interesting because it uses a specialized compiler to translate the old machine code into native instructions on the fly, ensuring the system runs fast enough for daily use. It even includes safety features that protect your hard drive images from crashes. This is a fascinating look at how complex legacy hardware can be revived on today's systems.

📰 https://news.ycombinator.com/item?id=47940203

🆔 @hackernewsgithubprojects

📦 hars-21/reqsh

The Persistent HTTP Shell

Reqsh remembers your API session even after you close the terminal. Built in Rust, this interactive shell saves your base URL, headers, and variables to a local file, so you pick up exactly where you left off without retyping anything. It acts like a smart command line for testing web services, letting you build requests line by line and instantly replay them. Because it handles state persistence automatically, it solves the annoyance of losing your test setup between restarts. It is a fast, lightweight tool for developers who need a quick way to hit APIs without writing code.

🆔 @hackernewsgithubprojects

📦 ashishps1/awesome-engineering-articles

Inside the Engineering Blogs of Tech Giants

Discover how the biggest tech companies solve massive problems by exploring a curated collection of over three hundred engineering articles. This resource gathers insights from teams at places like Airbnb, Discord, and Meta, showing you exactly how they build systems that handle billions of requests. You will learn practical lessons on performance, databases, and artificial intelligence, all written by the engineers who built those tools. It is a direct window into the real world of large scale software, helping you understand the decisions behind the apps you use every day.

🆔 @hackernewsgithubprojects

📦 phase-rs/phase

Phase: A Full Magic Rules Engine Built by One Developer

A single developer built an open source rules engine that runs the complex rules of Magic: The Gathering entirely in your browser. Phase uses Rust to handle the logic, compiling it to Web Assembly so it runs fast on desktop, in a web browser, or over a network with friends. The creator started this project because the official game is too complicated for his six year old son, but wanted to build something capable enough for real players. It handles everything from card casting to combat resolution with an AI opponent included.

📰 https://news.ycombinator.com/item?id=49405728

🆔 @hackernewsgithubprojects

📦 shy3130/tick-stock-panel

Tick Stock Panel: Your Local AI Stock Lab

Run your own local stock lab that scans the entire A-share market in milliseconds, letting you test trading strategies before risking a cent. Tick Stock Panel is a self-hosted workspace that pulls real market data, calculates complex indicators, and lets you build, monitor, and backtest your own rules. The coolest part is that it uses AI to help you write these strategies, so you don't need to be a coding wizard. It even plays audio alerts when your signals trigger, acting like a quiet analyst watching the screens for you. It is not a magic money maker, but it is a powerful tool for learning how markets actually move.

🆔 @hackernewsgithubprojects

📦 agenticenv/agent-sdk-go

Agent SDK Go: Crash-Proof AI Agents

Agent SDK Go is the development framework that ensures your AI agents survive crashes and restarts without losing a single step. This open-source Go library solves a critical pain point by using durable execution backends like Temporal or Restate to keep agent state, tool calls, and execution loops intact even if your process dies. You can run agents locally with zero setup, but the real power lies in production, where the system automatically reconnects and resumes exactly where it left off. It supports multiple LLM providers, human-in-the-loop approvals, and sub-agents, making it a complete toolkit for building reliable, resilient AI applications in Go.

🆔 @hackernewsgithubprojects

📦 microsoft/thinkingbox

ThinkingBox: Testing AI Agents

Microsoft’s ThinkingBox gives AI agents a sandbox where they can practice using tools like email and banking systems before facing real users. This framework connects language models to mock environments, allowing them to attempt tasks while the system checks if the actions were actually correct. It helps developers generate high-quality training data and find subtle failures that simple text checks miss. By isolating these digital rehearsals, teams can build more reliable assistants. It turns vague model confidence into measurable, verifiable performance.

🆔 @hackernewsgithubprojects

📦 fraunhoferhhi/ripepp

RIPEpp: Matching Images Without Negative Data

RIPEpp is the vision tool that trains image matchers using only positive examples. Usually, teaching a computer to link points in two photos requires telling it what matches and what does not. This project flips that logic by learning entirely from successful pairs, removing the need for negative training data. It acts as a complete pipeline that spots key points and links them across images. By relying solely on positive pairs, it simplifies the training process while still delivering strong results. If you work with visual data, this approach could change how you build matching systems.

🆔 @hackernewsgithubprojects

📦 rampstackco/claude-skills

Claude Skills: The Website Brain

Claude Skills is the playbook collection that turns your AI assistant into a full-stack web expert. Most AI tools just write text, but this repository packs over forty distinct skills that teach Claude how to handle the entire lifecycle of a website. Think of it as giving your AI a senior team of designers, SEO specialists, and developers to consult. The coolest part is that it handles the unglamorous but critical stuff, like auditing your site for accessibility or figuring out why your marketing emails are bouncing. You do not need to be a code wizard to use it.

🆔 @hackernewsgithubprojects

📦 fulappios/agent-quest

Agent Quest: Your AI Agents in a Pixel Village

Agent Quest turns your terminal AI sessions into a live pixel-art village where every running agent becomes a wandering hero. Instead of staring at command logs, you watch characters walk to the library when reading code, the forge when editing files, or the arena when running commands. It automatically finds your active Claude and Codex installations, so you can glance at your screen and instantly see what your bots are doing without leaving your desk. It is a simple, open source tool that makes invisible background work visible through a playful medieval interface, giving you a clear picture of your automation workflow at a glance.

🆔 @hackernewsgithubprojects

📦 yunhao-feng/vera

Vera Finds Hidden AI Agent Risks

Vera successfully attacks over ninety percent of modern AI coding agents in large-scale safety tests. This framework acts as a relentless security auditor, automatically generating thousands of unique scenarios to expose vulnerabilities that manual reviews miss. It spins up isolated digital environments where AI agents can interact with fake emails, code repositories, and banking services. The system then attempts to trick these agents into breaking their own safety rules, verifying every result with hard data rather than trusting the AI’s word. This project matters because it turns abstract safety concerns into concrete, reproducible evidence, helping developers understand exactly where their tools might fail before a real-world incident happens.

🆔 @hackernewsgithubprojects

📦 fanfulla/ocr-buddy

OCR Buddy: The Chrome Extension That Refuses to Lie

OCR Buddy grabs text directly from your screen, including code, formulas, and tables, without ever sending a single byte to a server. Most optical character recognition tools rely on large, cloud-based AI models that guess what they see, often inventing words when the image is blurry. This extension takes a different approach by running a specialized, classic OCR engine entirely on your device. It reads the pixels exactly as they appear, ensuring that if it cannot read a character clearly, it simply leaves it blank rather than hallucinating a plausible but wrong answer. This makes it perfect for capturing precise data like prices or identifiers where accuracy is critical.

📰 https://news.ycombinator.com/item?id=48824434

🆔 @hackernewsgithubprojects

📦 gaolii/quad-imaginarium

Robots Dancing Without Motion Capture

Imagine teaching a robot to do backflips without hiring a single motion capture artist. Quad-Imaginarium is a massive library of quadruped robot moves, all created by artificial intelligence. It generates thousands of acrobatic clips for the Unitree Go2 robot, turning video into real, physical movement. The team used a clever video model to invent these dance moves, then made sure the robot could actually copy them. In fact, almost every move worked perfectly on the real hardware. This proves you can create endless robot behaviors simply by describing them in words, rather than filming a human or an animal to learn from.

🆔 @hackernewsgithubprojects

📦 herrgottmargott/herrgotts-h3-infinite-continuation-suite

Infinite H3 Video Continuation for ComfyUI

Chain together short clips into a long, seamless video where the audio and picture never stutter. This ComfyUI tool connects segments of a generated video by directly copying the previous visual and sound data into the next one, rather than trying to guess how they should blend. It intelligently finds a safe point in the video that isn’t frozen or glitchy and locks the sound to continue naturally across that cut. This means you can build long, continuous scenes without awkward jumps or dropped words. It’s a clever way to extend your creative flow without fighting the machine.

🆔 @hackernewsgithubprojects