📦 ielab/skim-search-agent

SkimSearchAgent: Deep Research for Your Data

SkimSearchAgent lets you build intelligent research assistants that dig through your own document collections to answer complex questions without you writing code. Instead of guessing answers from a static database, this tool gives an AI model a loop where it can actively search, inspect results, and fetch specific sections of your documents to piece together accurate answers. It is fascinating because it separates every part of the process, meaning you can swap out different search methods, document structures, or even the AI brain driving the logic without breaking the whole system.

🆔 @hackernewsgithubprojects

📦 petroni-lab/librarian

AI Librarian: Natural Questions to Scientific Evidence

Ask a messy biology question like whether metformin extends lifespan and get back the exact sentences that answer it, not just a list of papers. This project acts as a smart librarian for artificial intelligence, taking your natural language query and turning it into a series of precise search commands across the Europe PMC database. It doesn't just dump results; it reads through the actual full text of open-access studies, filters out the noise, and extracts only the specific evidence snippets that matter. Think of it as a research assistant that skips the fluff and hands you the hard proof.

🆔 @hackernewsgithubprojects

📦 198808xc/dramasr-lrm

DramaSR-LRM: AI That Guesses Who Is Talking in TV Dramas

DramaSR-LRM turns large language models into sharp listeners that figure out who is speaking in long, messy TV dramas. Traditional speech recognition fails when voices change or characters disappear off-screen. This project solves that by training a model to think before it speaks. Instead of guessing instantly, the AI calls tools like audio comparison and character relationship maps to build a solid argument. It learns to weave together visual clues and voice patterns, catching subtle shifts that standard software misses. The result is a system that actually understands context, not just sound. It is a clever way to make AI pay attention to the story, not just the audio.

🆔 @hackernewsgithubprojects

📦 apphane-dev/nehir

Nehir: Scrolling Window Manager for macOS

Nehir lets you manage your macOS desktop using a horizontal scrolling layout that feels like a river of windows. Instead of stacking apps on top of each other or jumping between static workspaces, Nehir arranges your open windows into columns that slide left and right across your screen. It brings the popular Niri workflow to Mac, meaning you get smooth, animated transitions between apps without losing context. You can customize hotkeys, app rules, and monitor setups with simple text files that update instantly. It handles multiple displays well, keeps your dock out of the way, and even lets you trace exactly what caused a glitch if something goes wrong.

📰 https://news.ycombinator.com/item?id=49194123

🆔 @hackernewsgithubprojects

📦 raven-july/cfpo

CFPO: Smarter Multimodal Reasoning

CFPO is a counterfactual reinforcement learning framework that finally forces large vision-language models to actually look at images instead of guessing from text. Standard AI methods often let models cheat by relying on linguistic shortcuts rather than visual evidence, and CFPO solves this by testing whether the prediction changes when critical visual cues are suppressed. It builds a counterfactual path that blocks high-saliency visual signals, then rewards the model only if its reasoning genuinely depends on the image rather than blind language patterns. This approach consistently beats standard training baselines on math and logic benchmarks without needing external reward models.

🆔 @hackernewsgithubprojects

📦 larryvrh/comfyui-minimax-h3-turbo

ComfyUI-MiniMax-H3-Turbo

Create synced video and audio in just four generation steps using the ComfyUI-MiniMax-H3-Turbo project, which transforms how you handle multimedia generation. The repository provides a custom sampler and LoRA loader that bypass the usual twenty-step process, cutting generation time to a fraction of what it normally takes. It solves the tricky problem of audio distortion that usually wreals havoc when you rush video models, ensuring your sound stays crisp even at ultra-low step counts. By dropping two simple nodes into your existing workflow, you get fast results without complex tweaks. This is a neat trick for anyone who wants to experiment with fast-paced multimedia creation without waiting around.

🆔 @hackernewsgithubprojects

📦 aaddrick/attention-control

Attention Control: Air Traffic Discipline for AI

Attention Control forces coding agents to talk like air traffic controllers, stripping away the usual fluff so you get immediate, actionable instructions. It treats your attention like a pilot’s focus, demanding that every response lead with the next executable command instead of a polite preamble. This style was built for readers with ADHD but helps anyone drowning in context switches, by replacing vague suggestions with concrete steps and hard deadlines. You stop guessing and start running code. If your AI assistant is currently writing essays, this style forces it to get out of the way and let you work. It is a simple shift that makes complex help actually usable.

🆔 @hackernewsgithubprojects

📦 beatapi/awesome-minimax-h3-prompts

Awesome Minimax H3 Prompts: Curated Video Generation Library

This repository is a curated gallery of over one hundred ready-to-use text prompts designed specifically for the MiniMax H3 video generation model, turning simple descriptions into playable video clips. Instead of just listing text, it provides full production details including camera movement, lighting styles, and character interactions, covering diverse genres like action, anime, and product demos. It solves the problem of trial-and-error prompting by offering tested examples with credited sources, allowing creators to quickly generate cinematic sequences, commercials, or viral shorts without guessing the right technical phrasing.

🆔 @hackernewsgithubprojects

📦 1038lab/comfyui-minimax-h3-promptor

Create Cinema Prompts for MiniMax H3

Create cinema-quality video prompts automatically by feeding images into ComfyUI. This tool splits the work into two steps. First, it analyzes your pictures using any AI model you want, like OpenAI or Claude. It extracts details about lighting, characters, and mood, then sends that text report to a second node. That second node takes your creative idea and the visual report to write the perfect prompt. It even guesses whether you want a text-to-video or image-to-video result based on your inputs. You get precise, professional results without paying extra for the AI to look at your images twice.

🆔 @hackernewsgithubprojects

📦 guillermolg00/morphicons

Morphicons: Universal Icon Morphing

Morphicons lets any icon morph into any other using spring physics. Instead of manually declaring rotations, it calculates the optimal path so arrows rotate naturally instead of shrinking. The library handles strokes from major icon sets with zero dependencies and a tiny file size. It works seamlessly in React, Vue, Svelte, and React Native. State lives outside while the component handles the animation details. You get clean server rendering and accessibility defaults out of the box. It solves the messy transition problem with pure math. The core never touches the DOM, keeping everything fast. Even interruptions feel smooth thanks to velocity preservation.

🆔 @hackernewsgithubprojects

📦 frozenpepper/deepseek-and-destroy

DeepSeek and Destroy: Autonomous Coding Agent

Give a coding agent a complete plan, and it executes the entire multi-phase project without stopping to ask for your permission after every single task. DeepSeek and Destroy is a tool that splits work between a smart boss and cheap workers. The boss handles big decisions, while inexpensive agents do the heavy lifting, write code, review each other, and fix bugs automatically. It runs in a strict loop of implement, review, repair, and verify until the job is fully done. This setup saves money and time because the expensive AI only steps in for major choices, leaving the repetitive grinding to faster, cheaper models.

🆔 @hackernewsgithubprojects

📦 openclaw/ffmpeg-wasm

ffmpeg-wasm: Media Processing in Your Browser

ffmpeg-wasm is the browser-native media processor that turns your web app into a full video studio without downloading heavy native binaries. Instead of relying on massive server-side installs, this project compiles the essential parts of FFmpeg into a tiny, lightweight WebAssembly package that runs right inside your browser or Node environment. You can easily extract audio, grab thumbnails, or reformat videos using simple commands that feel just like working with the traditional command-line tools. It handles the tricky heavy lifting of video conversion locally, keeping things fast and private while avoiding the need for complex infrastructure setups.

🆔 @hackernewsgithubprojects

📦 mintdotgg/mint-playground

Mint Playground: Open-Source 3D Experiences

Mint Playground is the open-source collection that turns simple web code into stunning three-dimensional worlds you can actually play with. Instead of just reading about cool visuals, this repository lets you grab ready-made experiences like arcade basketball games, fantasy action adventures, or interactive furniture showrooms and run them instantly in your browser. It solves the boring problem of building complex 3D sites from scratch by providing polished templates that work out of the box. This is useful because you get to see exactly how developers create immersive interactive media without needing to reinvent the wheel every time.

🆔 @hackernewsgithubprojects

📦 tencent/tencentdb-agent-memory

TencentDB Agent Memory

TencentDB Agent Memory is the team library that turns past work into reusable brainpower for artificial intelligence assistants. Most tools just store chat logs, but this system actually learns from your history to build a living skill library. It scans your old conversations and code to extract reusable instructions, structured knowledge, and code maps, then packages them into a portable save file. You can hand this file to any assistant, and it instantly understands your project context, preferences, and past decisions without being retrained from scratch.

🆔 @hackernewsgithubprojects

📦 0xsero/deepseek-v4-flash-0731-spark-sparkinfer

DeepSeek V4 Flash on Spark

Run a massive twenty-six thousand token language model on a single portable NVIDIA Spark laptop using this ready-to-go Docker setup. It packs the heavy DeepSeek V4 Flash engine into a compact container that handles long conversations and strict code generation without crashing. The project carefully manages memory by using smart weight compression and a special drafting technique to keep responses fast and accurate. You simply clone the folder and start the server with one command to get a reliable AI assistant running locally. This makes powerful artificial intelligence accessible on compact hardware without needing complex cloud infrastructure.

🆔 @hackernewsgithubprojects

📦 mudler/vllm.cpp

vllm.cpp: The Fastest C++ AI Engine

vllm.cpp is the C++ engine that brings blazing fast AI responses to any device without needing a server farm. This project recreates the powerful features of the popular Python-based vLLM but strips away the heavy Python layer to run in raw C++. The biggest surprise is its memory management. It uses a system called paged attention, which treats your computer’s memory like a spreadsheet, letting you fit huge AI models onto smaller graphics cards by shuffling data in and out as needed. This means you can run advanced language models on your laptop or even a phone, not just in massive data centers.

🆔 @hackernewsgithubprojects

📦 cursor/sdk-bridge

Cursor SDK Bridge Explained

The cursor sdk bridge lets you run complex AI agents locally without writing cloud infrastructure. It is a tiny server that wraps the official cursor sdk and talks to your app over a simple standard protocol. You spawn it, wait for a ready message, and then send commands to create agents, chat with them, or manage files. It handles authentication, error tracking, and even lets your code define custom tools the AI can call. It is basically a safe, local gateway to cursor intelligence. Check out the docs if you want to build something clever.

🆔 @hackernewsgithubprojects

📦 b1z0n/homelander

Homelander: Auto-Applies to German Rentals

Paste a German apartment search link and watch Homelander automatically poll for new listings while silently filling out the contact forms for you. It runs locally on your computer, so all your data stays put, and it even handles those annoying captcha hurdles so you never miss a new apartment drop. Forget refreshing pages all night or paying for expensive cloud bots; just set your preferences and let this desktop app do the heavy lifting. It is clever, efficient, and genuinely solves the nightmare of competing for housing in busy German cities. Check it out if you want to beat the clock and secure that next place.

🆔 @hackernewsgithubprojects

📦 arteemg/sky-agent

Sky Agent: The Open-Source Superagent Framework

Sky Agent is the open-source framework that lets you deploy your own cloud-based AI agent in just one click. It solves the problem of needing complex infrastructure to run custom assistants by packaging everything into a simple Python tool that connects directly to Telegram or a web interface. The most fascinating part is that it automatically stores your conversations in a local database so it remembers past chats without needing expensive cloud storage. You can also drop text files into a special folder to give your agent extra knowledge instantly. It even includes a benchmark system where an agent rewrites its own instructions to improve its performance over time.

🆔 @hackernewsgithubprojects

📦 akcodez/hackingtool-plugin

HackingTool Plugin: 183 Tools for Claude

HackingTool Plugin is the Claude extension that gives your AI assistant a massive library of 183 professional security testing utilities right out of the box. It wraps the popular HackingTool suite, letting you run complex pentesting and OSINT commands with simple natural language requests. The real magic lies in its smart environment handling: it automatically detects your operating system and decides whether to run tools natively, through Windows subsystems, or inside isolated Docker containers. This means you get the power of advanced vulnerability scanners and network mappers without ever touching a complex installation script or dealing with missing dependencies.

🆔 @hackernewsgithubprojects