📦 vercel-labs/agent-desktop

How Agent Desktop Sees Your Computer

Agent Desktop is the computer-use tool that finally lets AI agents operate your real desktop apps without guessing from pixels. It reads the actual structure of your windows through the operating system’s accessibility tree, giving the agent a clear map of every button and field. Instead of relying on unstable screen captures, it uses stable references to click and type safely, ensuring actions stay reliable even if the interface changes slightly. This makes automation far more predictable and less prone to breaking. It works across Mac, Windows, and Linux, turning complex screen interactions into simple, verifiable steps.

🆔 @hackernewsgithubprojects

📦 flupkede/codesearch

CodeSearch: The Offline AI Codebase Brain

An AI coding agent can now read an entire company of software repositories at once without sending a single byte to the cloud. CodeSearch is a lightweight server written in Rust that plugs into tools like Cursor and Claude Code to answer questions about your codebase using plain English. It solves the problem of agents getting lost in large projects by chunking your code into logical pieces and searching for meaning, not just keywords. Because it runs entirely offline on your own computer, it keeps sensitive data safe while letting you ask complex questions like where is this function used across five different projects.

🆔 @hackernewsgithubprojects

📦 lalipa2003-arch/jailbreakhub

JailbreakHub: The Prompt Vault

JailbreakHub is the research hub that catalogs exactly how to bypass current AI safety filters. It solves the problem of outdated information by tracking the latest, verified prompt injection techniques for major models like Claude, Gemini, and DeepSeek. The most interesting part is the Multi-AI pipeline, which chains different models together to re-encode a request so one AI helps another break its own guardrails. This collection is a snapshot of how language models are currently being tested for weaknesses. It is not a tool for malicious use, but a clear map of the gaps in our digital defenses.

🆔 @hackernewsgithubprojects

📦 raph651/open-source-canvas

Open Source Canvas: The Graph Is the Studio

Your entire creative workflow runs on a single infinite canvas where every connected node is a working piece of your project. This open source tool turns a simple drag and drop board into a full studio for generating images and video. It connects directly to a local server on your own machine, meaning no data leaves your computer and you have total control over the process. The interface feels like a design tool, with clean connections between prompts, models, and final outputs. It is a rare project that combines the logic of a diagram editor with the power of local artificial intelligence.

🆔 @hackernewsgithubprojects

📦 n1neman/sapper-v2

Sapper V2: The AI Civilization

Sapper V2 is the autonomous security testing system that treats AI not as a single smart tool, but as a whole society. It takes a group of smaller, cheaper models and organizes them into a legion with a commander, generals, and soldiers, all working together to find vulnerabilities in software. The real magic is how these agents share knowledge through a layered memory system, allowing the group to learn from each other’s mistakes and successes without a human micromanaging every step. They won a major hacking competition by using far fewer tokens than their competitors, proving that collective coordination beats raw individual power.

🆔 @hackernewsgithubprojects

📦 dbreunig/nesformer

Running the NES on a Language Model

Play Super Mario Bros using a language model that actually executes the console hardware. This project, called Nesformer, compiles the entire NES CPU directly into the weights of a standard Qwen model. No training happened here; a compiler calculated every single weight to turn the model's internal memory into the chip's actual working memory. When you run a game, the model predicts the processor's execution log token by token, effectively acting as the brain of the retro console. It is surprisingly slow, taking minutes to generate a few seconds of gameplay, but the result is exact.

📰 https://news.ycombinator.com/item?id=50306442

🆔 @hackernewsgithubprojects

📦 datawhalechina/agentic-rl-lab

Agentic RL Lab: Train AI Agents for Pennies

Turn a paper on agent learning into a working experiment for less than the cost of a drink. Agentic RL Lab breaks down complex training methods like GRPO and Search-R1 using plain code and a remote cloud service. Instead of wrestling with eight expensive graphics cards or tangled engineering frameworks, you run these models on a single consumer GPU. The project focuses on one key idea: proving that the math works with just twenty training steps. This low-cost approach lets you verify real improvements in agent behavior without a massive budget. It is the perfect way to understand how language models learn to act, not just generate text.

🆔 @hackernewsgithubprojects

📦 ferluht/fm1-sampler

Turn Your Phone Into a Hardware Sampler with FM1 Sampler

Turn your phone into a full-featured, eight-track sampler by pairing it with a small piece of hardware called the M-VAVE FM-1. This project, named FM1 Sampler, handles all the complex audio processing on the phone while the hardware unit acts as a tactile control surface with physical knobs, buttons, and a tiny screen. The real magic here is the tape-style looper. You can record loops that all share a single moving pointer, meaning when you turn a knob to change the speed, every track speeds up or slows down together, just like spinning a physical tape reel. It’s a unique way to manipulate time in music without using a computer.

🆔 @hackernewsgithubprojects

📦 chris124567/cfsolve

How to Beat the Bot Check Without a Browser

Cloudflare Turnstile is the security gate that decides if you are a human or a script, and most solvers need a full web browser to get through it. CFSolve is a Rust library and command line tool that cracks this code by pretending to be a real iPhone running Safari. It does not open a window or render a page; instead, it performs a full reverse engineering of the challenge logic to answer every question directly. It reconstructs the exact timing, fingerprints, and network signals that a real browser would send, all without launching a single tab.

📰 https://news.ycombinator.com/item?id=50277379

🆔 @hackernewsgithubprojects

📦 joeblau/subvox

Read Your Silent Words with AirPods

Turn quietly mouthed words into text using only the motion sensors inside your AirPods Pro. Subvox is a Mac prototype that ignores audio entirely, relying instead on head and jaw movement to recognize a small set of silent phrases. It uses simple pattern matching and statistical tests to determine if the physical motion actually carries meaningful information, proving that you can communicate without making a sound. It is not a voice assistant, but a proof of concept for silent typing. The takeaway is that your head movement contains enough data to decode speech, opening the door to private, silent communication on your computer.

🆔 @hackernewsgithubprojects

📦 aalmenar/baikal-docker

Baikal: Your Own Calendar Server

Baikal is the self-hosted calendar server that finally puts your schedule in your hands. Instead of renting your free time to a cloud giant, this project gives you a ready-to-go Docker container that runs the entire Baikal software on your own hardware. It handles the tricky parts like database connections and email invitations, so you just have to point your phone at your home server and start syncing. This is perfect if you want to own your data without becoming a full-time sysadmin. You get a private, reliable place for your events that actually works with standard calendar apps. Stop worrying about privacy or service outages.

🆔 @hackernewsgithubprojects

📦 inclusionai/idm

IDM: Ultra-Fast Image Generation

IDM is the image generation method that finally makes one-step speed look as good as slow, multi-step results. Instead of relying on an external, heavy encoder to judge quality, this approach uses the model’s own internal features to guide the creation process. The result is stunning, legible text and crisp details generated in a fraction of a second. It solves the age-old problem where speed usually meant blurry or broken characters, making it perfect for instant creative workflows. You get high-fidelity portraits and posters without the wait. If you care about fast, accurate visuals, this is the tool to watch.

🆔 @hackernewsgithubprojects

📦 niazmorshed2007/reanimate

Reanimate: The AI Tool That Clones Any Website Animation

Reanimate is the coding tool that finally lets you copy a website animation exactly. Point it at any moving piece on a live site or video, and it watches every frame to learn how it moves. It then writes the code to recreate that motion in your own project, using the tools you already have. The coolest part is that it checks its own work. It records your new version and compares it to the original until the timing and movement match perfectly. It stops guessing and starts proving the copy is right. You get a perfect starting point to make your own.

🆔 @hackernewsgithubprojects

📦 thomaslehmann1234/datenerfassung-ortsnetz-auslastung

Datenerfassung Ortsnetz Auslastung: Your Home Grids Data

Your smart meter quietly tracks voltage spikes, yet your house battery stays in the dark about grid stress. This open-source collection fixes that gap by linking devices like Shelly plugs, Victron inverters, and Home Assistant directly to a public grid monitoring service. The system sends anonymous voltage readings every five minutes without needing any API key or personal login. In return, the service calculates a live traffic light status and tells your battery exactly when to charge or discharge to avoid peak load. This simple data loop helps stabilize the local power grid using equipment you already own.

🆔 @hackernewsgithubprojects

📦 cloudflare/moltworker

Moltworker: Your Personal AI Assistant in the Cloud

Moltworker is the cloud deployment that finally makes your personal AI assistant truly hands free. Cloudflare built this experimental project to run OpenClaw, a flexible assistant that connects to chat apps like Telegram and Slack, directly inside their Sandbox containers. Instead of keeping a computer on all day, the assistant lives in the cloud, handling tasks like browsing the web or managing messages without your local hardware working. The most surprising part is how it handles memory. A simple backup system automatically saves your chat history and settings to remote storage every five minutes.

🆔 @hackernewsgithubprojects

📦 0xsero/ascii-art

ASCII Art: Photos Turned Into Animated Dust

A static photograph can be transformed into a living, breathing loop where the subject literally dissolves into wind-blown dust. This is how the ASCII Art project works. It is a complete toolkit that turns any image into a stylized figure made entirely of dots on a fixed grid. The system analyzes the light and shadow of your photo to create a two-tone map, preserving the core shape while discarding unnecessary color details. From there, it builds a self-contained web page that animates these dots. You can watch a bust crumble into particles, or see tentacles grow from stone, all rendered in a seamless, hypnotic loop.

🆔 @hackernewsgithubprojects

📦 margrop/wrong-notebook

Wrong Notebook turns test mistakes into lessons

Wrong Notebook is the self-hosted app that turns your test mistakes into a real study plan. It solves the problem of simply copying down wrong answers by using AI to recognize the question from a photo, provide a full explanation, and tag it with specific knowledge points. The coolest part is that it generates step-by-step animated solutions with voice narration, so the app literally talks you through the fix. Because you run it on your own hardware, your study data stays private. If you want to stop memorizing errors and actually understand them, this tool makes the process interactive and clear.

🆔 @hackernewsgithubprojects

📦 kirklandsig/halo2-decompiled

Halo 2 Decompilation

Take the entire retail version of Halo 2, which is just a blob of machine code on an old Xbox disc, and rebuild it from scratch into readable C and C++ code that produces the exact same bytes, every single time. This is Halo 2 decompiled, and it is a massive, ongoing effort to reverse engineer the game so it can eventually run natively on PC, Linux, or handheld devices. It is not a mod or a hack; it is a forensic reconstruction of the original source code, verified to match the released binary perfectly.

🆔 @hackernewsgithubprojects

📦 wici-ai/inferssd

Run Massive AI on Your SSD

Run a seven hundred billion parameter AI model directly from your solid state drive, bypassing the need for massive amounts of random access memory. This tool is built on top of the popular Llama point Cpp engine and specializes in running specific types of large language models called mixture of experts. Instead of trying to squeeze the entire model into your computer’s main memory or graphics card, it cleverly streams only the parts of the model needed for each prediction from your fast storage drive.

📰 https://news.ycombinator.com/item?id=50308376

🆔 @hackernewsgithubprojects

📦 muaaz9329/cheap-ag

Cheap Agent: 50x Your Claude Usage

Stop paying your most expensive AI to type boilerplate. Cheap Agent fixes the org chart by letting Opus plan the work while a dirt-cheap model handles the actual coding. This tool runs those cheap tasks in a private copy of your project, forcing the model to pass a real test before any code lands in your main folder. The result? You get the same shipping quality with over sixty percent fewer tokens, meaning your twenty-dollar plan suddenly feels like a hundred. It is the perfect way to stretch your budget without sacrificing quality. Your code stays clean, the tests prove the work, and the expensive model stays free to think.

🆔 @hackernewsgithubprojects

📦 trufae/parla

Parla: Chat via Email, Natively

Chat with anyone using just their email address, without handing your messages to a big tech server. Parla is a native desktop app for Gnome that makes this encrypted, decentralized chat feel like a local program instead of a website. It runs on Linux, macOS, and Windows, giving you a familiar interface where you can pick your own layout and send stickers that stay on your device. The coolest part is that it runs small, offline web apps directly inside your chat, turning a simple message into a functional tool. It’s a clever way to keep your conversations private and in your hands.

🆔 @hackernewsgithubprojects