📦 wangeai/k3flight

Running a 2.8 Trillion Parameter AI on CPU

K3 Flight lets you run a massive two-point-eight-trillion-parameter AI model locally on your CPU, without needing a GPU. The model file is nearly a terabyte, but this tool uses a smart flight plan to load only the tiny piece of the brain needed for each word, keeping active memory around fifty-five gigabytes. It turns a hardware impossibility into a workable local server, proving that even the largest models can run on your desktop if you manage the data flow correctly.

🆔 @hackernewsgithubprojects

📦 fsmargoo/dsh-at-file

Point Your AI at Any File

Stop typing out long file paths for your AI. This plugin lets you type an ampersand in your prompt to search your entire project folder with one click. It acts like a smart autocomplete that finds any document, code file, or asset you need. The coolest part is that it doesn't stuff the file content into your chat. It just tags the file's location, letting the AI agent read it only when it actually needs to. This keeps your prompts clean and your context window tidy. It’s a simple, clever way to keep your AI grounded in your real work without the mess of copy-paste.

🆔 @hackernewsgithubprojects

📦 vivekhaldar/seed

The AI Agent That Builds Itself

A tiny AI agent starts life with just one tool: the ability to run shell commands. That is the entire starting kit for Seed. It does not come with pre-made skills, memory, or codebases. Instead, it grows its own brain. It writes notes, creates tools, and saves them in a local folder that acts as its persistent mind. Because it starts so small, you can see exactly how it thinks and changes. It is not a rigid framework; it is a living system that rewrites its own instructions after every chat.

📰 https://news.ycombinator.com/item?id=49384113

🆔 @hackernewsgithubprojects

📦 drstranded/co-rl

Co-RL: How AI Models Teach Each Other To Reason Without Labels

Train language models to reason by letting them grade each other's work without using any human labels. Co-RL, or Collaborative Reinforcement Learning, solves the problem of AI models getting stuck in bad loops when they only learn from their own feedback. Instead of training alone, this method uses a diverse group of different AI models to act as a team. Each model learns by looking at the answers of its peers, which breaks the cycle of repeating mistakes. This simple shift allows the system to figure out complex logic on its own. The result is a smarter model that improves without needing expensive, hand-made data.

🆔 @hackernewsgithubprojects

📦 embeddings-benchmark/embedders-dilemma

The Embedder's Dilemma: Why Cheap Beats Smart

This open-source benchmark puts ten large language models and twenty-six embedding models on the same stage for thirty-seven text tasks, but it stops there. It doesn't just measure accuracy; it tracks every dollar spent. The surprising result? While the biggest language models often win on complex reasoning, they are orders of magnitude more expensive than simple embedding pipelines. In fact, some tiny embedding models deliver nearly identical performance for mere cents. The repository provides the raw code, data, and scripts to reproduce every single chart from the paper without needing a GPU or an API key.

🆔 @hackernewsgithubprojects

📦 leonxlnx/unlazy

unlazy

Force your coding assistant to actually finish the job instead of cutting corners. AI agents have a nasty habit of stopping at eighty percent and confidently claiming victory while the code is still broken. Unlazy fixes this by making completion a strict, mechanical requirement. Before any work starts, the agent writes specific acceptance criteria into a file. Then, it has to run actual commands to prove each box is checked. It cannot just say it is done; it has to show the evidence. If the checks fail, the agent is blocked from ending the task until the work is truly complete.

🆔 @hackernewsgithubprojects

📦 thisislt/dapm

DAPM: The Drone Depth Breakthrough

A single photo from a tilted drone now reveals exact depth, solving a problem that has plagued aerial robotics. DAPM is the first model built specifically for drones to estimate depth and camera pose simultaneously, no matter how wildly the angle shifts. Existing tools fall apart when the camera tilts or zooms, but DAPM uses clever geometric tricks to keep the math steady. It even includes a new dataset called UAPD to test these extreme scenarios. If your drone navigation relies on accurate 3D maps from a single lens, this is the framework that finally handles the chaos.

🆔 @hackernewsgithubprojects

📦 lumosai8/comfyui-orbitsheets

ComfyUI OrbitSheets: The Fix for Inconsistent AI Character Views

Stop fighting inconsistent character designs by using ComfyUI OrbitSheets. This tool turns a single image into a perfect character reference sheet by rendering a continuous camera move around your subject. The key is that it doesn't just guess new angles; it captures one unbroken shot and then uses a vision model to pick the best frames. This ensures every view of your character matches perfectly, solving the problem where AI usually changes details like clothing or features between different shots. It is a clever way to get stable, professional-looking turnaround sheets without manually fixing errors in every single frame.

🆔 @hackernewsgithubprojects

📦 longin-yu/padoc

PaDoc: The Document Parser That Thinks in Trees

PaDoc is the document parsing tool that lets a single AI model read a whole page at once instead of crawling through it line by line. It works by having the model identify the layout first, then instantly branching off to read the text in each section simultaneously. This means you get the full content of a complex document much faster, without needing a second model or a complicated setup. It’s a clever way to make AI understand where things are on a page and what they say, all in one smooth motion.

🆔 @hackernewsgithubprojects

📦 zifacorp/zifamem

ZifaMem: Emotional Memory for AI Companions

ZifaMem is the memory framework that lets AI companions actually feel like they remember you. While most systems just store raw transcripts, this tool extracts emotional signals so your agent understands trust, conflict, and shared history. It doesn't just save facts; it lets memories fade or strengthen over time, creating a living relationship instead of a static database. This means your AI stops repeating old context and starts evolving with you, remembering what truly mattered in your dynamic. If you are building an agent that needs to feel genuinely personal, this is the layer you have been missing.

🆔 @hackernewsgithubprojects

📦 zcag/tela

Tela: The Wiki Your AI Agents Actually Use

Tela is the self-hostable team wiki built specifically for AI agents to be first-class teammates. While most tools keep human docs and AI data in separate silos, Tela bridges that gap by embedding the Model Context Protocol server directly into the core backend. This means your coding assistants can securely read, search, and write directly to the same markdown pages your team uses, without needing brittle external plugins. It even includes Atlas, an engine that automatically drafts wiki pages from your code so your documentation never rots.

🆔 @hackernewsgithubprojects

📦 sangrokjung/claude-forge

The AI Code Reviewer That Actually Catches Bugs

A second AI agent that has never seen the first one's reasoning is now catching real bugs in the codebase. This is the core of Claude Forge, a massive toolkit that turns a standard AI coding assistant into a coordinated team. It gives your AI sixteen specialized agents, thirty-five quick commands, and twenty-one safety hooks to handle complex workflows automatically. The latest update focuses on reliability. It uses an adversarial review loop where an independent agent actively tries to break the code to find hidden flaws before you do. This project solves the problem of trusting an AI to check its own work by enforcing a strict separation of duties.

🆔 @hackernewsgithubprojects

📦 bmb12d3/minimax-h3-prompt-composer

The Single File Prompt Composer for MiniMax Video

A single HTML file lets you build complex video prompts for MiniMax without leaving your browser. The H3 Prompt Composer is a free, offline tool that turns visual filmmaking choices into structured text for ComfyUI. Instead of writing raw code or guessing syntax, you use guided controls to manage camera paths, character references, and scene timing. It handles tricky workflows like inserting new subjects into existing footage or editing specific details while keeping the rest of the video unchanged. The app checks your prompt for errors before you generate anything, ensuring your setup is ready. It runs locally with no servers, so your creative data stays private.

🆔 @hackernewsgithubprojects

📦 jolars/panache

Panache: The Smart Formatter for Markdown

Panache is the language server that finally gives your Markdown, Quarto, and R Markdown documents a serious, professional polish. While most editors just show you text, Panache parses your files with the same deep precision used by major publishing tools, catching subtle structural errors and formatting issues as you type. It acts like a quiet editor in your corner, ensuring your headings, tables, and citations are perfectly aligned before you even hit save. This means your technical writing stays clean, consistent, and ready for publication without the usual manual cleanup.

📰 https://news.ycombinator.com/item?id=49398113

🆔 @hackernewsgithubprojects

📦 cyberstrikeus/cyberstrike

CyberStrike Automates Your Red Team

CyberStrike turns a standard command prompt into an autonomous offensive security team that executes complex attack chains while you watch. This open-source harness connects thirteen specialized AI agents to over one hundred and fifty large language model providers, allowing you to run full penetration tests against web, cloud, or mobile targets. The system does not just guess vulnerabilities; it follows established frameworks like the Open Web Application Security Project to plan, exploit, and report findings with evidence. You can even offload heavy scanning tasks to remote servers using a secure protocol, keeping your local machine light.

📰 https://news.ycombinator.com/item?id=49398125

🆔 @hackernewsgithubprojects

📦 ubicomplab/embedded-arena

Embedded Arena: Real Hardware Tests for AI Agents

Embedded Arena forces AI agents to prove their coding skills on physical hardware, not just in code editors. This benchmark lets agents tweak firmware or models, then the system compiles and flashes the changes to real chips. The score isn't a simulated number; it comes from actual measurements of power, heat, and energy usage. By using real instruments to give feedback, the framework helps agents learn what actually works on tiny devices. It solves the problem of software that looks fine on a computer but fails on a real circuit board. This is a clear way to see how well agents can handle physical constraints.

🆔 @hackernewsgithubprojects

📦 debpalash/voicestudio

VoiceStudio: Your Local Voice Lab

VoiceStudio is the local voice studio that lets you clone, dub, and create audio entirely on your own hardware. While most voice tools require cloud servers and subscriptions, this open-source app keeps your data private by running sixteen text-to-speech engines and eleven transcription engines locally. You can clone a voice from a short clip, dub videos in hundreds of languages, or generate audiobooks without sending a single byte to the internet. It works on macOS, Windows, and Linux, offering a private, offline alternative to hosted services that gives you full control over your creative output.

🆔 @hackernewsgithubprojects

📦 molbal/comfyui-gguf

ComfyUI GGUF: Run Huge AI Models on Small GPUs

ComfyUI GGUF lets you shrink massive AI image models so they actually fit on your modest graphics card. Without this tool, those heavy transformer models would crash your computer or crawl at a snail’s pace. The project compresses the data into a special format, keeping quality high while slashing memory usage. It even includes a handy tool to convert your existing files in one go. This means you can test expensive AI models without buying a new workstation. It is a clever workaround that saves both money and space. Give it a try if you want to experiment with big models on a budget.

🆔 @hackernewsgithubprojects

📦 ogulcancelik/herdr-browser

Herdr Browser: Watch Your Agents Browse

Herdr Browser is the terminal plugin that finally lets you see what your AI agents are actually doing in the browser. Usually, browser automation is a blind spot where you just guess if it worked or stare at static screenshots later. This tool renders a live, interactive Chrome view directly inside your terminal layout. The best part is that while an agent drives the browser, you can grab the mouse and type right over its shoulder without stopping the process. It is like having a co-pilot who can take the wheel at any moment. No more detached windows or invisible headless runs.

🆔 @hackernewsgithubprojects

📦 shy3130/tickflow-stock-panel

The Self-Hosted Stock Analyzer

TickFlow Stock Panel runs a full quant desk on your own computer, handling stock picking, live monitoring, and backtesting without any cloud dependency. The standout feature is its AI-powered strategy builder, which lets you describe a trading idea in plain language and generates a testable rule instantly. It then scans the entire A-share market in milliseconds to find matches. You can run this locally via Docker, making it a private, zero-maintenance tool for serious hobbyists. It is not a stock picker, but a playground to test your own logic safely. If you want to understand how quant tools actually work, this open-source project is the perfect place to start.

🆔 @hackernewsgithubprojects