📦 raymondhuang210129/llama.cpp-adaptive-kv-streaming

Run Massive LLMs on Small GPUs

This fork of Llama CPP lets you run massive language models on tiny graphics cards by moving the memory burden to your computer's main memory. Normally, long conversations require huge amounts of video memory, causing small cards to crash, but this version intelligently streams data between your hard drive and graphics card in real time. It keeps your GPU busy without freezing, allowing a twenty-seven billion parameter model to run fully on just sixteen gigs of video memory. This is a huge leap for anyone who wants local AI without buying expensive enterprise hardware, proving that smart software can beat raw hardware limits.

🆔 @hackernewsgithubprojects

📦 stanford-tml/robot_keyframe_kit

Robot Keyframe Kit: The Universal Robot Animator

Robot Keyframe Kit is the tool that finally lets any researcher animate any robot without writing code. It plugs into MuJoCo, the standard physics engine for robotics, and gives you a simple visual interface to pose, move, and save a robot’s motion. The standout feature is its mirror mode: move one arm, and the opposite arm mirrors the motion with the correct joint directions handled automatically. That single feature turns hours of manual tweaking into seconds of intuitive design. It’s built for speed, works with any standard robot model, and hands you clean, playable motion data ready for research or simulation.

🆔 @hackernewsgithubprojects

📦 pedrovieira/holeberry

Holeberry: Control Pi-hole from your Mac Menu Bar

Holeberry lets you unblock the specific website you are viewing without turning off your entire network’s ad blocking. It is a native macOS menu bar app that acts as a remote control for your Pi-hole instances, saving you from opening a browser tab every time you need to check stats or tweak settings. The coolest feature is its smart browser integration; it detects the domain in your current tab and allows you to unblock just that one site with a single click. This keeps the rest of your network protected while fixing your immediate issue.

🆔 @hackernewsgithubprojects

📦 anemll/dspark-vllm-gx10

DSpark: Run DeepSeek V4 on Two Desk-Top GPUs

DSpark is the two-node setup that finally makes massive AI models runnable on compact desktop hardware. The main project, DSpark vLLM for two DGX Spark nodes, solves a specific hardware problem: the DeepSeek V4 model is simply too big to fit on a single compact computer, but it runs smoothly when you connect two of them. This repository handles the complex technical setup required to bridge the software and hardware, ensuring the system actually works without crashing. It includes a live dashboard so you can watch performance in real time and provides a fully reproducible deployment process.

🆔 @hackernewsgithubprojects

📦 roadmaus/comfyui-continuity

One Node for All Your AI Video

Render finished AI videos with sound using one single node inside ComfyUI. The Continuity pack strips away the complex wiring and lets you just type a prompt, attach reference images, and hit render. It handles the heavy lifting for six different model families, driving them entirely locally without any API keys or cloud costs. This solves the headache of managing separate workflows for each new model by putting everything in one place. You get the raw creative freedom of open source tools with the simplicity of a single button.

🆔 @hackernewsgithubprojects

📦 amirmushichge/machine-vision

Turn Any Video Into Generative Art with Machine Vision

Machine Vision transforms your local videos into dynamic, generative graphics right in your browser without sending a single frame to a server. It analyzes motion, edges, and light in real-time, then overlays a deterministic field of points, lines, and labels that react to your footage. The coolest part is the privacy: everything stays on your device, letting you experiment with a completely private, creative coding playground. You can tweak the chaos, pick presets, or fully randomize the look, then export a five-second composite clip in seconds.

🆔 @hackernewsgithubprojects

📦 vrgamegirl19/vrgdg-seedvr2-tensorrt-studio

Fix Old Videos Locally with SeedVR Studio

Restore blurry videos without sending them to the cloud. This local tool uses your computer's graphics card to repair and upscale footage, giving you a real-time preview so you can compare the before and after side by side. It is fast enough to handle long clips, and if your render stops, you can pick up right where you left off. It is the perfect way to breathe new life into old home videos or rough game footage, all while keeping your data safe. Just drop in a clip, tweak the settings, and let your GPU do the heavy lifting.

🆔 @hackernewsgithubprojects

📦 alpbahadur/49-ide

49 Agents IDE: The 2D Canvas for AI Terminals

Forget terminal tabs. 49 Agents IDE replaces your cluttered command line with a zoomable, infinite canvas where every terminal lives side by side. You can arrange panes spatially, zoom out for the big picture, and watch AI agents work across multiple machines without ever touching SSH. It’s like a living map of your entire coding setup, accessible from your laptop or phone. Instead of alt-tabbing through hidden windows, you see everything at a glance, with live stats and file editors right on the surface. It turns chaotic agent management into a calm, visual workspace that actually fits your workflow.

📰 https://news.ycombinator.com/item?id=49514828

🆔 @hackernewsgithubprojects

📦 crispstrobe/crispasr

CrispASR: Run Any Speech Model Locally

One of the most surprising facts about modern AI is that you can run complex speech recognition and text-to-speech models directly on your own laptop, completely offline. CrispASR solves the problem of fragmented tools by acting as a single, unified C++ runtime that handles a huge variety of multilingual models, from Whisper to Parakeet and Voxtral, without needing a cloud connection. This is incredibly useful for developers who need fast, private, and on-device audio processing, as it removes the heavy setup usually required to run these diverse models locally. If you are building apps that need instant voice interaction, this is the infrastructure that makes it happen.

📰 https://news.ycombinator.com/item?id=49515564

🆔 @hackernewsgithubprojects

📦 openclaw/wacrawl

wacrawl: Private WhatsApp Backups

wacrawl turns your Mac’s local WhatsApp files into a private, searchable archive without touching the cloud. It copies your chat history into a secure local database so you can hunt down old invoices or lost messages using fast, offline commands. The standout feature is its encrypted backup system, which locks your data with strong cryptography before pushing it to a Git repository, ensuring your private conversations remain unreadable to anyone else. This tool is perfect for developers who want fast local search and safe, verifiable backups without relying on third-party servers.

🆔 @hackernewsgithubprojects

📦 reactor-team/infinite-livestream

This AI Stream Never Stops

Stream an endless AI video show where viewers literally dictate the plot. The project takes a simple chat message, turns it into a script, and renders real video with audio using a powerful model. The coolest part is that the stream never cuts out. It runs continuously, filling gaps with filler content until your viewer's request is ready, creating a broadcast that feels alive and reactive. It’s a neat way to see how we can merge live audience interaction with generative video tech. Try it out and see how your ideas become a scene in real time.

🆔 @hackernewsgithubprojects

📦 krispuckett/murmur

Murmur: AI Orbs That Breathe

Murmur renders living glass orbs in Swift that physically react to your voice and typing. It is a collection of parametric shaders, not static icons, so the visuals change shape based on real-time signals. This tool solves the problem of boring, flat loading states by giving your app a dynamic presence. You can switch between eighteen unique glass styles, like ink, water, or auroras, and they all animate smoothly at sixty frames per second. The package is designed to drop into your code with a single view, requiring no external assets. It turns a simple waiting indicator into a responsive companion that feels alive.

🆔 @hackernewsgithubprojects

📦 hash-7777/hashcortx

HashCortx: The Local AI Workspace

HashCortx bundles a full autonomous coding agent, 3D model builder, and finance tool into a single desktop app that runs entirely on your machine. No servers, no accounts, just code. The most surprising part is that it includes a real Python sandbox and a coding agent that can actually edit your files and run commands, but it is strictly locked down by a Rust permission guard. This means the AI can't sneak data out or mess with sensitive system files without your explicit permission. It handles everything locally, even running offline models, so your work stays private.

📰 https://news.ycombinator.com/item?id=49516181

🆔 @hackernewsgithubprojects

📦 amancode22/deeperseeker

DeeperSeeker: Your DeepSeek Account as an OpenAI API

DeeperSeeker is the local server that turns your existing DeepSeek web account into a fully compatible OpenAI and Anthropic API gateway. If you want to use DeepSeek’s powerful reasoning and vision capabilities with tools like Claude Desktop or your own Python scripts without paying for separate API credits, this is the bridge. It handles the complex session management, automatic rate limit recovery, and tool calling parsing for you, so you can just point your favorite apps at a simple local address.

🆔 @hackernewsgithubprojects

📦 8borane8/s2pipe

Play Switch 2 from Your Browser

Your Nintendo Switch two can actually run as a free cloud gaming server inside your own home. This project, s2pipe, grabs the video signal from the console and streams it to your web browser using standard web technologies. The coolest part? A tiny Raspberry Pico chip acts as a bridge, turning keyboard inputs into four distinct USB gamepads that the Switch recognizes. No expensive subscriptions, no cloud fees, just your local network and some clever hardware wiring. If you have a Linux PC and a spare Pico, this is a surprisingly deep dive into how gaming hardware actually talks to your computer.

🆔 @hackernewsgithubprojects

📦 ahmadpiran/restoredrill

Prove Your Backups Work: restoredrill

restoredrill solves the awkward truth that untested backups are just hopeful data. It spins up a throwaway database container, restores your latest backup file, and runs your specific checks to see if the data is actually there. The real magic is the timestamped JSON report it generates. This file proves exactly when the restore happened and how long it took, giving you concrete evidence for compliance audits instead of just saying you did it. It turns a scary manual task into a single command with a clear pass or fail result. If you care about data safety, this tool gives you the receipts you need to sleep better at night.

📰 https://news.ycombinator.com/item?id=49465291

🆔 @hackernewsgithubprojects

📦 spark-arena/recipe-registry

Recipe Registry: Run AI Models in One Command

Recipe Registry is the collection of ready-made instructions that makes running large language models on NVIDIA DGX Spark devices shockingly simple. Instead of wrestling with container images, memory limits, or complex command lines every time you want to test a new AI model, you just type a single command. This repository holds the official, tested configurations that tell the system exactly how to launch and run those models so they actually work. It turns a tedious setup process into a quick, reliable action. No more guessing your settings. Just pick a recipe and let it handle the heavy lifting for you.

🆔 @hackernewsgithubprojects

📦 vectal-labs/agentic-productivity

Is Your AI Hype Actually Productive? This Mac Tool Tells The Truth

Agentic Productivity is a clever macOS tool that measures your actual output by tracking git commits and AI coding sessions over time. It quietly scans your local machine to count how many times you use agents like Claude Code or Cursor, comparing that activity against real code commits. Every morning, it sends you three simple charts to Discord, showing you if all that prompt typing is actually turning into shipped work. The privacy focus is strong, as it only shares daily totals and never leaves your specific prompts or file paths. It is a great way to check if you are just busy or genuinely building.

🆔 @hackernewsgithubprojects

📦 zz1358m/agentic-esopt

Agentic ESOpt: Fine-Tuning Agents Without Heavy GPUs

Agentic ESOpt fine-tunes large language models without needing massive GPU clusters. Traditional training methods store heavy data to update weights, but this project uses a simpler trick. It makes tiny random changes to the model and tests if those changes improve the agent's performance. If the score goes up, the system keeps the change. If not, it reverts. This means you can improve complex AI agents using the same amount of computer memory as just running them. It is a clever way to teach software new skills without expensive hardware. If you are working with AI agents, this memory-friendly approach is worth exploring.

🆔 @hackernewsgithubprojects

📦 evermind-ai/skillcorpus

SkillCorpus: The Secret to Smarter AI Agents

AI agents are failing at tasks they should be good at because they are missing specific, step-by-step instructions. SkillCorpus solves this by building a massive, searchable library of procedural knowledge harvested from open-source code. It curates these scattered skill files, filters out bad or unsafe code, and trains a model to find the exact right instruction for any given task. The result is a huge boost in performance, with agents improving by up to fifteen points on complex benchmarks just by having access to these vetted steps. It’s not about making the model bigger, but about giving it the right context at the right moment.

🆔 @hackernewsgithubprojects