📦 zli12321/lhtb

Long Horizon Terminal Bench

Test how well artificial intelligence actually handles real, grueling work by dropping it into a terminal for hundreds of steps. Instead of letting agents write one quick script and brag about it, lhtb forces them to operate in a live container where their work is graded by hidden, objective verifiers. They cannot just say they are finished; they have to keep working and adapting to feedback until they truly pass or run out of time. Even the best models only solve about a quarter of these tasks, and half are completely unsolved. It is the ultimate reality check for AI agents.

🆔 @hackernewsgithubprojects

📦 w3cj/how-llms-work

how-llms-work

You can train a decoder-only GPT transformer entirely from scratch in your web browser with zero machine learning libraries. An interactive web application called how-llms-work runs a complete educational pipeline locally using raw TypeScript. Instead of hiding the math inside complex frameworks, it implements every single step of modern language models by hand, from simple pattern matching and backpropagation to subword tokenizers and word embeddings. You can inspect active neural network layers, watch word vectors cluster together in real time, and trigger multi-threaded training operations inside a clean interface. It makes the actual mechanics behind generative artificial intelligence transparent and accessible.

🆔 @hackernewsgithubprojects

📦 inferock/inferock-bench

Inferock Bench

Audit your AI bills and track actual token costs locally by routing your LLM traffic through a secure, self-hosted proxy. Instead of blindly trusting monthly invoice totals, inferock-bench acts as an independent observer that sits between your app and major providers like OpenAI, Anthropic, Gemini, and OpenRouter. It generates precise, per-call receipts that capture exact token usage, latencies, and silent errors, calculating exactly how much cash you lose on responses that timed out, cut off mid-sentence, or failed entirely. Set it up in minutes, swap two SDK lines, and start keeping your AI vendors completely honest.

📰 https://news.ycombinator.com/item?id=48868354

🆔 @hackernewsgithubprojects

📦 blaizzy/nativ

Nativ

Run frontier artificial intelligence models completely locally on your Apple Silicon Mac without any cloud subscriptions or account signups. This open-source macOS app bundles an optimized machine learning server to let you download, manage, and run highly efficient local models for chat, vision, and coding. You can use it as a private, secure desktop assistant or turn your machine into a local API server compatible with popular developer tools. It provides live performance metrics directly on your dashboard, showing memory pressure and decoding speeds. Grab this free desktop workspace to unlock your hardware's true offline intelligence.

🆔 @hackernewsgithubprojects

📦 avasol/galadriel-public

galadriel-public

Imagine having an AI assistant that actually remembers every single decision, bug, and conversation it has ever had, and can even edit its own code to fix its own bugs. Meet galadriel-public, a self-hosted Claude agent harness that builds a local, verbatim memory palace on your own machine. Unlike other agents that run up massive API bills re-reading past conversations, this project stores your history locally to query it for absolutely zero token cost. Even better, it exploits prompt caching to slash your overall Claude API costs by ninety percent. It is the ultimate playground for running a truly continuous, self-improving assistant.

🆔 @hackernewsgithubprojects

📦 wzk1015/worldcuparena

WorldCupArena

WorldCupArena is the benchmarking suite that finally puts artificial intelligence and deep-research agents to the ultimate real-time test of football prediction. Instead of evaluating models on static, outdated knowledge bases, this project challenges them to predict actual future matches before kickoff, scaling from Champions League games to the entire World Cup. The most fascinating aspect is how it measures a model's ability to research, forcing agents to browse current injury news, tactical setups, and market odds while strictly checking for information leakage. By evaluating everything from final scorelines to the exact minute of a substitution, it provides a fascinating, objective scoreboard of machine intelligence.

🆔 @hackernewsgithubprojects

📦 responsible-ai-labs/rail-score-sdk

rail-score-sdk

rail-score-sdk is the Python library that finally makes it easy to measure and guarantee the safety of your artificial intelligence applications in real time. Instead of just guessing how your language models are performing, this tool plugs directly into your existing setup to score text across eight critical dimensions, including safety, privacy, and fairness. What makes it incredibly cool is the safe regeneration loop. If a model output falls below your quality threshold, the system can automatically catch it, write a correction prompt, and run it again until it passes your safety bar. It is like having an automated editor built right into your code.

🆔 @hackernewsgithubprojects

📦 southleft/figma-console-mcp

Figma Console MCP

Connect your AI assistant directly to your design canvas to build, inspect, and tweak layouts using natural language. Instead of manually copying values, this tool lets you extract entire design systems, create polished components, and sync tokens back and forth in real time. You can even lock your AI agent onto a specific file so it keeps working in the background while you focus on other screens. It is a brilliant way to turn Figma into a fully programmable space, helping teams bridge the gap between static designs and actual code without the usual back-and-forth friction.

📰 https://news.ycombinator.com/item?id=48989470

🆔 @hackernewsgithubprojects

📦 shinmegamiboson/openplanter

OpenPlanter

An autonomous AI investigation agent called openplanter cross-references massive, messy public databases to expose hidden connections in seconds. Built with a desktop interface and a recursive engine, it takes raw data like corporate registries, campaign finance files, and government contracts to build a live knowledge graph of linked entities. Instead of manually parsing spreadsheets or writing queries, you simply ask questions in plain English. The agent then spins up recursive sub-agents to resolve duplicate names, map relationships, and flag conflicts of interest. It is a powerful, local way to turn scattered public records into clear visual evidence.

📰 https://news.ycombinator.com/item?id=48989627

🆔 @hackernewsgithubprojects

📦 ethanfel/comfyui-lora-optimizer

ComfyUI LoRA Optimizer

Stacking multiple LoRAs in ComfyUI often ruins your AI images with terrible oversaturation, weird visual artifacts, or completely lost details. The comfyui-lora-optimizer custom node suite completely fixes this by automatically calculating the absolute best way to blend your LoRAs together. Instead of blindly piling them on, this tool analyzes exactly where your model layers conflict, surgically resolves those styling clashes, and dynamically scales down their strengths so the final output looks clean. It even normalizes different trainer formats so they can merge seamlessly. It is the ultimate set-it-and-forget-it tool for clean, multi-LoRA generations.

🆔 @hackernewsgithubprojects

📦 yejunzhang/geomix

geomix

geomix is the 3D matching framework that finally solves visual localization without relying on traditional image descriptors. Instead of getting bogged down by matching complex visual textures, it uses smart geometric connections and local spatial layout to align camera images with 3D map points. By training simultaneously on multiple keypoint detectors like SIFT and SuperPoint, geomix learns to be incredibly adaptable to whatever features your system throws at it. It is a fantastic tool for anyone interested in camera tracking, robotics, or building 3D maps because it proves you can find your way using pure geometry instead of complex visual signatures.

🆔 @hackernewsgithubprojects

📦 xianzhi-ma/weaveearth

WeaveEarth

WeaveEarth is the training-free framework that finally helps vision-language models understand massive, ultra-high-resolution satellite images without choking on detail. Normally, computer vision models either scale up resolution and blow their computational budget, or zoom in blindly and lose all context. WeaveEarth solves this by acting like a smart magnifying glass. It identifies a tiny, high-impact set of localized details across a landscape, and then neatly weaves those details, spatial metadata, and layout rules together. This gives standard open-source AI models global-local joint reasoning instantly, with zero extra training. Check it out to see how it makes satellite analysis incredibly precise and fast.

🆔 @hackernewsgithubprojects

📦 yilailiu-hku/slotmem

SlotMem

SlotMem is the video generation tool that finally keeps characters looking identical across long, complex stories. If you have ever tried generating a multi-scene video, you know AI characters constantly shape-shift between cuts. This project clever-ly solves that by creating dedicated, character-specific memory slots. Think of it like a digital wardrobe and face-file that the AI checks before rendering any scene, ensuring your protagonist has the same hair, jacket, and features from start to finish. It is a massive leap forward for anyone trying to tell consistent visual stories without the constant, frustrating visual glitches.

🆔 @hackernewsgithubprojects

📦 pku-sec-lab/jetson-pi

Jetson-PI

Jetson-pi is the robot control system that finally makes advanced AI decision-making fast enough to run directly on low-power onboard hardware. While modern robots use massive vision and language models to figure out what to do, running these giant brains on a standard compact computer normally causes lag, making the robot slow, clumsy, and prone to mistakes. This project solves that by predicting what the robot's environment will look like a few steps ahead while the main AI is still thinking. This clever foresight trick lets a smaller, faster action expert command the robot instantly and smoothly without waiting around. Now, you can run high-level intelligence locally on your hardware...

🆔 @hackernewsgithubprojects

📦 danish296/codevibes

CodeVibes

Scan your GitHub repositories and get an instant sanity check on your code quality using a single AI tool. CodeVibes runs deep security and performance checks on your files, but instead of burying you in confusing dashboard metrics, it translates complex bugs and potential security holes into easy-to-read warnings and a single, unified Vibe Score. It is perfect for solo developers who want a quick, smart code review without setting up heavy enterprise infrastructure. Think of it as a friendly, sharp second pair of eyes that keeps your projects clean and secure so you can ship with confidence.

🆔 @hackernewsgithubprojects

📦 gongnyang/gongnyang-prompt-kit

Gongnyang Prompt Kit

Gongnyang-prompt-kit is the Claude Code skill that finally turns your vague, one-liner ideas into gorgeous, fully-formed image prompts. If you have ever asked an AI to make you a poster, you know that simple requests usually return boring, generic results. This clever tool acts like a code compiler, taking a quick scribble and translating it into a hyper-detailed, structured prompt complete with lighting, camera angles, and precise color palettes. It even includes a built-in validator to catch junk words and illegal settings before you hit generate. It is the ultimate shortcut to getting exactly what you pictured in your head on the very first try.

🆔 @hackernewsgithubprojects

📦 krauwu/point-as-skeleton

Point as Skeleton

Point as Skeleton changes how we test self-driving cars by generating photorealistic, interactive camera feeds from 3D points. Instead of relying on static video replays, this codebase acts as a closed-loop simulator that builds editable scenes using background point clouds and foreground vehicle templates. As a virtual driver updates its steering and speed, the system projects these point skeletons into camera views and uses diffusion models to render a matching video stream in real time. This allows developers to test how an AI driver reacts to novel, interactive scenarios. It is a massive step forward for safe, scalable autonomous vehicle training.

🆔 @hackernewsgithubprojects

📦 hulianyuyy/colt

colt

colt is the latent reasoning framework that finally gives multi-modal AI models a hidden way to think before they speak. Normally, if you want an AI to solve a complex visual puzzle, it has to write out a painfully long, slow chain of thoughts textually, which takes forever and eats up massive computing power. This project changes the game by training models to process those complex logical steps entirely in hidden, invisible mathematical vectors instead of writing out words. The result is a massive speedup, dropping visual reasoning times by over twenty times while actually scoring higher on tough benchmarks. Check it out to build faster visual AI.

🆔 @hackernewsgithubprojects

📦 datalab-to/chandra

Chandra

Turn messy PDFs, complex tables, handwritten notes, and dense math equations into clean, beautifully structured markdown or HTML. While traditional text scanners fail when they hit multi-column layouts, checkboxes, or diagrams, this smart tool extracts everything perfectly while keeping the exact layout intact. It supports over ninety languages and even handles cursive handwriting and intricate financial spreadsheets. You can run it locally or use a fast-serving engine to process hundreds of pages. It is the ultimate way to convert chaotic paperwork into neat, machine-readable digital files that are ready to search, analyze, or feed directly into your databases.

🆔 @hackernewsgithubprojects

📦 agentscope-ai/agentteams

agentteams

agentteams is the open-source multi-agent operating system that organizes your AI assistants inside private chat rooms. Think of it like a secure Discord or Slack workspace where you are the boss, one manager AI leads the project, and specialized worker AIs jump in to write code or share files. Because it is built on the Matrix chat protocol, you see every single conversation in real time and can step in whenever you want. There are no black boxes here. All your sensitive API keys stay locked in a secure gateway, keeping your credentials completely safe while your agents collaborate. It runs entirely on your own machine with a single command.

🆔 @hackernewsgithubprojects