📦 zhou-weichen/geosubprobe

Geosubprobe

Geosubprobe is the diagnostic toolbox that finally reveals exactly how self-supervised vision models understand the physical, three-dimensional geometry of the images they see. Instead of treating these advanced vision models like black boxes, this tool trains specialized probes to dissect their internal layers and analyze how they represent spatial concepts like depth and surface angles across different datasets. By inserting targeted interventions directly into the model's internal subspaces, researchers can pinpoint where 3D awareness actually forms during training. It is an invaluable way to audit and compare different vision backbones to see which ones build the most accurate mental map of our physical world.

🆔 @hackernewsgithubprojects

📦 yuzu-hub/appscreen

Appscreen

Appscreen is the browser-based design tool that finally makes creating gorgeous App Store screenshots incredibly simple. Instead of wrestling with complex graphic design software, you just drag and drop your raw app images directly into your browser. The coolest feature is its interactive 3D mode, which lets you wrap your screenshots around an iPhone mockup and tilt, rotate, or position it in 3D space with simple mouse drags. You can easily customize gradients, add polished text overlays, and even translate your marketing copy into multiple languages automatically using built-in AI. It is the perfect, hassle-free way to build professional app store graphics in minutes.

🆔 @hackernewsgithubprojects

📦 mims-harvard/athena

Athena AI

Athena is the medical AI assistant that actually researches patient treatment decisions before answering. Instead of relying on raw memorization or guessing like typical chatbots, this system connects directly to a digital library of over two hundred validated biomedical databases and tools. When you ask a complex medical question, the agent takes a step-by-step approach, systematically looking up drug interactions, clinical warnings, and patient dosage guidelines to build its logic. It is so effective that it outperforms massive commercial AI models on treatment-planning tests. It is a fantastic tool for researchers wanting to see how specialized AI can safely reason through complex science.

🆔 @hackernewsgithubprojects

📦 rohitg00/tailclaude

TailClaude

TailClaude is the secure bridge that finally frees Claude Code from the terminal and brings it straight to your web browser. Instead of squinting at SSH terminals on a tiny phone screen, this clever tool lets you publish your coding assistant interface to any device over a secure connection. You can instantly access a beautiful, touch-friendly UI complete with streaming markdown answers, model switching, and real-time cost tracking just by scanning a QR code on your terminal. It is the ultimate way to code on the go without sacrificing the visual comfort of a modern web application.

🆔 @hackernewsgithubprojects

📦 shweta-mishra-ai/tokenmizer

Tokenmizer

Keep your AI assistant's memory alive across different chat sessions without wasting thousands of tokens. This local proxy sits between your favorite coding tools and any large language model to build a live knowledge graph of your project decisions. Instead of re-uploading massive files or re-explaining your code history every time you start a new chat, it compresses your context and saves your active decisions into a structured memory graph. You can resume an entire day's work using just a fraction of the usual context window, saving you time and keeping your API costs remarkably low.

🆔 @hackernewsgithubprojects

📦 dgominski/treematch

Treematch

Counting millions of individual trees from space usually requires flawless, hand-labeled data that takes years for experts to create. A new framework called treematch changes this by using mathematical optimal transport to train AI models on a mix of highly accurate expert maps and highly imperfect, noisy computer-generated guesses. By applying a smart slack mechanism, the algorithm automatically down-weights the noisy labels, meaning it can ingest massive satellite datasets without being thrown off by errors. This opens the door to tracking global forest health and carbon capture on a massive scale using easily accessible satellite imagery.

🆔 @hackernewsgithubprojects

📦 wassermanproductions/motion-previs-studio

Motion Previs Studio

Motion Previs Studio is the open-source desktop app that extracts perfectly clean camera movement and actor poses from any video clip. If you are experimenting with AI-video pipelines, you know how hard it is to get consistent control layers. This app runs completely locally to solve camera tracking while smart-masking out the subject, meaning you get the exact background camera pan or tilt separated from the actor. It instantly exports full production packs with OpenPose skeletons, depth maps, and even Python scripts to recreate the exact setup in Blender. It is an absolute powerhouse for previsualizing your shots before burning generation credits.

🆔 @hackernewsgithubprojects

📦 amilich/isometric-city

Isometric City

Build and run your own bustling metropolis right in your web browser with isometric-city. This project lets you design a lively, simulated world complete with functioning roads, railways, and flight paths, all rendered beautifully on an HTML5 canvas without relying on bloated external game engines. You will watch autonomous cars navigate traffic lights, see pedestrians walk the streets, and manage zoning to grow your city's economy. The best part is that it even features a built-in theme park mode where you can construct custom roller coasters and watch simulated guests run around. It is a wonderfully polished sandbox that shows just how powerful modern web tools can be.

🆔 @hackernewsgithubprojects

📦 vinayaklatthe/microsoft-security-skills

Microsoft Security Skills

Your AI assistant can now give precise, expert-level Microsoft security guidance instead of generic safety advice. The microsoft-security-skills project is a packaged capability layer that loads eighty-eight curated security skills directly into compatible AI coding tools like GitHub Copilot, Claude Code, and Cursor. It solves the massive headache of AI hallucination by grounding your assistant in official, public Microsoft documentation. Rather than giving vague recommendations, your agent gets step-by-step decision trees and guardrails to design solid conditional access baselines, set up data loss prevention, and configure firewalls without making costly configuration mistakes. It is an instant expertise upgrade for your daily workflow.

🆔 @hackernewsgithubprojects

📦 muratcankoylan/agent-skills-for-context-engineering

Agent Skills for Context Engineering

Your AI agents are likely wasting their limited attention budget on useless tokens, causing them to forget critical instructions mid-conversation. The agent-skills-for-context-engineering repository solves this by providing a highly optimized collection of plug-and-play skills designed to manage an LLM's context window. Instead of basic prompting, it introduces smart strategies like progressive disclosure, which only loads full skill details when absolutely necessary. This keeps the agent's attention sharp and prevents performance degradation over long sessions. It is a brilliant way to build highly efficient, reliable agent systems that actually stay on track.

🆔 @hackernewsgithubprojects

📦 miguok/fable-harness

Fable Harness

Fable Harness is the behavioral protocol that finally makes your AI coding assistant work like a disciplined engineer. Instead of letting your agent make sloppy, unverified guesses, this clever tool injects a strict routine directly into your coding sessions. Its signature move is a multi-party adversarial review. Before making any critical decision, it spins up three independent sub-agents in parallel—a skeptic, a red-team, and a simplifier—who aggressively cross-examine the plan from every angle. It even blocks the agent from finishing a task if it modified code without running an automated test, keeping your pipeline safe and your agent honest.

🆔 @hackernewsgithubprojects

📦 techenthusiast167/d4rk_intel-osint-investigative-toolkit

D4rk Intel OSINT Investigative Toolkit

An open-source intelligence toolkit can piece together a complete digital footprint using only publicly available internet records. The d4rk_intel-osint-investigative-toolkit brings together a curated collection of scripts, resources, and investigative methodologies to help analysts track down open-source information. Instead of wasting hours manually searching across dozens of disconnected databases, investigators and journalists can use these structured resources to quickly gather and analyze public data for ethical research. This toolkit simplifies complex online investigations, making it easier than ever for researchers to uncover hidden connections across the web and verify digital evidence efficiently.

🆔 @hackernewsgithubprojects

📦 dicklesworthstone/franken_ocr

franken_ocr

You can now run a powerful three-billion parameter document-parsing model on a basic CPU using nothing but a single, lightweight Rust binary. The project is called franken_ocr, and it is a pure-Rust, CPU-only engine designed specifically to run Baidu’s Unlimited-OCR model. Instead of relying on bloated machine learning frameworks, Python, or heavy GPUs, it uses custom-written, highly optimized kernels and a specialized mixed int4 and int8 quantization format. By stripping away general-purpose framework overhead and focusing on this single model’s exact shapes, it delivers incredibly fast, memory-safe text extraction on standard hardware. It is a brilliant glimpse into the future of ultra-portable, self-contained local AI.

🆔 @hackernewsgithubprojects

📦 it-baer/proxmorph

ProxMorph

You can now customize your Proxmox dashboard with twenty-two stunning native themes like Dracula, Catppuccin, and Nord. ProxMorph integrates directly into the built-in color theme selector for Proxmox VE, Backup Server, and Datacenter Manager. It goes way beyond basic style sheets, featuring a clever hybrid engine that injects custom JavaScript to style your charts and even adds real-time hardware sensor monitoring for temperatures, fan speeds, and UPS status right on your node summary. Best of all, an automatic update hook keeps your themes perfectly patched even after system updates. Run the quick one-liner installer and make your hypervisor look incredible.

🆔 @hackernewsgithubprojects

📦 maximqaxd/wince-dc

wince-dc

Boot a fully functional, multitasking windowed desktop directly on a Sega Dreamcast. The wince-dc project takes the classic console's hidden, on-disc Windows CE runtime and transforms it into a self-contained operating system you can actually play with. It delivers a custom desktop shell with a Start menu, taskbar, file explorer, and multitasking apps running in their own processes, all navigated using the Dreamcast controller. It even includes a reconstructed networking stack to help revive dead game servers. You can easily build a bootable disc image from source with a single command, bringing retro console hardware to life in a whole new way.

📰 https://news.ycombinator.com/item?id=48785840

🆔 @hackernewsgithubprojects

📦 sgl-project/sglang-omni

Sglang Omni

Sglang-omni runs native, high-fidelity forty-eight kilohertz speech models with streaming speed by splitting generation across specialized hardware stages. Normally, running advanced voice assistants requires cramming speech synthesis, text generation, and audio processing into a single slow loop. This project solves that by managing the entire pipeline, letting different models and hardware stages work together. It coordinates everything from encoders and autoregressive engines to streaming vocoders, passing data smoothly through shared memory. It means you can build voice agents that talk back instantly without lag, using open models through a standard, easy-to-use API.

🆔 @hackernewsgithubprojects

📦 agentphone-ai/agentphone-mcp

agentphone-mcp

agentphone-mcp is the developer toolkit that finally bridges the gap between local AI assistants and the actual telephone network. Instead of wrestling with complex telecom APIs, you just plug this Model Context Protocol server into your setup to let your assistant buy real phone numbers, send texts, and place voice calls using simple English. It connects directly to your favorite clients, meaning you can literally tell your AI to purchase a local number, assign it a custom voice, and call to book your next appointment. It handles all the messy backend routing so you can build fully functional voice agents in minutes.

🆔 @hackernewsgithubprojects

📦 shiva108/ai-llm-red-team-handbook

ai-llm-red-team-handbook

Security teams are discovering that traditional penetration testing methods completely fail when applied to generative artificial intelligence. The ai-llm-red-team-handbook is a massive, forty-six-chapter operational curriculum designed to solve this problem by mapping out every known angle of AI vulnerability. It compiles deep strategic guides with hands-on tactical playbooks covering specialized attack vectors like training data poisoning, model theft, and prompt injection. Instead of just theorizing, the repository provides concrete testing methodologies and automated Python scripts to help security consultants and engineers audit large language models before they deploy. This is the ultimate open-source manual for securing the future of AI.

🆔 @hackernewsgithubprojects

📦 shlokkhemani/rabbithole

Rabbithole

Rabbithole is the infinite learning canvas that transforms plain text selections into branching nested documents. When you are studying a complex topic, you simply open your document, highlight any phrase, and ask a question. The answer immediately streams in as a new child document on your screen, letting you branch off into deeper topics as recursively as you want. It runs entirely locally on your machine as an MCP server for terminal assistants, connecting your AI agent's logic to a beautiful browser canvas with auto-layout, canvas zoom, and follow-up chat. It is a brilliant way to explore tough concepts without losing your train of thought.

🆔 @hackernewsgithubprojects

📦 eric-terminal/etos-llm-studio

ETOS LLM Studio

Run custom open-source AI models entirely offline right on your iPhone or Apple Watch. With etos-llm-studio, you can load local files to run private conversations without any internet connection, or connect to cloud services like OpenAI and Claude if you prefer. The app packs advanced features like cross-device syncing, smart widgets, and a system that pulls together custom tools and personal notes. It is a completely self-contained setup built with Swift that lets you carry a fully customizable assistant right on your wrist or in your pocket. Grab your own model files and start chatting privately today.

🆔 @hackernewsgithubprojects

📦 frenzymath/danus

Danus Mathematical Reasoning Swarm

Danus is the multi-agent system that orchestrates a swarm of autonomous codex workers to solve research-level mathematical proofs. Instead of relying on a single AI to get a complex proof right on the first try, this clever engine divides and conquers. A main agent coordinates the strategy while autonomous workers focus on proving tiny, individual lemmas. Every single claim must pass a strict, stateless verifier before it enters a shared, content-addressed fact graph. Once all the pieces are verified and linked, the system automatically compiles the entire chain into a publishable LaTeX paper or a clean progress report.

🆔 @hackernewsgithubprojects