📦 jarrettyu/alphamemo

AlphaMemo: The AI That Remembers How to Find Money

AlphaMemo is a self-evolving framework that teaches artificial intelligence to hunt for financial formulas without forgetting its mistakes. Traditional search methods start from scratch every time, but this system keeps a structured memory of its entire search process. It remembers what failed, what worked, and how to adjust its confidence levels as it goes. By acting like a persistent analyst rather than a blind guesser, it finds profitable patterns in stock data far more effectively. It turns a chaotic search into a disciplined, learning loop, which is genuinely impressive for anyone interested in how machines can improve their own problem-solving skills.

🆔 @hackernewsgithubprojects

📦 geroldmeisinger/comfyui-outputlists-combiner

ComfyUI Outputlists Combiner: Batch Process AI Images

ComfyUI Outputlists Combiner is the workflow node that finally lets you batch process hundreds of AI prompts in a single run. Instead of clicking generate over and over, this tool takes simple lists from spreadsheets, JSON files, or even just a block of text, and combines them into every possible variation. It handles the heavy lifting of feeding different LoRA weights, samplers, or parameters into your workflow automatically. You can even create complex grids of images to compare results side by side. If you are tired of running one image at a time, this is the bridge between your data and your AI engine.

🆔 @hackernewsgithubprojects

📦 hoaxisr/awg-manager

The Router Control Panel You Didn't Know You Needed

Turn your Keenetic router into a full-fledged network command center with A W G Manager. Instead of juggling scattered config files, this tool gives you a clean web interface to manage your Amnezia W G tunnels directly from your browser. It handles the complex routing, server settings, and system diagnostics that usually require deep command line knowledge, making it surprisingly simple to keep your private network running smoothly. If you are looking for a way to take control of your home network without the usual headache, this is the tool to watch.

🆔 @hackernewsgithubprojects

📦 mukul975/cve-mcp-server

CVE MCP Server: Instant Security Intelligence

CVE MCP Server is the security plugin that finally turns your AI assistant into a full-time threat analyst. Instead of tab-hopping through twenty different websites to check if a vulnerability is actually dangerous, you just ask your AI one question. It instantly fans out to trusted sources like the National Vulnerability Database and CISA, checks for active exploits, and calculates a single, clear risk score. The real magic is that it doesn't just give you raw data; it tells you exactly what to patch first, saving hours of tedious research. It’s like having a senior security expert who never gets tired and never forgets to check the latest threat intel.

🆔 @hackernewsgithubprojects

📦 kuberwastaken/hp-laser-1008a-macos

HP Laser 1008a on macOS

HP Laser 1008a on macOS is the driver that finally makes an aging office laser printer behave on your Apple Silicon Mac. You know the frustration? You buy a cheap, reliable laser, only to discover HP never built a driver for modern Macs, so it just sits there refusing to print. This project fixes that by reverse engineering the printer’s specific quirks, specifically how it misreads page sizes, and patches a known open-source engine to speak its language. It bypasses the usual clunky workarounds like virtual machines or heavy Python scripts, using lightweight native code that just works.

📰 https://news.ycombinator.com/item?id=49508251

🆔 @hackernewsgithubprojects

📦 pbanddev/comfyui-node-organizer

One Click to Tidy ComfyUI

ComfyUI Node Organizer snaps hundreds of scattered nodes into perfect rows with a single click. When building complex image workflows, the canvas often turns into a chaotic tangle of boxes and wires that makes tracking data flow nearly impossible. This extension instantly tidies that mess by automatically aligning your nodes into clean, logical layouts. You can even tag specific groups with simple labels to force them into exact grids or columns. It works entirely within your browser, so your messy creation process stays perfectly reversible with standard undo shortcuts. If you build with ComfyUI, this tool is the fastest way to keep your visual logic clear and your headspace free.

🆔 @hackernewsgithubprojects

📦 futuregene/future-os

FutureOS: The AI Agent That Works Everywhere

FutureOS gives you a single AI agent that operates seamlessly across your terminal, desktop, mobile devices, and chat apps. The standout feature is its trust-first design, where every action the agent takes is gated by your explicit approval, ensuring nothing runs silently in the background. Built on a high-performance Rust core, it connects all these different interfaces to one shared memory and session history. This means you can start a complex task on your computer and check the results on your phone later, without losing context. It is open source and designed for users who want powerful automation without sacrificing control or security over their personal data.

📰 https://news.ycombinator.com/item?id=49417312

🆔 @hackernewsgithubprojects

📦 rookedsysc/kanvibe

KanVibe: The Kanban Board for AI Coding Agents

KanVibe is the keyboard-first desktop app that turns your scattered AI coding terminals into a single, organized Kanban board. The coolest part? It watches your AI agents work. As Claude Code, Codex, or Gemini sessions run in your background terminals, KanVibe automatically tracks their progress and moves your tasks from TODO to DONE without you touching a single card. When a task is finished, it even cleans up the temporary git branch for you. It is like having a project manager for your AI assistants, keeping your workflow tidy while you focus on coding. Check it out if you use AI coding tools and want to stay organized.

📰 https://news.ycombinator.com/item?id=49509626

🆔 @hackernewsgithubprojects

📦 mitkox/stolen-thoughts

Stolen Thoughts: The Hidden AI Vulnerability

A research project called Stolen Thoughts proves that the hidden thinking inside top-tier AI models is not actually safe. The team built this tool to show how encrypted reasoning traces can be intercepted and decoded using a cheaper, weaker AI model as a decryption oracle. This repository offers a working demo and a ten-point validation suite to verify these security gaps. It clearly explains how proprietary AI providers can fail to protect their internal logic. Understanding this risk is vital for anyone building systems that rely on these advanced AI services.

🆔 @hackernewsgithubprojects

📦 sudoingx/qwen38-mtp

Qwen3.8 MTP: The Free Speed Boost

Qwen3.8 MTP is the performance recipe that finally makes your existing model run significantly faster without buying new hardware. The project reveals that a specific command line flag unlocks a massive speed increase by activating a feature already buried inside your current model file. Essentially, the model was built to predict multiple words at once, but the software ignored it until now. By simply adding one flag, you turn on this internal draft mechanism, resulting in a thirty percent or better speed boost on standard consumer graphics cards. This repository documents the exact settings and benchmarks from a global community, helping you find the perfect configuration for your specific setup.

🆔 @hackernewsgithubprojects

📦 zapier/wade-skills

Wade Foster's AI War Council

Zapier CEO Wade Foster has open-sourced the exact prompt files he uses to let an AI agent act as his executive staff. The repository, called Wade Skills, contains three specific workflows that turn your AI assistant into a high-level advisor. The most surprising feature is the War Council. It forces your AI to role-play a panel of ruthless, opinionated experts, like a skeptical CFO or a contrarian board member, to aggressively stress-test your business ideas before you commit. The other tools handle mundane but critical tasks, such as automatically extracting action items from your meeting notes and generating a forced-ranked weekly agenda for your leadership team.

🆔 @hackernewsgithubprojects

📦 blazux/qwen3.8-flash-dgx

Running a 176B Model on One Desktop Chip

A 176 billion parameter language model now fits on a single desktop computer, and it runs fast enough to be useful. The project, Qwen 3.8 Flash DGX, solves the problem of a massive lookup table that simply did not fit in the device’s unified memory alongside the active computation. By streaming that table from fast storage on demand, the system frees up enough memory for the model to actually work, turning a hardware limitation into a manageable engineering task. This repository acts as a complete recipe, including the necessary software patches to fix specific bugs that made the model unstable on this hardware.

🆔 @hackernewsgithubprojects

📦 pottergravity/microsoft-access-craft

Microsoft Access Craft

Organize your daily work with a Windows desktop tool that handles scheduling, note taking, and goal tracking all in one light place. Microsoft Access Craft focuses on staying out of your way, offering quick setup and low resource usage for day to day productivity. It supports offline mode so you can keep working without internet, then sync automatically later. You can also connect popular apps for email and cloud storage, plus use pre built templates for meetings and projects. The goal is simple, fast organization that fits into your existing routine without heavy software overhead. Download it, extract the files, and start managing your tasks more efficiently today.

🆔 @hackernewsgithubprojects

📦 patchzyy/wiicompiled

Mario Kart Wii Without an Emulator

Wiicompiled runs Mario Kart Wii on your PC by compiling the entire game into native code, meaning zero emulation, zero interpreters, and zero lag from interpreting instructions. The most surprising part is that the physics are one hundred percent identical to the original console, proven by ghost replays syncing perfectly with real hardware. It is not just a port; it is a complete static recompile that lets you play at high resolutions with unlocked framerates and frame interpolation for smoother visuals. You need your own clean PAL disc image, and the setup compiles everything locally on your machine.

🆔 @hackernewsgithubprojects

📦 alexgreensh/token-optimizer

How to Stop Your AI from Forgetting

Your AI assistant is likely wasting a massive chunk of your budget on invisible data that vanishes every time the conversation gets long. Token Optimizer acts as a background guardian that spots these ghost tokens, fixes bloated settings, and ensures your work survives even when the system tries to compress memory. While other tools only trim command outputs, this one digs deep into the structural waste that quietly eats your context quality. It automatically checkpoints your progress so nothing gets lost, then shows you a live dashboard tracking exactly where every dollar went.

🆔 @hackernewsgithubprojects

📦 nadimtuhin/claude-token-optimizer

Claude Token Optimizer: Fix Your AI Context Waste

Cut your Claude token bill by ninety percent with Claude Token Optimizer. This tool stops your AI assistant from wasting precious context on outdated documentation. It restructures your project files so only the four essential markdown documents load at startup, keeping everything else in a zero-cost archive until you actually need it. The tool automatically detects your tech stack, whether it's Python, Go, or JavaScript, and sets up a smart loading system in about thirty seconds. You get a massive boost in available memory for actual coding logic.

🆔 @hackernewsgithubprojects

📦 mibayy/token-savior

Token Savior: The AI Efficiency Trick

Token Savior is an MCP server that helps AI coding agents work significantly faster and more accurately without using up your expensive token budget. Instead of making the AI read entire files just to find one specific function, it builds a smart map of your codebase so the model can navigate by pointer rather than guessing. It also remembers key decisions and conventions from previous sessions, so it doesn't forget context when a chat restarts. In one benchmark, this tool helped an AI model score nearly 98 percent on a coding test while using eighty percent fewer tokens and finishing tasks in a fraction of the time.

🆔 @hackernewsgithubprojects

📦 ooples/token-optimizer-mcp

Token Optimizer MCP: Cutting AI Costs by 90%

Token Optimizer MCP cuts your AI coding agent costs by up to ninety percent. It works by intercepting the massive amounts of text and code that these agents send to large language models. Instead of paying to process the same files repeatedly, it uses smart caching and compression to remove the fat. Think of it as a personal manager that only hands the AI exactly what it needs, not the entire library. This is especially useful if you use multiple different AI tools, as it creates a shared knowledge graph across sixteen clients. Your agents learn from each other's mistakes and optimizations, so you stop paying for redundant work.

🆔 @hackernewsgithubprojects

📦 n1tr00-10/tiktok-signature

TikTok Signature Generator

Generate the hidden security codes that TikTok uses to verify every web request. This open source project recreates the exact algorithm TikTok uses to create what they call X-Gnarly, X-Bogus, and X-Dynosaur signatures. These are essentially digital fingerprints attached to your data. Without them, the TikTok website won't even talk to its servers. The project is a pure Python tool that mimics the encryption TikTok uses. It uses specific encryption methods like ChaCha20 and RC4 to build these codes exactly the way the official system does. It is interesting because it shows the engineering behind a system we normally just see working in the background.

📰 https://news.ycombinator.com/item?id=49258849

🆔 @hackernewsgithubprojects

📦 tonyd2wild/deepseek-v4-flash-vision-exp-dspark-1m-nvfp4-kv-2x-dgx-spark

DeepSeek V4 Flash Vision: 1M Context on Two Sparks

DeepSeek V4 Flash Vision is the two-node recipe that finally brings one-million-token image understanding to a modest DGX Spark pair. The single most surprising thing it does is keep the model fast and accurate while reading entire books or massive codebases, all while accepting direct image input, without melting your hardware budget. It works by squeezing the memory footprint of its context cache using a special low-precision format that fits a million tokens into just under twenty gigabytes of shared GPU memory. The recipe handles the tricky wiring of two machines acting as one giant brain, fixing silent bugs that usually cause half your speed to vanish without a warning.

🆔 @hackernewsgithubprojects