📦 itsmeadarsh2008/opencode-senses

OpenCode Senses

Give your text-only coding models the ability to see by running a local vision layer right on your own machine. By connecting opencode-senses to your session, any attached screenshot or mockup instantly gets analyzed, giving your model exact text extraction, coordinates for user interface elements, and precise pixel measurements. Because it runs locally using the Moondream model on your own graphics card, your private files and desktop images never leave your computer. You can easily spot design changes, pull code straight from a mock image, and debug visual layout errors in seconds. Grab the package from node package manager to give your workspace a powerful, private set of eyes.

📰 https://news.ycombinator.com/item?id=49289890

🆔 @hackernewsgithubprojects

📦 charlesmod/infinite-tapedeck

Infinite Tapedeck

Your local music collection is about to become a radio station that never ends, powered entirely by your own graphics card. Infinite-tapedeck uses a clever local artificial intelligence pipeline to analyze your existing music folders, mapping out your exact personal taste in beats, keys, and vibes. It then automatically generates entirely new tracks matched to your library using local music models. A built-in critic automatically scores and rejects any tracks that do not fit your style before you even hear them. It is a brilliant way to enjoy customized, endless offline music, all running through a beautiful, responsive cyberpunk cassette deck interface.

🆔 @hackernewsgithubprojects

📦 zylwithxy/hrdit

HRDiT: Training-Free 4K AI Images

Hrdit is the smart tool that finally lets you generate massive four-K images using standard diffusion transformer models without any extra training. Usually, if you try to force models like Flux to generate giant resolutions, you get a chaotic, repeating mess, and your computer slows to a crawl. This project brilliantly solves both problems. First, it aligns positional data so the AI actually understands where things go in a massive frame. Second, it cleverly prunes unnecessary attention computations to keep generation times incredibly fast. You get gorgeous, ultra-high-resolution images straight out of the box with zero fine-tuning.

🆔 @hackernewsgithubprojects

📦 meta-models/meta-oss-cookbook

Meta OSS Cookbook

Build and deploy fully offline AI agents on the graphics card you already own. The meta-oss-cookbook repository is a hands-on guide for running Meta's Muse Glimmer, an open-weight model designed specifically for local, multi-step agent tasks. It gives you the exact recipes to set up tool-use loops, handle structured outputs, and configure model servers like ollama or llama cpp without needing cloud APIs. The best part is the step-by-step agent loop that lets the model plan, call local tools, and self-correct entirely on your machine. Grab the code, run the quickstart, and start building private, self-correcting agents today.

🆔 @hackernewsgithubprojects

📦 localgpt-app/localgpt

localgpt

You can actually build fully explorable 3D worlds with natural language using localgpt. This clever Rust project lets you prompt a local engine to generate complete interactive scenes with customized geometry, PBR materials, lights, spatial audio, and active behaviors. Instead of relying on heavy cloud services, it runs entirely on your own machine, combining a local AI assistant with a powerful game engine backend. You get a secure, sandboxed playground with smart memory that learns your style over time. It is perfect for quickly bringing wild concepts to life, giving developers a fast, private way to prototype 3D spaces.

🆔 @hackernewsgithubprojects

📦 validatedev/brewservicesmanager

Brew Services Manager

Control your background databases and web servers directly from the macOS menu bar without opening a terminal window. This neat utility sits quietly in your menu bar, giving you an instant visual status of all your Homebrew services. Instead of typing command line instructions to start, stop, or restart background tasks, you can manage them all with a simple click. The absolute best part is how it automatically scans running processes to detect open ports, letting you save custom shortcuts directly to local web interfaces like admin dashboards. It is the perfect, stress-free way to keep your local development environment organized.

🆔 @hackernewsgithubprojects

📦 p2r3/convert

Convert

Most online file converters force you to upload your private files to a distant server, but this brilliant browser-based tool processes everything directly on your own device. The project, called convert, is a truly universal web-based file converter that keeps your data completely secure by running entirely in your browser. Using advanced web technologies like WebAssembly, it can translate files across completely different mediums, such as transforming vector graphics, audio files, and even code documents without a server intermediating. It provides a private, fast, and remarkably flexible way to manage your files. Try hosting it locally to handle your next tricky conversion entirely offline.

🆔 @hackernewsgithubprojects

📦 titanwings/colleague-skill

Colleague Skill

Distill the knowledge, communication style, and decision-making framework of any real person into a custom artificial intelligence skill that you can converse with directly inside your favorite developer agents. By analyzing written materials like design documents, emails, public interviews, or even exportable chat history, this system extracts both a person's technical workflow and their unique conversational personality. It supports three distinct families, letting you preserve a departing coworker's professional playbook, replicate a public figure's mental models, or even capture the familiar voice of a close friend. It is an amazing way to keep essential workplace context alive and interactive.

🆔 @hackernewsgithubprojects

📦 whiteguo233/openbiliclaw

openbiliclaw

Your personal computer can now run an intelligent scout that learns exactly what you like and hunts down the best content across platforms like YouTube, Reddit, and Bilibili. This project, called openbiliclaw, is a local, privacy-first artificial intelligence assistant that gets to know your taste without sending your private data to corporate servers. Instead of relying on addictive platform algorithms to feed you whatever keeps you clicking, this smart assistant proactively searches the web and your favorite social networks to bring you curated recommendations. It is like having a loyal digital companion that works entirely for you. Grab it on GitHub and reclaim your feed today.

🆔 @hackernewsgithubprojects

📦 tiann/hapi

Run AI Coding Agents Remotely with hapi

Run your favorite terminal-based AI coding assistants locally on your computer and control them remotely from your phone or web browser without losing any progress. This clever tool wraps your existing command line agents like Claude Code or Codex, giving you a secure, real-time connection to your terminal from anywhere. If you need to step away from your desk, you can monitor the AI's progress and approve file edits with a single tap on your mobile device. It even supports voice commands so you can talk to your workspace hands-free. Keep your coding sessions alive and stay in control wherever you go.

🆔 @hackernewsgithubprojects

📦 liustack/modlens

ModLens

ModLens acts as a plug-in vision engine that gives sight to text-only artificial intelligence models like DeepSeek and GLM. Instead of forcing you to save an image and manually pass a file path, this tool lets you paste images straight into the chat to get instant, structured JSON evidence back. It operates with a light touch, requiring no local proxy running in the background and no changes to your existing configurations. By setting up a free Gemini key or utilizing its built-in engine chain, your text-only AI agent can read graphics, diagrams, and code snippets on the first try.

🆔 @hackernewsgithubprojects

📦 decionis/agent-safe-pipeline

Agent-Safe Pipeline

AI agents are incredibly smart at proposing actions, but they should never be allowed to approve them. The agent-safe-pipeline project solves this security challenge with an elegant reference architecture that completely isolates agent reasoning from actual execution. Instead of giving an agent direct access to downstream credentials, this system captures its proposed action as an immutable intent, passes it to an independent policy engine for a clear verdict, and then verifies human approval before a sealed executor runs the action. It completely removes the risk of rogue AI decisions by making sure that agents can only suggest actions, while strict policy and human oversight always have the final say.

🆔 @hackernewsgithubprojects

📦 justvugg/lumabri

Lumabri

Lumabri is the decentralized artificial intelligence engine that lets you run massive mixture of experts models across a swarm of ordinary computers. Instead of requiring massive, expensive graphics cards to host giant neural networks, this pure C tool splits artificial intelligence tasks at the expert level. When you chat with a model, only a tiny four-kilobyte activation travels to whichever peer holds that specific active expert. The network can recruit from ordinary processors and solid-state drives, making decentralized hosting accessible to everyone. Every single byte and calculation is cryptographically verified for absolute safety, delivering identical results to a local run.

📰 https://news.ycombinator.com/item?id=49293523

🆔 @hackernewsgithubprojects

📦 raiyanyahya/superlearn

Superlearn

You can now transform any complex topic into a fully interactive, local learning dashboard directly from your terminal using Claude Code. Superlearn is a clever plugin that takes your prompt, researches the live web for papers and videos, and designs a gorgeous local web app tailored to that specific subject. It is perfect for visual learners, featuring auto-generated mindmaps, interactive charts, and runnable code blocks right on the page. You can even drag cards around on a spatial whiteboard canvas or study cross-board flashcards. Just type a slash command in your terminal to start building your own self-hosting knowledge library today.

🆔 @hackernewsgithubprojects

📦 bariona/fact

FACT Robot World-Action Model

Robots learn much faster when they are allowed to fail and understand the consequences of their mistakes. A new system called fact turns failure into a teaching tool by using a single causal diffusion transformer that lets a robot act first, then imagine what happens next. Normally, models get confused trying to predict actions and future video at the same time, but fact keeps them separate so the robot can plan actions without getting distracted by its own imagination. Even better, it trains on failed attempts without copying the mistakes, helping the robot visualize what went wrong so it can perform better in the real world.

🆔 @hackernewsgithubprojects

📦 jacckma/pipo_verl

pipo_verl

Large language models can now dynamically adjust their own learning pace during reinforcement training by analyzing their own real-time progress. The pipo_verl repository introduces a plug-and-play training framework that fine-tunes artificial intelligence policies using direct closed-loop feedback. Instead of relying solely on static rewards, it measures performance gains between training runs to selectively boost or suppress learning signals on the fly. This adaptive approach works across complex mathematical reasoning, competitive programming, and tool-use tasks. By adding this layer of self-correcting feedback, developers can train smarter, more resilient models that actively correct their course during the training cycle.

🆔 @hackernewsgithubprojects

📦 opensandbox-group/opensandbox

OpenSandbox

Run untrusted code and heavy-duty browser automation securely inside isolated containers without risking your host machine. OpenSandbox gives you a ready-to-go runtime designed specifically for AI agents, packing everything from python code execution to full desktop VNC environments into protected sandboxes. Instead of worrying about an AI accidentally deleting your files or leaking api keys, this platform keeps everything locked down using strong container isolation. It even features a built-in credential vault to inject secrets safely. Grab their command line tool or multi-language SDKs to spinning up isolated environments instantly. Check out opensandbox and build safer AI apps today.

🆔 @hackernewsgithubprojects

📦 tnm/zclaw

zclaw

Run a personal artificial intelligence assistant directly on a tiny microcontroller. With zclaw, you can flash a complete voice and text agent onto inexpensive boards like the Seeed Studio Xiao, keeping the entire footprint under eight hundred eighty-eight kilobytes. Instead of managing complex servers, this firmware lets you interact with models from OpenAI, Anthropic, or local endpoints like Ollama over Telegram. From there, you can use natural language to control physical pins, schedule tasks, read temperature sensors, and even write persistent memory. It is a incredibly clever way to bridge cloud intelligence with real-world physical hardware.

🆔 @hackernewsgithubprojects

📦 hackingdave/btrpa-scan

btrpa-scan

btrpa-scan is the Bluetooth scanner that unmasks devices trying to hide behind randomized addresses. When nearby gadgets rotate their MAC addresses using privacy protocols, this clever tool uses identity resolving keys to decode and track them in real time. It is perfect for hunting down specific hardware, showing you estimated distances, signal history, and even pinpointing locations with GPS integration. You can watch it all unfold on a beautiful, matrix-themed browser radar screen with live signal sparklines and proximity alerts. It is the ultimate playground for anyone curious about the invisible wireless traffic buzzing all around us.

🆔 @hackernewsgithubprojects

📦 a-streetcoder/agent-deck

Agent Deck

Running terminal coding assistants on a Mac usually means dealing with messy web interfaces or tracking hidden changes. Agent Deck solves this by wrapping the Pi coding assistant in a completely native macOS desktop workspace. Built natively with Swift and SwiftUI, it lets you visually track what your agent is doing in real time. You can organize parallel tasks, manage custom skills, and see exactly which files are being modified before any writes happen. It is a clean, visual command center that gives you total control over terminal-based AI coders without the clutter.

📰 https://news.ycombinator.com/item?id=48361456

🆔 @hackernewsgithubprojects