📦 huiliyi37/dsh-tianshu-tui

dsh-tianshu-tui: The Terminal Interface Redefined

The DSH Tianshu TUI renders complex AI agent conversations as smooth, live streaming text directly inside your command line without any browser overhead. This open source plugin transforms the official DeepSeek Harness into a polished interactive workspace that feels more like a modern desktop app than a raw terminal. Instead of staring at static logs, you get real time updates, custom color themes, and built in tools for memory and code diagnostics. It solves the friction of reading dense AI output in plain text by making every response visually structured and easy to scan.

🆔 @hackernewsgithubprojects

📦 yifanzhang-pro/recurrent-looped-tranformer

Recurrent Looped Transformer

Recurrent Looped Transformer is the research project that gives language models unbounded temporal depth by carrying hidden state across every single token. Instead of treating the prompt and response as separate chunks, it loops the internal memory through every word so the model can keep a continuous line of reasoning without forgetting what it just saw. This design creates a path that grows longer as the sequence gets longer, allowing the model to maintain context over much longer distances than standard architectures permit.

🆔 @hackernewsgithubprojects

📦 browser-use/browser-use-pi

How Browser Use Pi Writes Its Own Code

Browser Use Pi is the web agent that writes its own JavaScript to control your browser. Instead of relying on rigid commands, it spins up a live coding environment where it can invent custom tools on the fly. It looks at the page, writes the exact code needed to find a button or read data, and executes it directly inside Chrome. This means it handles complex, weird websites by building its own helpers rather than getting stuck on standard actions. It is a small, fast tool for developers who want their apps to navigate the web with human-like flexibility.

🆔 @hackernewsgithubprojects

📦 iconidentify/chonkstep

Chonkstep: The Linux Window Manager That Speaks Hyprland

A Linux window manager can speak the exact same configuration language as Hyprland without sharing a single line of code. Chonkstep is a traditional floating compositor for the Omarchy ecosystem that acts as a direct drop-in replacement for Hyprland. It solves the migration headache for users who want the familiar feel and setup of a popular tiling manager but prefer the flexibility of a floating layout. The project is built in Rust and includes extensive benchmarking data to prove its performance across different graphics hardware.

🆔 @hackernewsgithubprojects

📦 bulutarkan/mac-mcp

Mac MCP: Give Your AI a Hands-On Mac

Give your AI agent the ability to actually use your computer. Mac MCP is a local server that connects coding assistants directly to your Mac, allowing them to run commands, manage files, and control apps. The most surprising part is how it handles your browser. It can open tabs and read data in the background without stealing focus or hiding windows. You stay in control of your screen while the agent works quietly on a separate tab. This means you can have AI research or fill forms right next to your own work.

🆔 @hackernewsgithubprojects

📦 arnsri33/embedflow

EmbedFlow: Zero Downtime Embedding Upgrades

EmbedFlow is the migration tool that lets you upgrade your search model without taking your database offline. Usually, swapping out an embedding model means re-processing every single document in your index before you can use the new one, which takes forever and costs a fortune. This project flips that script. It uses a clever trick where the new model scores a smaller list of candidates found by your old, existing index. While that happens, it quietly builds the new index in the background, so your search stays live the whole time. It works with major vector databases like Faiss and Pinecone.

📰 https://news.ycombinator.com/item?id=49636147

🆔 @hackernewsgithubprojects

📦 legendapp/legend-apps

Legend Apps: Code, Music, and Docs on One Platform

Build native apps that feel fast and stay local, without the usual cross platform bloat. Legend Apps is a workspace where you can write your own music player, markdown editor, or code viewer using one shared engine. It uses React Native and a custom native shell to talk directly to your operating system, meaning no browser layer slowing you down. The coolest part is that the same code runs on your Mac, iPhone, or Android, so your apps look identical everywhere. If you want to learn how to build serious desktop software that actually respects your files, this is the playground to study.

🆔 @hackernewsgithubprojects

📦 tier4/meteor

METEOR: The AI That Wrote Its Own Driving Brain

Watch as a single network processes eight camera feeds to map lanes, detect cars, and plan a driving path all at once on a tiny computer chip. The most surprising part is that no human wrote a single line of code for this system. Instead, artificial agents built the models, generated the training labels, and ran the tests autonomously. It turns raw video into a full understanding of the road without any human annotation. This proves that machines can now handle the entire process of creating and testing complex software, marking a huge step toward self-maintaining technology.

🆔 @hackernewsgithubprojects

📦 qisutu/qisutu

Qisutu: The Ticket System That Talks Eleven Languages

Qisutu is the open source ticket system that speaks eleven full interface languages right out of the box. Most help desk tools force you to install complex plugins to support global teams, but Qisutu includes complete translations for German, English, French, Italian, Spanish, and six others in its core package. This means you can deploy a support portal where your agents and customers interact in their native tongue without extra configuration. It runs on standard web servers and databases, making it easy to host privately. You get a stable, production ready tool that respects your users' language preferences from day one. No extra plugins needed.

📰 https://news.ycombinator.com/item?id=49449440

🆔 @hackernewsgithubprojects

📦 andreumassanet/impasto

Impasto: The Linux Desktop That Paints Itself

Impasto turns your Arch Linux desktop into a living painting. Pick a wallpaper, and the entire system repaints itself to match those exact colors. The terminal, your editor, even the taskbar all shift hue instantly to blend with the art. It’s not just a theme; it is a complete shell with a dynamic island that morphs into your notes, games, and widgets. One script installs everything, meaning you get a cohesive, artistic workspace without the usual configuration headache. It is a clever way to make your computer look as good as the art on your screen.

🆔 @hackernewsgithubprojects

📦 taoliveaigc/taomate-h3

TaoMate H3: Real-Time Video Gen

TaoMate H3 is the streaming video generator that finally keeps your audio and visuals in perfect sync without the usual wait time. Built by the Alibaba TaoLive team, this tool solves the frustrating problem of generating long videos where the sound and picture drift apart or take forever to render. It works by creating tiny five-second chunks of video and audio at the same time, ensuring that every sound you hear matches the movement you see exactly. This is cool because it lets you make continuous, high-definition clips up to a thousand and eight pixels tall in a fraction of the time normally required.

🆔 @hackernewsgithubprojects

📦 nirholas/three.ws

three.ws

You can now give a chatbot a physical body that lives entirely inside your web browser. Three dot W S is an open-source tool that lets you turn text prompts into interactive three-dimensional characters without needing heavy graphics cards or complicated installations. You type a description, the system generates a model, and then you plug in an artificial intelligence brain so it can think and remember. The coolest part is that these avatars run natively on the web, meaning no downloads or plugins are required to see them move.

🆔 @hackernewsgithubprojects

📦 msu-video-group/freeflow

FreeFlow: The Bias-Free Optical Flow Model

FreeFlow calculates how objects move between video frames without any hard-coded rules about how motion works, letting the network learn everything from scratch. This bias-free approach allows it to handle complex scenes better than older methods that assumed motion follows specific patterns. It takes two images and maps the movement of every pixel, creating a detailed flow map. This is useful for tasks like video stabilization or understanding action in videos. The model is flexible and can be adapted to different datasets. It’s a clever way to let the computer figure out motion dynamics on its own, rather than telling it what to look for.

🆔 @hackernewsgithubprojects

📦 showlab/show-harness

Show-Harness: The AI That Plays Robots

Show-Harness is the robot control tool that lets a vision language model play with hardware like a video game. It solves the messy problem of teaching machines to move by replacing complex coding with a simple set of actions. The coolest part is GUMI, a browser tool where you demonstrate a task by playing the robot with your keyboard. This records your moves as training data instantly, no special hardware needed. It works on real arms and simulations alike. This means anyone can teach a robot new tricks by simply playing in their browser. It is a fresh, human way to build smart machines. Try it out.

🆔 @hackernewsgithubprojects

📦 op7418/pilot-harness

Pilot Harness: Desktop AI Client

Pilot Harness lets you run your AI assistant in a proper desktop window, not just a terminal. It wraps DeepSeek Harness in a sleek, CodePilot-style interface so you can manage models and providers with your mouse instead of typing commands. Think of it as taking a powerful command-line tool and giving it a friendly, organized dashboard. You get clean settings cards, easy file navigation, and a visual way to pick exactly which AI brain you want to use for your project. It keeps the underlying power but removes the intimidating terminal clutter.

🆔 @hackernewsgithubprojects

📦 xiaomoboy/pi-bluebook

Pi Bluebook: Master AI Agents in Chinese

Pi Bluebook is the Chinese learning guide that turns confusing AI agent concepts into clear, checkable steps. It solves the problem of knowing how to install tools without knowing what a successful result actually looks like. The project walks you through real tasks, showing exactly what to see and how to verify completion at every stage. You get a structured path from your first simple command to complex workflows, ensuring you truly understand the logic behind the agent actions. It is not just a manual, but a practical curriculum that proves your work is done.

🆔 @hackernewsgithubprojects

📦 eternityspring/reelbench-skills

Reelbench Skills: The Shot Breakdown Tool

Turn any finished film into a detailed shot-by-shot table without manually logging every cut. Reelbench Skills measures exact durations and camera moves, then generates a single interactive page you can open offline. It highlights the current shot as the video plays, letting you see exactly how pacing and rhythm build the narrative. It is not a code generator or a simple editor; it is a structural analysis tool that breaks down the mechanics of visual storytelling. If you are trying to understand why a specific scene feels the way it does, this is the missing piece of the puzzle.

🆔 @hackernewsgithubprojects

📦 skyeshark/eanpa-sky

Eanpa Sky: The Realistic Skybox That Actually Works

Want your 3D games to have skies that actually look like they belong in a movie? Eanpa Sky does exactly that for Three.js developers. It creates a fully real-time, volumetric sky that handles everything from clear blue days to dark, stormy nights. The coolest part? You get to choose your planet. Stick with Earth, or jump to a sci-fi world with a red giant star and a shattered moon. It is not just pretty pictures, though. It features live weather fronts, lightning that actually hits the ground, and realistic cloud shadows that play with your scene. It is a free engine that makes your virtual world feel alive.

🆔 @hackernewsgithubprojects

📦 jiejie567/jepa-policy

JEPA Policy: The Robot Brain That Ditches Diffusion

JEPA Policy is the imitation-learning framework that skips the heavy diffusion model to train robots faster. Instead of the complex noise-based steps usually required, this system uses a single shared Transformer to predict the next robot movement and a rough sketch of what the scene looks like a moment later, all at once. This joint prediction allows the visual encoder to learn from scratch while staying efficient, meaning you get sharp, reliable motion without the computational overhead of traditional sampling loops. It is a streamlined way to teach machines physical tasks like picking and placing objects directly from video demonstrations.

🆔 @hackernewsgithubprojects

📦 eclipse-ai4science/gradsolve

How Grad Solve Makes GPU Math Faster

Grad Solve is the GPU solver that finally makes massive differential equation ensembles fast and differentiable. It runs thousands of simulations simultaneously on your graphics card, which is great for testing how changes in one parameter ripple through a whole system. The real magic is in how it handles gradients. Instead of painfully recalculating every step to find the derivative, it records the path during the forward run and simply replays it. This record-and-replay trick cuts calculation time by up to fourteen times compared to standard tools.

🆔 @hackernewsgithubprojects