📦 seesee75-commits/comfyui-minimaxh3-director

MiniMax H3 Director: The ComfyUI Timeline Editor

MiniMax H3 Director turns your ComfyUI workflow into a visual timeline editor, letting you build video clips by dragging media onto tracks instead of writing long text prompts. It works by compiling your visual edits into a storyboard that the AI follows shot by shot. You can anchor scenes with specific start and end images, add background music, and even preview the actual video rendering in progress. The tool also handles retakes, where it regenerates only the parts of a clip that need fixing while keeping the rest intact. This makes creating consistent, narrated video scenes much easier and more intuitive than typing instructions one by one.

🆔 @hackernewsgithubprojects

📦 mvparakhin/ml-tidbits

ML Tidbits: Clean Machine Learning Recipes

ML Tidbits is the practical collection that turns complex machine learning research into clean, readable code you can actually use. It solves the problem of academic papers being dense and hard to replicate by providing simple Python and C++ examples for real tools. The standout feature is its focus on deterministic models that let you generate realistic data by swapping parts instead of guessing randomly. This approach makes complex math predictable and easy to understand for anyone building smart software. It is a brilliant resource for developers who want to see exactly how these ideas work in practice without getting lost in theory.

🆔 @hackernewsgithubprojects

📦 markrussinovich/doompaint

DoomPaint: Play Doom in MS Paint

DoomPaint runs the actual DOOM shareware inside Microsoft Paint. The Python script renders frames headlessly, then copies each image to your clipboard and pastes it into Paint as a real document edit. This means every frame is an undo step. You can press Control and Z to rewind time and un-die. It works on Windows and plays sound effects while the game runs.

📰 https://news.ycombinator.com/item?id=49155026

🆔 @hackernewsgithubprojects

📦 mrinalwadhwa/unramble

Unramble: Dictate Incognito

Unramble runs a speech-to-text engine and text cleaner entirely on your Mac, so your voice never leaves your device. Press a hotkey, speak naturally, and polished text appears in any app without sending audio to the cloud. It uses small local models to remove filler words and fix punctuation while keeping your data private. You can switch to a cloud mode for better language support, but the default setup works completely offline and costs nothing. It is a simple way to dictate clearly while keeping your conversations secret.

📰 https://news.ycombinator.com/item?id=47411467

🆔 @hackernewsgithubprojects

📦 scenemaai/scenema-audio

Scenema Audio: Emotional Voice Cloning

Scenema Audio is the expressive voice cloning tool that finally turns text into acting performances instead of robotic reading. Most text-to-speech software just pronounces words correctly, but that is useless when you need a voice that sounds angry, exhausted, or terrified. This project changes the game by letting you describe exactly how a character should feel, and then cloning their voice from just a ten second audio clip to match that performance. You provide a short reference recording for the identity, and then write stage directions to control the pacing and breath. The result is realistic speech with natural emotional arcs, even if the original speaker never made those sounds.

🆔 @hackernewsgithubprojects

📦 revylai/greenlight

Check App Store Rules Before You Submit

greenlight is an open-source tool that checks your mobile app code and build files against Apple App Store and Google Play rules before you hit submit. It scans your project offline, looking for things like missing privacy explanations, hard-coded secrets, or missing account deletion buttons that could get your app rejected. The most interesting part is its verify mode, which doesn't just check if a delete button exists in the code, but actually runs it on a cloud device to prove the account gets deleted. It saves developers from painful rejection cycles by catching these issues early, all without uploading your private data anywhere.

📰 https://news.ycombinator.com/item?id=49187437

🆔 @hackernewsgithubprojects

📦 amazingang/old-coder

Old Coder: Let AI Prove It Works

Old Coder flips the script on coding agents. Instead of you reading thousands of lines of generated code, the AI must make its work pass a brutal gauntlet of checks. It writes a test plan you approve, then runs its code through type checks, mutation tests that plant fake bugs, and property tests that blast it with random inputs. You only read two documents: the approved plan and the final evidence report with real proof numbers. This approach builds trust through constraints, not inspection, letting you rely on the results without getting lost in the code.

🆔 @hackernewsgithubprojects

📦 crabbuild/compass

Compass: A Knowledge Graph for Your Codebase

Compass lets you turn a messy codebase into a searchable graph, letting you instantly find dependencies, trace impact, and understand architecture without needing vector databases or Python. Built in Rust for speed, it runs locally, keeping your data private and queries fast. The real magic is its versioned history feature, which captures exact snapshots of your code at every Git commit, allowing you to compare how the project evolved and spot breaking changes before they hit production. It is basically a time machine for your software structure, giving you precise, traceable answers instead of guessing.

📰 https://news.ycombinator.com/item?id=49178759

🆔 @hackernewsgithubprojects

📦 jarynclouatre/bindery

Bindery: Drop Files, Get Converted E-Books

Drop e-books and comics into a folder and watch them automatically transform into perfectly formatted files for your Kindle or Kobo, all handled by a single Docker app that runs quietly in the background. It combines two powerful conversion tools into one self-hosted service, so you never have to mess with complex command lines or manual settings. You simply put your digital reading material into a designated folder, and Bindery takes care of the rest, optimizing page layouts and shrinking file sizes without losing quality. A clean web interface lets you track progress, upload files directly from your phone, and tweak preferences without touching any code.

🆔 @hackernewsgithubprojects

📦 zadid6pretam/gotabpfn

Gotabpfn: Smarter AI for Complex Data Tables

Gotabpfn is the smart tool that finally helps artificial intelligence understand messy, complex data tables by reordering their columns for better clarity. Most AI models struggle when they face huge spreadsheets with hundreds of columns because the random order of information confuses them. This project fixes that by using a graph method to arrange your data features into the most logical sequence before the model even starts learning. It also shrinks that data into a compact format that fits easily into powerful foundation models, making prediction tasks much faster and more accurate.

🆔 @hackernewsgithubprojects

📦 primeintellect-ai/prime-agent

Prime Agent: The AI That Learns From Its Own Mistakes

Prime Agent is an open-source coding assistant that actually improves itself while you work. Instead of just answering your questions, it runs your code in a persistent Python environment, watching what happens and storing those lessons as durable memories. You can tell it to refine its own approach, and it applies small, evidence-backed updates to its working knowledge without wiping the slate clean. This means it remembers past mistakes and successful patterns, so it gets sharper with every session. It can also run complex, long-running tasks in the background, sending messages to its own helper agents to coordinate the work while you sleep.

🆔 @hackernewsgithubprojects

📦 xiaomirobotics/xiaomi-robotics-1

Xiaomi-Robotics-1 Explained

Xiaomi-Robotics-1 learned over one hundred thousand hours of robot movement just to master new tasks with almost no help. This repository contains the code for a foundation model that acts like a general-purpose brain for physical robots. Instead of teaching a machine to fold laundry, open drawers, or pack boxes from scratch, engineers train it on massive amounts of diverse, real-world video data first. The result is a system that can look at a messy kitchen and figure out how to clean it up, or adapt to a completely new room and object arrangement without being reprogrammed.

🆔 @hackernewsgithubprojects

📦 autom8ai/open-higgsfield-ai

Open Higgsfield AI: Free Cinematic Studio

Open Higgsfield AI packs over two hundred creative models into one free desktop app. You get full control over cinematic camera angles, generate video from images, and animate portraits with lip sync. Unlike closed platforms, this open-source tool lets you self-host everything so your data never leaves your machine. It works right out of the box on Mac and Windows. Download the app and start building professional visuals without paying a dime or worrying about vendor lock-in.

🆔 @hackernewsgithubprojects

📦 tr1dae/comfyui-minimaxh3_latentupscaler

MiniMax H3 Latent Upscaler

Upscale video resolution between MiniMax H3 sampling passes while keeping the audio clean. Most upscaling tools break audio when resizing, but this custom ComfyUI node handles the tricky NestedTensor data properly. It doubles the video resolution and then carefully remakes the audio noise based on your settings. You can lock the first pass audio completely, just lightly polish it, or let the second sampler fix it entirely. It even upscales reference images to match the new size so everything stays in sync. Just wire two samplers together with this node in between to get a sharp, coherent result.

🆔 @hackernewsgithubprojects

📦 herenn/infralens

Infralens Kernel Network Map

InfraLens turns your messy infrastructure into a clear live map by tracing network traffic directly from the operating system kernel. You get a real-time visualization of exactly which services talk to each other without touching any code or adding heavy helper containers. It works on Linux servers and inside Kubernetes clusters, identifying everything from database connections to DNS queries with almost no performance impact. The tool even uses artificial intelligence to automatically explain what each service does just by clicking on it. It is the fastest way to understand your architecture without rewriting your code or waiting days for reports.

🆔 @hackernewsgithubprojects

📦 faraz-portfolio/demo-2026-r3f-inspector

demo-2026-r3f-inspector

demo-2026-r3f-inspector is the interactive playground that turns the tricky parts of ThreeJS WebGPU development into a visual game. It lets you tweak geometry, lighting, and materials in real time using a familiar control panel instead of rewriting code every five seconds. The highlight is the built-in inspector that hooks directly into your React components, letting you adjust sliders and colors while the scene updates instantly. It feels like magic because you get immediate feedback on how those changes affect the rendering pipeline without restarting anything. This makes experimenting with new graphics features way less frustrating and way more fun.

🆔 @hackernewsgithubprojects

📦 cloudflare/cloudflare-os

Cloudflare OS

Cloudflare OS turns your company’s messy data into a secure, AI-powered workspace where every app runs in its own private sandbox, keeping your information safe while letting AI agents build tools on the fly. Instead of trusting a central server with your secrets, the system creates individual instances of software just for you, so a slide deck or a game never leaks data to the outside world. The coolest part is the Gatekeeper security layer, which lets AI agents plan their work in the background and then asks you to approve big moves like sending emails or updating databases all at once, rather than freezing up after every single step.

📰 https://news.ycombinator.com/item?id=49182551

🆔 @hackernewsgithubprojects

📦 optim-agent/optim-plans

Optim Plans

Optim-plans is the human-in-the-loop planning plugin that finally stops coding assistants from rushing into the wrong implementation. Instead of letting an AI agent blindly write code, this tool acts as a dual-ring safety net. First, it forces a planning-and-refining loop to turn vague ideas into clear, human-reviewed markdown plans while asking clarifying questions. Once approved, the executing-and-validating ring safely tests, audits, and verifies the changes inside a secure git workspace before anything is integrated. It is an incredibly smart way to give systems like Claude and Codex the explicit execution gates and safety controls they need to build software reliably.

🆔 @hackernewsgithubprojects

📦 dingduff/dingduff-public

DingDuff Public: Free Legal Research for Lawyers

DingDuff Public gives lawyers instant access to millions of court opinions and statutes right inside the Claude chat they already use. Built by attorneys for attorneys, this project acts as a connector that lets you ask legal questions in plain English and receive answers backed by real, verifiable citations from federal and state sources. Instead of memorizing complex search commands, you simply describe what you need, and the tool pulls the exact texts of case law, rules, and PACER filings to help you build accurate research memos.

🆔 @hackernewsgithubprojects

📦 attemorysystem/attemory

Attemory: AI Memory That Actually Remembers

Attemory is a retrieval engine that lets AI agents search their own long-term memory using attention, not just vector similarity. It indexes documents and code into reusable key-value state, then runs a local model to reason over that data and pull up the exact evidence needed. This approach cuts coding-agent token usage by nearly half while keeping search quality high. Instead of relying on compressed embeddings, Attemory lets the model read raw text directly, making it much better at finding specific details in huge context windows. It works locally on Linux and macOS, supporting CPU and GPU backends.

🆔 @hackernewsgithubprojects