📦 getgat-dev/gat

Gat: Git for Huge Files

Track massive files like video clips or model weights directly in your git history without bloating your repository. Gat handles the heavy lifting by storing these large assets in cheap cloud storage while keeping your code and version control completely standard. It works like a smart file manager that automatically syncs the right version of every big asset whenever you check out a branch or pull changes from a teammate. You do not need a special server or complex setup, just point it at any standard cloud bucket and start committing.

📰 https://news.ycombinator.com/item?id=49868710

🆔 @hackernewsgithubprojects

📦 zzzz7788990213-ops/evovlm

EvoVLM: How to Make Vision AI Faster

EvoVLM makes a small vision model answer chart questions twice as accurately without touching its weights. This is a research tool that treats the model like a frozen brain and instead evolves the software around it. It uses another AI to tweak prompts, image handling, and thinking time, testing different combinations to find the sweet spot. It’s basically training a recipe for getting the best answer quickly, rather than retraining the model itself. The result is a faster, smarter way to handle scientific charts. It’s a clever hack for efficiency. If you care about making AI lighter and smarter, this is how you do it.

🆔 @hackernewsgithubprojects

📦 infinri/writ

Writ: The Bouncer for AI Coding Agents

Writ blocks an AI coding agent from writing a single line of code until you explicitly type a specific word to approve the plan. It treats AI instructions as suggestions rather than guarantees, stepping in to enforce strict rules the moment a tool tries to run. The system pauses work, demands a human review, and refuses any self-approval to keep you in control. It also tracks every decision so you never lose context when a long session gets compressed. If you want to trust your AI pair programmer, you need a layer that physically stops it until you say go.

🆔 @hackernewsgithubprojects

📦 codepdbh/nfsmw-android

Run NFS Most Wanted on Your Phone

Turn your smartphone into a console for the classic two thousand and five racing game Need for Speed Most Wanted. This project takes the original Xbox code and rebuilds it from scratch to run natively on Android phones using modern graphics technology. Instead of simply emulating the old console which is slow and heavy, this port rewrites the engine so it speaks directly to your phone's processor. The result is a high-speed racing experience with touch controls and smooth video playback that feels surprisingly close to the original hardware. It is a massive piece of work that lets you pick up and play a cult favorite from your pocket.

🆔 @hackernewsgithubprojects

📦 mini-yifan/dsh-orb-cordis

dsh-orb-cordis

A small floating ball sitting on your screen edge can literally click your computer buttons for you. This unofficial add-on for the DeepSeek Harness desktop app installs a persistent orb that sees your screen, understands what you ask, and performs visible actions like filling forms or opening apps directly in front of you. If you have a heavy task like writing code or organizing files, it quietly hands that work to a background session so you can keep chatting without waiting. The whole system runs as a plugin, leaving the original application completely untouched while giving you a tiny, always-on assistant that actually does things on your desktop.

🆔 @hackernewsgithubprojects

📦 rbrown101010/bluey-by-riley

Bluey By Riley

Your phone has become a physical cursor for your Mac, controlled by a silly blueberry character named Bluey. It is a menu bar app and iPhone companion where you talk to the fruit, and he actually clicks, types, and navigates your desktop. The coolest part is that he has his own pointer, so he can interact with your screen without messing up where you left your mouse. You press and hold the phone to ask him questions, and he answers in a speech bubble. It uses an AI model to understand what you say, but the real magic is that he can open apps, scroll, and type for you.

🆔 @hackernewsgithubprojects

📦 jakeschincariol/replica-skill

Clone Any App With Replica Skill

Replica Skill is the Claude toolkit that rebuilds any application from scratch while fixing what users hate. Instead of copying code, it studies the original app's public screens to map out features, then builds a fresh version using your own database and design. The standout feature is its ability to read real user reviews, identify common frustrations, and apply specific fixes to your new build. It handles everything from database planning to branding, ensuring your final product is legally distinct and ready for launch. If you want to create a competitive app without starting from zero, this is your blueprint.

🆔 @hackernewsgithubprojects

📦 marimo-team/marimohub

Own Your Notebooks: Marimohub

Host your entire data workflow on your own machines with marimohub. Instead of relying on a cloud vendor to manage your code and data, this platform lets you plug in your own storage, compute, and login systems. The coolest part is that there is no separate database. Everything lives in your object store, meaning your backup is just a file transfer. It is built for teams who want full control without the heavy lifting of managing infrastructure. If you like owning your data, this is your new home base.

📰 https://news.ycombinator.com/item?id=49949471

🆔 @hackernewsgithubprojects

📦 charlenemercury/roster-match

Roster Match: The Honest Name Matcher

Roster Match is the name-matching tool that admits when it is unsure. It takes a spoken Chinese name and checks it against a roster, handling tricky surnames and regional accents by comparing the sounds rather than just the spelling. The coolest part is its refusal to guess; if two students sound identical, it flags the ambiguity instead of picking a winner. This prevents the frustrating scenario where a teacher accidentally credits the wrong student. It is a zero-dependency library that prioritizes accuracy over speed, giving you the candidates and the reasoning so you can make the final call.

🆔 @hackernewsgithubprojects

📦 chrissotraidis/bluewake

BlueWake: The Wind Waker on iPhone

Play The Legend of Zelda: The Wind Waker directly on your iPhone, iPad, or Windows computer without an emulator. This project takes the original GameCube game code and translates it into native instructions for your device, so it runs at the game’s native speed without slowing down your phone. The most impressive part is that it runs at a steady thirty frames per second on an iPad Pro, matching the original console’s performance. It handles touch controls, external game controllers, and high-resolution texture packs, all while using your own legally purchased disc. Since it translates the code ahead of time, it feels like a genuine app rather than a simulation.

🆔 @hackernewsgithubprojects

📦 remocn/remocn-studio

Remocn Studio: The Video Editor for AI

Edit video by talking to an AI agent instead of dragging clips on a timeline. Remocn Studio is a desktop app built on Remotion where you describe what you want in plain language, and an AI assistant writes the code to build the animation for you. It comes pre-loaded with motion design rules and templates for product demos, so your AI editor actually understands pacing, easing, and how to make things look professional. You can preview your video in real-time inside the app, and even point at specific elements to tweak them.

🆔 @hackernewsgithubprojects

📦 yusuke710/manim-skill

Manim Skill: AI Video Generation

manim-skill is the coding agent extension that turns plain text prompts into fully produced educational videos with synchronized voiceovers. It solves the problem of complex video production by having the AI agent handle every step, from writing the narration script to generating local audio and coding the final animation. This is useful because it removes the need to manually sync timing between voice and visuals, allowing you to simply describe a concept and receive a polished, explainer-style video that you can easily refine by giving specific feedback on individual scenes.

🆔 @hackernewsgithubprojects

📦 syncso-inc/syncso

SyncSo: Live City Data for AIs

SyncSo plugs your AI assistant into a live feed of actual events happening right now in New York City. Instead of guessing based on old training data, it scans thousands of real listings for live music, comedy shows, and local classes, returning precise times, prices, and direct booking links. You can ask for a quiet first date spot or a free outdoor activity, and the system filters the entire city to find what fits that exact moment. It turns vague questions into bookable plans with real-world details. So next time you ask your chatbot for date ideas, you get fresh answers, not recycled guidebook entries.

📰 https://news.ycombinator.com/item?id=49950236

🆔 @hackernewsgithubprojects

📦 gesta-run/subpool

Subpool: One Key for Every Team AI Subscription

Subpool is the self-hosted manager that finally lets one API key access every AI subscription in your company. It solves the chaos of juggling separate logins for tools like Codex and GitHub Copilot by pooling those accounts behind a single, secure endpoint. Instead of every employee managing their own access, administrators create shared pools and hand out unique keys that automatically route requests to healthy accounts. The system tracks remaining usage so you never hit a wall mid-project, and it never stores your actual conversation data. You get centralized control, clear auditing, and a simple way to share expensive AI licenses without exposing raw credentials to individual users.

📰 https://news.ycombinator.com/item?id=49535569

🆔 @hackernewsgithubprojects

📦 ythx-101/live-panel-skill

Live Panel Skill: Animated Architecture Diagrams

Turn a single text file into a living system map that actually runs. This project generates moving infographics where data packets travel along wires, counters tick upward, and status bars flip colors as states change. It solves the problem of static diagrams by creating a video that looks like a live monitoring dashboard. Every frame is a complete picture, so you can pause it anywhere and still understand the whole system. The result is a short video file or a webpage that loops forever. You just describe your setup in a simple config file, and the tool handles the rest.

🆔 @hackernewsgithubprojects

📦 adityajha2005/yc-outreach

YC Outreach: The 80 Percent Email Trick

Turn any Y Combinator batch into a ready-to-send list of personalized cold emails, all from your browser. This project scans public company pages to pull founder names and websites, then guesses their professional email addresses using common naming patterns. The surprising part is that this simple guesswork is right about eighty percent of the time. You write your email once in a single template, and the tool fills in every founder’s name and company details automatically. It’s free, requires no server setup, and keeps your data private in your browser, making it a surprisingly effective way to reach out to startups without paying for expensive lead generation tools.

🆔 @hackernewsgithubprojects

📦 zoolapp/aime

AIME: The Open Source Chinese Input Method for macOS

Type Chinese on your Mac with a tool that respects your privacy. AIME is an open-source input method built on the RIME engine, meaning your daily typing happens entirely on your device without any hidden data collection. The standout feature is its optional AI layer. You can invoke translation or text polishing directly from your keyboard using a simple shortcut, letting you refine sentences instantly while keeping full control over what gets processed. It blends the flexibility of a traditional engine with modern, on-device intelligence.

🆔 @hackernewsgithubprojects

📦 ymuskrat/arcticdb_mcp

arcticdb_mcp: Your AI's New Data Sidekick

arcticdb_mcp gives your AI assistant direct hands-on control over your financial data, turning complex database tasks into simple conversation. You can ask your bot to check stock prices, update records, or manage versions, and it will just get the work done. This project solves the problem of needing to write manual scripts for every small data change, letting you automate your workflow with built-in safety features like versioning. It is basically a bridge that lets smart assistants talk directly to your data store without breaking anything. If you are tired of writing repetitive code for data tasks, this is the tool that lets you just talk to your database instead.

🆔 @hackernewsgithubprojects

📦 jayleaton/glm53-tensorfold-spark

GLM53 Tensor Fold Spark: 2x Faster AI on Spark

GLM53 Tensor Fold Spark is the inference engine that makes two small computers outperform big servers. It runs a massive AI model on two NVIDIA DGX Sparks, splitting the math between them to answer questions much faster than standard tools. The coolest part is its warm replay feature. If you send the same prompt again, it instantly jumps back to where it left off, answering in a fraction of a second instead of reprocessing everything. It feels like the computer has perfect memory. This proves you do not need a data center to get fast, local AI. Just two sparks and this engine give you serious power.

🆔 @hackernewsgithubprojects

📦 nand2mario/z486_mister

z486 MiSTer

A small FPGA board now runs Doom at nearly thirty frames per second using a processor design that never shipped in a box. The z486 MiSTer core is a digital brain written in code that runs on a tiny chip, turning a handheld device into a retro computer capable of handling 1990s graphics. It solves the problem of running old software on modern hardware without needing a bulky desktop tower. The result is a compact, battery-powered PC that handles classic games with surprising speed. It proves that modern electronics can recreate the muscle of machines from thirty years ago. The takeaway is simple: you can own a pocket-sized Windows computer.

🆔 @hackernewsgithubprojects