📦 lightorigins/lightnav-0

One Model, Four Robots: LightNav Zero

Drive four completely different robots through an unseen park using just a single camera and a sentence. That is exactly what LightNav Zero does. It takes a standard vision language model and teaches it to understand space, so it can follow instructions, find objects, and track people without needing special parts or depth sensors. The model looks at your camera feed, figures out where to go, and sends simple waypoints to the robot’s controller. You can try it right now in a browser simulator, no GPU required. If you are building autonomous agents, this is the bridge between knowing what is in a room and actually moving through it.

🆔 @hackernewsgithubprojects

📦 phoneuse-official/phoneuse

PhoneUse: Control Your iPhone With AI

PhoneUse turns your Mac and USB cable into a remote control for a real iPhone, letting you type plain English commands that an AI agent executes on the physical device. It solves the friction of testing mobile apps by letting you automate taps, scrolls, and text entry locally without cloud servers. The coolest part is watching the live screen mirror update in real time as the AI navigates your actual phone. It is a developer tool, not a magic trick, so you still need to trust the process.

🆔 @hackernewsgithubprojects

📦 nondeletable/comfylauncher

ComfyLauncher

ComfyLauncher is the desktop tool that finally stops portable ComfyUI from hijacking your main browser. Many artists run multiple separate versions of ComfyUI to avoid conflicts, but the standard setup forces a heavy, cluttered browser to open every time. This app replaces that with a dedicated, lightweight window that uses significantly less memory than a standard Chrome instance. It lets you pick from a list of saved builds, hide the messy terminal window, and even monitor the server status right from the app. If you run multiple instances, this gives you a clean, focused workspace without the extra system overhead.

🆔 @hackernewsgithubprojects

📦 ndom91/open-plan-annotator

Open Plan Annotator: Annotate AI Plans Like a Document

Open Plan Annotator stops your AI coding agent right before it writes code. Instead of a quick chat approval, it launches a browser window where you can mark up the plan exactly like a document. You can strike through parts you don't like, replace text, or add comments. The agent reads these notes and revises the plan until you approve. It runs entirely on your computer, keeping your data private. This tool turns a simple yes or no into a detailed, collaborative editing session. It is a clever way to take real control over your AI workflows. Check it out to see your plans come to life.

📰 https://news.ycombinator.com/item?id=49060901

🆔 @hackernewsgithubprojects

📦 blaiselabs/procedural-almanac

The Code Island: Procedural Almanac

A tiny piece of code can build a whole island, complete with rivers, storms, and walking deer, without using a single image file. That is what procedural almanac does. It is a collection of over one hundred sixty small tools for drawing and animating on a computer screen. Instead of giving you a finished engine, it gives you the specific math behind things like how water ripples or how a leg steps. The coolest part? Every single tool comes with a note telling you exactly where the math comes from. It is like a field guide for creating worlds from scratch.

🆔 @hackernewsgithubprojects

📦 floci-io/floci-az

Floci AZ: Your Free Local Azure

Floci AZ lets you run a fully functional copy of the entire Azure cloud right on your own laptop, free of charge. Instead of paying for servers or waiting for cloud credits, you just spin this up locally to test your apps against real Azure services like storage, databases, and even Kubernetes. The coolest part? It actually spins up real Docker containers for your databases and message queues, so your code behaves exactly like it would in production, but without the cloud bill. It handles authentication, networking, and deployments all in one lightweight package. No more worrying about breaking your actual cloud account during a test run.

🆔 @hackernewsgithubprojects

📦 tencentarc/looking-ahead

Looking Ahead

Make video AI that actually understands physics. Most video models just guess the next frame, which leads to weird stuff like objects vanishing. Looking Ahead fixes this by teaching the model to predict what happens a few seconds from now, not just the immediate next moment. It uses a technique called Future Chunk Prediction, which forces the model to look ahead and learn how things really move and change over time. The result is videos where things stay consistent, like milk pouring smoothly or dominoes falling in order. It’s a simple idea with a big impact, making generated video feel much more real and less glitchy.

🆔 @hackernewsgithubprojects

📦 platformatic/sqlite-js-stream

sqlite-js-stream

Keep your database safe even if your server crashes. This library, called sqlite js stream, constantly backs up your local SQLite file by copying its recent changes to a remote location like Amazon S3 or a simple folder. It does not move your whole database every time; instead, it tracks only the new data, compresses it into tiny, safe files, and uploads just those updates. This means your backup stays lightweight and fast, no matter how big your database grows. If something ever goes wrong, you can instantly restore your data from those uploaded files, guaranteeing you never lose your work.

🆔 @hackernewsgithubprojects

📦 bimsonz/keel

Keel: A Git Platform For AI Agents

Keel lets software agents push code to a Git platform that automatically checks for semantic conflicts before anything lands. Built on Cloudflare, it gives every agent a dedicated workspace and verifies that changes do not break existing assumptions. The standout feature is how it writes a test to prove a conflict exists, keeping that guard in place until a human makes a ruling. This creates a verified merge queue where only safe changes advance to the main trunk. It is an interesting look at how we might manage autonomous coding in the future.

🆔 @hackernewsgithubprojects

📦 chadyi/studyontv

StudyOnTV: Earn Your Screen Time

An entire family media center runs without a single cloud server. This Android application turns the television into a learning hub where kids must answer questions to earn their screen time. The most surprising part is how self-contained the system is. Everything from the question bank to the usage logs stays stored directly on the device. Parents simply scan a local code to manage the rules from their phone. It pauses the video player automatically during study sessions and restores it when done. It works entirely offline and keeps the data local.

🆔 @hackernewsgithubprojects

📦 kenkaiii/gg-framework

GG Framework: The Coding Agent That Saves You Money

Cut your AI bill by nearly half without sacrificing speed. GG Framework is a desktop coding agent that races ahead of the competition by using just fourteen cents worth of compute for tasks that cost other tools forty-nine. It achieves this by skipping the usual guesswork and instead searching the actual open-source code on your machine to find precise answers instantly. This means you get verified results much faster than the heavy alternatives that waste tokens on vague suggestions. If you are paying for an AI subscription, this lightweight tool ensures you squeeze the most work out of every dollar you spend.

🆔 @hackernewsgithubprojects

📦 spaquet/markify

Markify: Write and Render Markdown in One Pane

Stop juggling split screens and start writing directly on a living page. Markify is a native Mac app that lets you type Markdown and see the formatted result in the exact same spot, without any separate preview window. The coolest part is how it handles the switch between modes. You can write your document with headings, tables, and math fully rendered, then press one shortcut to instantly fade into the raw code view. It keeps your cursor and scroll position perfectly in place, so you never lose your rhythm.

🆔 @hackernewsgithubprojects

📦 brightbeanxyz/brightbean-chat

BrightBean Chat: Own Your Chat Automation

BrightBean Chat runs on your own server, not a third-party cloud, giving you full ownership of your audience data without per-contact fees. This open-source tool acts as a self-hosted chat marketing hub, letting you build visual message flows that trigger automatically across Instagram, WhatsApp, Telegram, SMS, and email. You can turn simple Instagram comments into private direct messages or schedule drip campaigns using a visual drag-and-drop builder. Since you host the software, there are no feature gates, no middlemen, and no monthly bills that scale with your subscriber count. It is a complete, free alternative to paid chat automation platforms.

🆔 @hackernewsgithubprojects

📦 ultrametricai/ultrametric

Ultrametric: The Open Startup Repo for the Agentic Era

Ultrametric treats AI agents as the primary users of business software, ranking every vendor from banks to legal services on how well agents can actually control them. It is an open repository that maps the entire startup journey, breaking down complex processes like incorporation into clear steps that can be handled by humans, forms, or automated agents. This solves the confusion of choosing tools that look good to humans but fail when machines need to operate them. The project also includes open code for standard calculations like runway and equity dilution, ensuring transparency.

🆔 @hackernewsgithubprojects

📦 jdpolasky/chief-of-staff-2

Chief of Staff 2: Your AI Chief of Staff in Plain Markdown

Grab your morning brief by letting an AI read the plain text notes of your life. Chief of Staff 2 is a digital assistant that tracks your calendar, commitments, and standing rules directly inside an Obsidian vault, so you stop holding everything in your head. Built specifically for people with ADHD, this system uses a swappable AI to reason about your day instead of just displaying a to-do list. It is all local markdown, meaning no hidden databases or subscriptions, just a folder you can open and read. The author stripped away complex machinery to make it fast and transparent.

🆔 @hackernewsgithubprojects

📦 grafeodb/grafeo

Grafeo: The Embedded Graph Database

Grafeo is a graph database written in Rust that runs entirely inside your application without needing a separate server or heavy infrastructure. It solves the problem of slow, memory-hungry database engines by using a highly optimized columnar storage format that keeps your data tightly packed on disk. This means you can run complex graph queries with a tiny memory footprint compared to traditional alternatives. You can query it using six major languages, including Cypher and SPARQL, directly from your code. Because it is built as a library rather than a standalone service, it eliminates network latency and simplifies your deployment pipeline.

🆔 @hackernewsgithubprojects

📦 68hctmzzjf-svg/tvport

TVPort: Play Switch Games on Apple TV

TVPort turns your Apple TV into a Nintendo Switch by running the actual game code directly on the Apple chip, bypassing the usual emulation overhead entirely. It pairs with a simple Mac application that takes the game files you legally own and compiles them into a standard Apple TV app, handling the complex graphics translation automatically. This means titles like Super Mario Odyssey and Tears of the Kingdom run natively on your living room screen without needing a separate handheld console nearby. The software even manages save data so it persists even if the system clears its temporary cache.

📰 https://news.ycombinator.com/item?id=50378708

🆔 @hackernewsgithubprojects

📦 leggedrobotics/terrain-grounded-imagination

Terrain-Grounded Imagination: How Robots Learn to Walk

Teach a robot to walk over broken ground by letting it dream about the terrain. This project trains a humanoid to navigate stairs, gaps, and stepping stones entirely through a learned world model. Instead of hard-coding the ground into the robot’s brain, it reads the terrain directly from a heightmap, allowing the robot to imagine its next steps and adjust its balance based on exactly what it feels under its feet. This approach makes the robot’s perception of gaps and edges precise, ensuring it knows when a foot is unsupported.

🆔 @hackernewsgithubprojects

📦 jrryzh/f2g-pose

F2G-Pose: The One-Shot Object Scanner

F2G-Pose lets you figure out exactly where an object is in three-dimensional space from just a single photo and depth scan. You know how it feels when you snap a picture of a mug on your desk and your phone’s camera can’t quite tell which way it’s facing? This project fixes that. It takes a standard image and a depth map, and in one go, it guesses the object’s exact position, its size, and even what the whole shape looks like, even if you only saw half of it. No labels, no special templates, just a quick guess.

🆔 @hackernewsgithubprojects

📦 harunrrayhan/content-machine

Content Machine: Own Your Content Pipeline

Capture ideas straight from your phone and turn them into scheduled social posts without ever opening a separate app. That is exactly what Content Machine does. It acts as a private hub where you can jot down thoughts via a simple text note or even a voice memo. The system then uses AI to organize those messy fragments into clear, actionable content. Once your idea is ready, you can schedule it to publish automatically across your social channels. You get a single place to manage your creative workflow, keeping everything in sync. No more juggling multiple apps.

🆔 @hackernewsgithubprojects

📦 apolinario/urnas-tardias

Urnas Tardias: The 5,000 Votes That Never Should Have Counted

A new data analysis reveals exactly how many votes were cast after the legal deadline in Brazil's 2026 second round. The project, called Urnas Tardias, digs into the raw machine logs published by the electoral court to track every single vote processed after five in the afternoon. It turns out that only about five thousand votes were actually counted late across the entire country. This tiny fraction represents less than one percent of the massive lead secured by the winner. The analysis proves that extending voting hours had virtually no impact on the final outcome. The takeaway is simple: the election result is mathematically immune to late vote processing.

🆔 @hackernewsgithubprojects