📦 cactus-compute/cactus-hybrid

Cactus Hybrid

Run small, fast AI models on your own device without worrying about incorrect answers. With cactus-hybrid, you get a smart system that actually knows when it is wrong. It embeds a specialized probe directly into the model checkpoint to calculate a precise confidence score for every single response. When the confidence is high, you get a lightning-fast local answer; if it dips too low, you can automatically hand the prompt off to a larger cloud model. It is a brilliant way to slash cloud costs and latency while keeping your application incredibly accurate. Give it a spin to build smarter, self-aware local apps.

📰 https://news.ycombinator.com/item?id=49010782

🆔 @hackernewsgithubprojects

📦 kuberwastaken/megaphone

Megaphone

Megaphone is the dictation app that finally brings incredibly fast, private speech-to-text directly to your Mac. Powered entirely on-device by Apple's latest speech analyzer, it bypasses slow, expensive cloud APIs and achieves a staggering level of accuracy that easily beats standard Whisper models. You just hold down a shortcut key, speak naturally, and let go to have the text immediately flow into whatever app you are using. Because it handles the entire transcription process locally on your Apple silicon, your voice recordings never leave your machine, making it completely private and free. It even uses local smart models to clean up stutters and self-corrections on the fly.

📰 https://news.ycombinator.com/item?id=48908349

🆔 @hackernewsgithubprojects

📦 gd03gd031/windows-gdid-changer

windows-gdid-changer

Microsoft secretly tracks your Windows installation using a hidden persistent ID called a GDID, which even survives major system updates. The windows-gdid-changer script solves this privacy issue by wiping out your local device registration sessions and forcing Microsoft's servers to think you just did a fresh install. It automatically triggers a re-registration process to issue your PC a completely brand-new identity in under ninety seconds. Just be careful because resetting this token will log you out of your Microsoft Store and Xbox apps, but it is an incredibly clever way to reclaim your digital privacy.

📰 https://news.ycombinator.com/item?id=48818707

🆔 @hackernewsgithubprojects

📦 z-lab/flashdrive

FlashDrive

NVIDIA's massive ten-billion-parameter AI models can now drive autonomous cars nearly five times faster with absolutely zero loss in accuracy. The magic is a library called flashdrive, which acts as an inference acceleration patch for NVIDIA's Alpamayo models. Instead of choking on raw sensor data, it combines smart streaming that reuses video frame memory, clever speculative reasoning to predict actions ahead of time, and tight math quantization to keep things lightweight. This drops decision latency from almost a second down to a blistering one hundred and fifty milliseconds. It is the perfect reminder that smart software optimization is just as important as bigger hardware.

🆔 @hackernewsgithubprojects

📦 luqp2/image-metahub

Image MetaHub

Image MetaHub is the desktop app that finally makes organizing local AI-generated art painless and completely private. If you have been creating artwork with Stable Diffusion or ComfyUI, you know how quickly your folders turn into a chaotic mess of random filenames. This clever tool scans your local folders to read generation metadata directly from your PNG, JPEG, and video files, lets you filter by prompt, model, or seed, and even lets you run generations directly from your local ComfyUI setup. Best of all, it keeps your library entirely offline. Keep your local AI generations organized and easily searchable today.

🆔 @hackernewsgithubprojects

📦 ovid/paad

paad

paad is the AI-assisted development safeguard that finally brings defense-in-depth to your coding workflow. Instead of blindly writing code based on bad assumptions, this clever tool introduces smart, multi-layered guardrails directly into your assistant. The absolute best part is its pushback feature. While most AI assistants happily write code for flawed or vague specifications, pushback critically reviews your requirements before a single line is written, pointing out logical gaps, security risks, or hidden design conflicts. It gives you a rigorous, honest reality check so you do not waste time building the wrong thing. You get more reliable, high-quality software with fewer expensive mistakes.

🆔 @hackernewsgithubprojects

📦 pawel-mazurkiewicz/comfyui-applesilicon-fp8

comfyui-applesilicon-fp8

Run massive FP8 and INT8 diffusion models on Apple Silicon without your machine crashing or slowing to a crawl. The comfyui-applesilicon-fp8 repository introduces a clever compatibility layer that patches Metal performance shaders at startup, allowing Macs to successfully run demanding models like FLUX, SD3.5, and Ideogram 4. Because Apple's graphics backend natively lacks eight-bit float support, this tool acts as a translator, decoding those weights on the GPU with bit-exact precision to keep your renders fast and stable. It eliminates common memory crashes, gets slow text encoders running back on the graphics chip, and boosts performance using custom attention kernels.

🆔 @hackernewsgithubprojects

📦 tunnetio/tunnet

Tunnet Mesh Networking

Most mesh networks keep their core coordination servers locked away behind proprietary code, but Tunnet changes the game by open-sourcing the entire stack. This clever Rust-based tool packages the agent, control plane, management API, and dashboard in a single repository you can fully self-host. It goes way beyond a standard VPN by combining private mesh, public tunnels, file transfers, secure identity-based SSH, and Kubernetes routing under a single access control system. It even lets you embed a mesh node directly inside your Node or Rust apps. It is the ultimate tool for taking complete control of your private network.

🆔 @hackernewsgithubprojects

📦 alex-durango/pingfusi

Pingfusi

You can now put a real human directly into your AI coding agent's development loop to stop them from generating generic designs. Pingfusi is a clever tool that acts like an on-demand review team for your autonomous coding agents. Instead of letting your agent loop forever trying to guess what looks good, it automatically publishes the mid-task draft and calls a real human reviewer. The reviewer pins exact comments directly to the layout, and the agent uses that precise feedback to iterate until it gets a green light. It is like having a design-savvy friend instantly guide your AI past the slop.

🆔 @hackernewsgithubprojects

📦 krakonjac300-pixel/podcast-shorts-factory

Podcast Shorts Factory

Podcast-shorts-factory is the automated editor that orchestrates ten cooperating AI agents to turn long YouTube podcasts into polished vertical videos for free. Rather than paying for expensive clipping subscriptions, this tool coordinates specialized digital workers to download episodes, find the most engaging highlights, apply smart face tracking, and burn in animated karaoke captions. Its standout strength is the finishing editor agent, which acts as a picky human reviewer by watching every render to catch errors like text covering faces or dead air before blocking bad clips. It is a brilliant way to keep a video channel running completely hands-free.

🆔 @hackernewsgithubprojects

📦 risa-labs-inc/bossconsole

bossconsole

Bossconsole is the native operator's console that finally gives your AI agents a unified, multi-threaded workspace to get real work done. Instead of letting an agent run wild on your machine, this tool wraps around your terminal, editor, and a built-in browser to give you absolute control. You can run any coding agent you want, and it instantly gains a massive suite of tools to safely read files, run commands, and automate web tasks. The coolest part is that it is built on the JVM with true multithreading instead of slow Electron, and you can hot-reload custom plugins on the fly. It is the ultimate dashboard for anyone who wants...

🆔 @hackernewsgithubprojects

📦 pnnbao97/vieneu-tts

VieNeu-TTS

vieneu-tts is the text-to-speech toolkit that brings human-like voice synthesis and instant voice cloning to Vietnamese applications entirely on your local device. It turns written text into natural speech with a high-fidelity output of up to forty-eight kilohertz, and it even lets you clone any voice using just a five-second audio clip. What makes this project really special is that it can run directly on a standard computer processor without needing a heavy graphics card. This means you can easily generate natural-sounding voiceovers, podcasts, and realistic dialogues offline. Check it out to build highly responsive, localized voice applications.

🆔 @hackernewsgithubprojects

📦 quintinshaw/openasr

OpenASR

Imagine transcribing audio and meeting recordings entirely on your computer without sending a single byte to the cloud. OpenASR is a local-first speech-to-text platform that lets you run twenty-eight different voice recognition models completely offline on your own CPU or graphics card. It transcribes audio files, generates subtitles with speaker labels, and even captures system audio directly from your screen. Because it runs entirely locally with zero telemetry, your private conversations and recordings never leave your machine. It is the perfect tool for private, high-speed transcription and translation right from your desktop or terminal.

🆔 @hackernewsgithubprojects

📦 tommyjepsen/awesome-ux-skills

Awesome UX Skills

You can now automatically inject professional user experience frameworks directly into Claude Code. A new open-source repository called awesome-ux-skills acts as an automated design companion by loading specialized prompt templates straight into your terminal workflow. When you ask Claude design questions, it automatically detects the right template to evaluate your UI against accessibility guidelines, run cognitive load checks, or structure user persona research. By translating complex design concepts into automated, situational prompts, this project bridges the gap between raw code and polished product design, helping you build more intuitive interfaces without having to manually feed best practices to your AI assistant.

📰 https://news.ycombinator.com/item?id=48766415

🆔 @hackernewsgithubprojects

📦 duolahypercho/codex-router

Codex Router

codex-router is the credential-isolating local gateway that finally lets you run premium external models like Claude, DeepSeek, and Grok directly inside your favorite AI desktop editors. It acts as a clever local translation proxy, converting your app's standard requests into the native language of whatever provider you choose, while keeping your API keys and login sessions completely isolated on your own machine. My favorite feature is that it even lets you use your editor without an OpenAI login by publishing external models under native aliases so the app still recognizes them. It is the perfect weekend upgrade to make your development environment truly model-agnostic.

🆔 @hackernewsgithubprojects

📦 pyang5166/gbro-collage-broll

gbro-collage-broll

gbro-collage-broll is the clever agent tool that turns your five-second voiceover scripts into gorgeous, paper-collage stop-motion video clips. Instead of burning real money on chaotic AI video generations, this project introduces a smart three-step approval gate. First, it cooks up a solid visual metaphor. Next, it generates a beautiful halftone paper-collage static image. Finally, it uses Gemini Omni Flash to assemble everything into a slick animation where elements slide and snap into place. It gives you complete creative control at every step, saving your budget while producing stunning, high-end editorial visuals for your videos.

🆔 @hackernewsgithubprojects

📦 slvdev/esp32-ai

ESP32-AI

Run a twenty-nine million parameter language model directly on an eight-dollar microcontroller with absolutely no internet connection. The esp32-ai project makes this possible on an ESP32-S3 chip by using a clever memory design called Per-Layer Embeddings. Instead of trying to squeeze the entire model into the tiny half-megabyte of fast RAM, it stores ninety percent of the model in slow, cheap flash memory. It only pulls the tiny handful of data rows needed for each word, allowing the chip to write coherent stories to a small screen at nearly ten words per second. It is a brilliant way to make tiny, cheap hardware punch way above its weight class.

🆔 @hackernewsgithubprojects

📦 escoffier-labs/brigade

Brigade

Artificial intelligence agents can execute complex loops and generate code, but keeping track of what they actually accomplished is notoriously difficult. Brigade solves this by acting as a lightweight, local control plane that creates verifiable file receipts for every single action your agents take. Running completely on your own machine without background daemons or platform lock-in, it lets different AI tools and frameworks easily share the Model Context Protocol, custom tools, and memory. Because it records everything with tamper-resistant receipts, you can easily prove what your agents did and use real program exit codes to systematically improve their performance.

🆔 @hackernewsgithubprojects

📦 nvidia-nemo/emerging-optimizers

Emerging Optimizers

Nvidia’s emerging-optimizers repository implements advanced training algorithms that actually speed up how fast massive AI models learn. Traditional training rely on simple adjustments, but this project packages sophisticated alternatives like Muon and Soap, which use advanced matrix operations and geometry to find the best settings much faster. In fact, these exact techniques were used to train the massive Kimi K2 language model. By integrating these tools directly into standard graphics processing unit training systems, developers can achieve quicker convergence and better final model quality. It is a powerful toolbox for anyone looking to push the boundaries of neural network efficiency.

🆔 @hackernewsgithubprojects

📦 maxwell2732/pai-econ-claude

pai econ claude

Empirical economists often struggle to bridge the gap between real-world data and formal mathematical modeling. The open-source project pai-econ-claude solves this by acting as a highly structured theoretical scaffolding for empirical researchers. Instead of building models from scratch, it matches real-world economic puzzles with classic theoretical frameworks like search models, moral hazard, or rational inattention. The tool walks users through rigorous checks, auditing assumptions, outlining proof sketches, and organizing findings into draft manuscripts. It is a brilliant way to make sure your empirical insights are grounded in solid, peer-reviewed economic theory before you write your paper.

🆔 @hackernewsgithubprojects

📦 shanghai-academy-of-ai-for-science/mkb

mkb

Imagine a single AI model that can design RNA sequences, forecast global weather, and segment medical images with expert-level precision. That is exactly what mkb does. Built on an eight-billion parameter language backbone, this system pairs specialized scientific encoders and decoders with a shared core. Instead of swapping out different models for different tasks, it routes biology, chemistry, and physics data through a single natural-language interface. It actually outperforms massive one-trillion parameter models on biological tasks and rivals physics-based forecasting systems. It is a brilliant, lightweight way to solve diverse scientific problems on a single graphics card.

🆔 @hackernewsgithubprojects