📦 apple-aiml-research/ml-lensvlm

LensVLM: The AI That Only Reads What It Needs

LensVLM is a nine-billion parameter model that solves the headache of processing massive documents without burning through your computer's memory. Instead of reading every single word, it scans low-resolution thumbnails of your pages and uses a tool to zoom in on only the specific text it needs to answer your question. It is essentially like a very focused librarian who only pulls the book from the shelf when absolutely necessary. This approach keeps the model fast and efficient while still finding deep details buried in long reports. If you are working with heavy documents, this is the smart way to handle them.

🆔 @hackernewsgithubprojects

📦 ilyap358/fluxcast

FluxCast: Stream Your Linux Desktop to the Big Screen

A standard Linux laptop can wirelessly mirror its entire desktop to a Smart TV using a feature called Wi-Fi Direct. This is what FluxCast does. It captures your screen and audio in real time and sends that video stream directly to a compatible TV over your local Wi-Fi network. No cables, no external hardware, just a direct connection between two devices. It is useful for anyone who wants to use their laptop as a media center for their living room, without buying an expensive streaming box or dealing with the hassle of finding a long HDMI cable.

📰 https://news.ycombinator.com/item?id=49835593

🆔 @hackernewsgithubprojects

📦 valentynkit/awesome-jev-typesafe

What is Awesome Jev Typesafe?

Awesome Jev Typesafe is the curated collection that finally maps the entire System One model ecosystem in one place. It solves the confusion of how to use this new kind of AI, which answers specific questions with a confidence score instead of writing long, rambling paragraphs. The repository acts as a central guide, listing official tools and community projects that use these quick, typed judgments for coding agents and browser automation. Instead of guessing what an AI wants, you give it a strict format and get a clear yes, no, or score back. This is how you build systems that make decisions without needing to read a novel every time.

🆔 @hackernewsgithubprojects

📦 dci-agent/dci-agent-lite

How AI Actually Searches Files

DCI-Agent-Lite is the search tool that reads your files directly instead of using complex indexes. Most search systems require you to pre-process documents into a database, which is slow and rigid. This project skips that entirely. It lets an AI agent use basic command-line tools to search, inspect, and cross-check your raw text files on the fly. This means your data stays private on your machine, and the agent can dig deep into long documents without losing track. It turns your computer into a simple, powerful research environment.

🆔 @hackernewsgithubprojects

📦 abdelstark/awesome-typesafe-jev

Awesome Jev: Your Field Guide to TypeSafe

Stop guessing if TypeSafe's System One model actually works. Awesome Jev is a hand-picked field guide that cuts through the noise by listing real SDKs, live demos, and honest evaluations. It doesn't just link code; it demands proof, forcing every entry to show its data and limitations so you know exactly what you're building. This is your shortcut to finding tools that actually function without falling for marketing fluff. You get a curated map of the ecosystem, not a cluttered directory. Check it out to build with confidence, not guesswork.

🆔 @hackernewsgithubprojects

📦 genai-security-project/agent-control-standard

Agent Control Standard

A Guardian Agent can literally edit an AI agent's plan before it runs, rewriting a dangerous command into something safe without stopping the whole process. This is the Agent Control Standard, an open specification that gives security teams a way to inspect, block, or modify what artificial intelligence systems are about to do. It works like a strict gatekeeper sitting between the brain of the system and its actual tools. Instead of just saying no, the gatekeeper can change the instructions to match your company rules, ensuring nothing harmful ever leaves the building. This standard creates a clear audit trail so you can see exactly why a decision happened.

🆔 @hackernewsgithubprojects

📦 0xsero/omarchy-local-ai

Run Local AI on Your GPU

Omarchi Local A I puts a powerful coding assistant directly on your own graphics card, removing the need to send your code to outside servers. The system automatically identifies your specific hardware and loads a model that has been proven to run correctly on it, so you avoid wasting time guessing at settings. Once the model is ready, you can launch a terminal-based agent to write code, and with a single click, you can share that model with your private network. This approach gives you full control over your data while keeping your workflow fast and local. You get a reliable, private AI helper without any complicated configuration.

🆔 @hackernewsgithubprojects

📦 dgreenheck/tidewater

Tidewater: The Browser Game That Should Not Exist

Tidewater is the browser game that puts a full 3D ocean simulation in your tab without downloading a single file. It is a complete coastal island where you cast a line, fight fish, and sell your catch to a local vendor, all running directly on your graphics card. The wild part is that the water actually behaves like water, with waves breaking on the sand and light filtering through the surface. It is built from scratch with no heavy frameworks, just raw code and some really clever math. If you want to see how much is possible in a web browser right now, this is the project to watch.

🆔 @hackernewsgithubprojects

📦 merijjeyn/jive

Jive: The Agent That Thinks in Graphs

Jive is the terminal coding agent that replaces linear step-by-step tool calls with executable graphs. Most AI agents execute one action at a time, waiting for each result before deciding the next move, which makes repetitive tasks slow and expensive. Jive solves this by letting the AI plan a whole workflow as a connected map of steps that can run in parallel or loop automatically. The surprising part is how it mixes deep thinking with fast, intuitive decisions so the agent doesn't waste time re-evaluating simple follow-ups. This approach dramatically cuts down on wasted tokens and processing time.

🆔 @hackernewsgithubprojects

📦 swarajbachu/zuse

Zuse: The AI Coding Hub

Zuse is the open source desktop app that finally unifies your scattered AI coding agents into one persistent workspace. It wraps major coding tools like Claude and Codex into a single chat-first interface, so you stop jumping between terminals and losing your work. The coolest part is that it gives every chat its own isolated git worktree, letting you experiment with code changes without ever touching your main project. If you use AI to code, this turns chaos into a calm, organized flow that actually keeps up with your brain.

📰 https://news.ycombinator.com/item?id=49837957

🆔 @hackernewsgithubprojects

📦 dcmmc/macpad

macPad: A Touch Mac in Your iPad

Run full macOS desktop applications directly on a jailbroken iPad, using the device's actual hardware graphics instead of slow translation layers. This project lets you use touch gestures like scrolling, multi-finger zoom, and three-finger swipes to control windows, just like you would on a real Mac. It bridges the gap between your tablet and your computer by making the iPad's screen act as a trackpad and the touch input drive the desktop interface. If you have ever wanted to run heavy Mac apps on the go, this is a fascinating look at how deep customization can go. It turns your pocket-sized device into a powerful workstation without needing a laptop.

📰 https://news.ycombinator.com/item?id=49809600

🆔 @hackernewsgithubprojects

📦 theoleecj/semif-openjev

SemIf: Instant AI Decisions

SemIf is the decision engine that reads answers directly from the model without generating text. Instead of waiting for a chatbot to type out a sentence that software has to parse, it reads the raw probabilities for specific options in a single pass. This cuts processing time from five seconds to just one second, making it ideal for routing requests or checking evidence in real time. It runs on standard hardware and even in your browser, proving that open models can handle complex logic without proprietary services.

🆔 @hackernewsgithubprojects

📦 typesafe-ai/system-one-adapter-python

System One Adapter: Test Any LLM Like a Human Evaluator

System One Adapter is the testing tool that lets you swap any large language model into the TypeSafe evaluation framework without rewriting your code. Instead of relying on a single proprietary service, this library acts as a universal translator, allowing you to point your evaluation questions at OpenAI, Anthropic, or Gemini providers just like a plug-and-play adapter. This is incredibly useful for developers who need to compare the cost, speed, and intelligence of different AI models side by side to see which one actually understands nuance best. By keeping the evaluation logic consistent while changing the underlying engine, you can make data-driven decisions about which model truly fits your project.

🆔 @hackernewsgithubprojects

📦 y0usaf/pi-jev

pi-jev: The Bouncer for AI Agents

Your AI coding assistant is about to run a command that could wipe your project, and it has no instinct for danger. pi-jev fixes this by acting as a strict bouncer for your AI agent. Before it executes any risky action, it asks a specialized model for a simple yes or no answer on whether the move is destructive or out of scope. If the risk is too high, it stops the action dead in its tracks and asks you for permission. It’s not about blocking your AI, it’s about giving it a safety net so you can trust it with real work without holding your breath.

🆔 @hackernewsgithubprojects

📦 nestrilabs/virtio-nvgpu

virtio-nvgpu

Render inside a virtual machine at the exact same speed as bare metal, while sharing the same physical graphics card. This project lets a guest system run the real, unmodified NVIDIA drivers, bypassing the heavy translation layer that usually slows down virtual graphics. Instead of serializing every single draw call across the virtual boundary, it forwards the low-level driver commands once, letting the guest write directly to the host's memory. The result is stunning: a guest machine renders within two percent of the physical limit, costs nothing in extra processor power, and even encodes video on the GPU itself.

📰 https://news.ycombinator.com/item?id=49824864

🆔 @hackernewsgithubprojects

📦 vigozhao/ai-visual-prompt-cookbook

The AI Style Cookbook

You can copy a single text file and make an AI image generator paint exactly like a specific magazine or comic book. A new project called the AI Visual Prompt Cookbook collects over a hundred of these ready-made recipes. Usually, getting an AI to draw in a consistent style takes hours of testing and tweaking. This project solves that problem by packaging the perfect words into a simple file you can just copy and paste. It works with popular tools like Midjourney and GPT. The collection is updated daily, so there is always a new look to try.

🆔 @hackernewsgithubprojects

📦 freestylefly/wechatbridge

WeChatBridge: The Mac App That Feeds WeChat Chats to AI

WeChatBridge is the Mac tool that finally stops your WeChat history from rotting in the app. It taps directly into the native share menu so you can grab a chat export and blast it right into your favorite AI agents or local notes. The coolest part? It turns messy group threads into clean, readable files for you or your AI to analyze, all without uploading anything to the cloud. It is the perfect bridge for developers and note-takers who want to leverage their digital life. Give it a shot to see how easy it is to make your chats actually work for you.

🆔 @hackernewsgithubprojects

📦 kyotofin/tax-doc-classifier

tax-doc-classifier: Classify 261 Tax Forms for a Penny

Identify exactly which IRS form every page in your tax documents belongs to, instantly and for almost nothing. This tool uses a specialized decision model to look at a page and pick from 261 specific forms, giving you a high-confidence answer without training a single AI model. It replaces expensive and slower large language models with a setup that is thirty-four times cheaper and six times faster. The entire classifier is just a generated list of form descriptions, so it is easy to update whenever the IRS changes a document. It is a clever, low-cost way to automatically organize massive stacks of tax paperwork.

🆔 @hackernewsgithubprojects

📦 undefined-ui/second-brain-os

Second Brain OS

An AI agent that actually organizes your notes for you is finally here. Most people save articles, bookmarks, and ideas with the best of intentions, but the manual filing and linking gets boring fast, so the system dies. Second Brain OS solves this by letting an AI agent do the tedious work. It reads your raw dumps of clips and transcripts, turns them into clean, linked markdown pages, and then automatically updates your existing notes when new information arrives. This means your personal knowledge base compounds in value instead of just growing messy. The agent handles the deduplication and linking so you can just ask it questions later.

🆔 @hackernewsgithubprojects

📦 agricidaniel/compass

Compass: Your Life in One Folder

One folder, not ten apps, now runs your entire life. Compass turns a plain text note app into a living command center where your journal, tasks, habits, and writing all talk to each other. The coolest part is the AI assistant built right into the notes. It reads your daily reflections and offers help, but it never changes a single word without your explicit permission. This keeps your personal data safe while giving you a smart helper for planning and writing. It is a simple, powerful way to organize your thoughts without losing control. Stop switching tabs and start using one honest system that actually works.

🆔 @hackernewsgithubprojects