📦 openmoss/moss-vl

moss-vl

Chat with an AI about a video stream while the camera is still rolling, instead of waiting for the recording to finish. The moss-vl project makes this possible by processing live video feeds and text simultaneously. This open-weight model lets you interrupt it with questions at any timestamp, and it answers instantly based on the frames it has seen so far. It can even correct its own previous statements as new frames arrive, or autonomously choose to stay silent if it needs to watch more. It is basically like having a smart friend watch a live feed right alongside you.

🆔 @hackernewsgithubprojects

📦 xiaomirobotics/xiaomi-robotics-0

Xiaomi-Robotics-0

Xiaomi-Robotics-0 is the vision-language-action model that brings lightning-fast execution directly to consumer-grade hardware. It is built to bridge the gap between complex artificial intelligence and actual robot control by translating visual camera feeds and simple text commands into precise physical actions in real-time. Instead of requiring a massive industrial server room, it is optimized to run efficiently on a single consumer GPU by using smart asynchronous execution to keep latency incredibly low. It handles multi-view camera inputs, like a wide desk view combined with a wrist camera, to figure out exactly how a robot arm should move to grasp and manipulate objects. It is a fantastic, accessible playground for anyone...

🆔 @hackernewsgithubprojects

📦 benoks/eshot

EShot: The ultimate open-source screenshot and recording tool

eshot is the lightweight tray application that packs an entire screen capture studio into a single utility for Windows and Linux. Instead of juggling separate tools for screenshots, screen recording, and text recognition, this project handles it all from one system tray icon. You can select a region, annotate it with arrows or highlight text, and instantly pin the capture above other windows. But the standout feature is its ability to directly extract text with built-in optical character recognition or trigger an instant visual search. It even lets you record selected regions straight to smooth GIFs or MP4 videos, keeping your daily workflow incredibly simple and self-contained.

🆔 @hackernewsgithubprojects

📦 littledivy/mimic

mimic

Capture the network traffic of any mobile app and automatically turn it into a clean Python library using AI. If you want to automate actions or fetch data from an app that does not have an official public API, mimic makes it incredibly easy. You just run a command to route your phone's traffic through a local proxy, use the app normally for a few seconds, and let mimic analyze the security headers and endpoints. It then uses Claude to generate a fully functional Python client file for you. You can literally call the app's internal functions with plain Python code.

🆔 @hackernewsgithubprojects

📦 clawkwork/clawk

clawk

Giving an AI coding agent full access to your terminal is incredibly risky, but clawk completely eliminates that danger by letting the agent work inside its own disposable Linux virtual machine instead of yours. It runs your code in a locked-down container-like environment, allowing Claude or other agents to install packages and run commands at full speed. Best of all, outbound network traffic is blocked by default, and your keys never enter the guest machine, keeping your personal files and tokens perfectly secure. If the agent somehow breaks the environment, a single command completely destroys the VM and spins up a fresh one instantly.

📰 https://news.ycombinator.com/item?id=48892859

🆔 @hackernewsgithubprojects

📦 camel-ai/seta

Seta

Scale your reinforcement learning setups for terminal-based AI agents without tearing your hair out. The seta repository provides a dedicated, highly scalable framework built for training and evaluating CAMEL terminal agents using real-world terminal environments. Managing multiple execution environments can easily bottleneck your training, but this project solves that by introducing remote CPU server clustering, distributed docker slot pools, and robust scheduling services. It allows you to run concurrent rollouts and test complex terminal interactions seamlessly across remote infrastructure. If you want to train terminal-operating agents efficiently at scale, download this repository to spin up your distributed agent training environment today.

🆔 @hackernewsgithubprojects

📦 nvidia-nemo/labs-molt

labs-molt

labs-molt is the agentic reinforcement learning framework that finally lets you scale complex, multi-turn AI agents to trillion-parameter models in pure PyTorch. Instead of wrestling with bulky, rigid platforms, this elegant library gives you a incredibly lightweight codebase that manages everything from multi-step tool use to vision-language tasks using a single actor. You can write your agent rewards in plain Python, while the system quietly orchestrates model training, async rollout queues, and weight syncing in the background. It is the ultimate playground for developers who want to run advanced AI training experiments without the engineering headache.

🆔 @hackernewsgithubprojects

📦 ramonvermeulen/whosthere

Whosthere Network Discovery

You can now scan and map your entire home network without typing complex commands or needing administrator privileges. A security tool called whosthere lets you discover active devices on your local network using a clean terminal interface. By utilizing clever techniques like reading the system's ARP cache and listening to local broadcast signals, it safely finds computers, smart home tech, and phones. The app even resolves manufacturer names and lets you trigger quick port scans directly from your keyboard. It is a fantastic way to easily spot unknown devices on your Wi-Fi and keep your home network secure.

🆔 @hackernewsgithubprojects

📦 potatameister/paperknife

PaperKnife

Your sensitive documents never leave your physical device when you edit them with PaperKnife. This privacy-focused PDF utility handles everything from merging, splitting, and compressing to signing and protecting files using a clever architecture that runs entirely in your local browser memory. By processing tasks client-side with sandboxed WebAssembly, it ensures your tax forms and contracts are never uploaded to a third-party server or stored in a database. It is a completely self-contained, offline-capable tool that gives you absolute security and control. Grab the Android app or load it in your browser to process your files with peace of mind.

🆔 @hackernewsgithubprojects

📦 lazovelko/neverclick

Neverclick

You can completely control your computer mouse using only your keyboard and a clever local computer vision engine. This is made possible by a desktop app called neverclick, which scans your screen and overlays letter tags on top of every clickable button, link, and text box. To click anything, you just type the letters on its label, and the app instantly moves your cursor and clicks for you. It runs entirely offline, requires zero configuration, and bypasses the mouse completely to save your wrists from strain. If you want to give your hands a break, it is a brilliant way to navigate.

📰 https://news.ycombinator.com/item?id=48877485

🆔 @hackernewsgithubprojects

📦 wan-video/wan-dancer

Wan-Dancer

Generate long, beautifully synced dance videos from any music track using a single reference photo. Most AI video tools fall apart after fifteen seconds, turning movement into a chaotic, glitchy mess. That is where wan-dancer comes in. It uses a clever two-step trick to plan the entire dance routine globally before going back to polish up the high-resolution details. This means you can create over a minute of continuous, high-definition choreography in styles like street dance, K-pop, or classical, all perfectly matched to the beat. Grab your favorite audio track and watch your static images come to life.

🆔 @hackernewsgithubprojects

📦 bigstationw/comfyui-conditioningnoiseinjection

ComfyUI Conditioning Noise Injection

Shake up your stale image generation workflows by forcing stubborn AI models to actually give you diverse results. Some fast-rendering models suffer from incredibly low seed variance, meaning they churn out almost the exact same image even when you change the seed. The comfyui-conditioningnoiseinjection custom node fixes this by injecting controlled chaos directly into your prompt embeddings at the very start of the process, before cleaning it back up. This simple trick forces the model to explore wildly different compositions while still sticking to your prompt. Grab this node to finally get the variety you actually asked for.

🆔 @hackernewsgithubprojects

📦 changetheconstants/seedvarianceenhancer

SeedVarianceEnhancer

The ComfyUI custom node that finally injects real variety into your image generation, seedvarianceenhancer stops your AI models from spitting out the same boring, repetitive compositions. Sometimes when you change the random seed, your image generator barely changes the actual picture. This clever node fixes that by adding a controlled splash of random noise directly to your prompt's text embedding during the initial steps. It gives the AI a gentle nudge to explore wildly different creative paths before it settles down to finish the details. It is the perfect tool for breaking out of creative ruts and discovering completely unexpected visual ideas.

🆔 @hackernewsgithubprojects

📦 microck/ordinary-claude-skills

ordinary-claude-skills

Upgrade your AI assistant instantly by loading hundreds of pre-made prompts and scripts that teach your AI new tricks without cluttering its memory. With ordinary-claude-skills, you get a massive local warehouse of over six hundred community-built capabilities, ranging from advanced coding helpers to niche scientific tools. Instead of explaining your context every single time, you simply load these bite-sized instructions only when you need them, keeping your chats clean and smart. You can search the organized web catalog or dump the raw files straight into your own setup. Grab this collection and give your AI some serious real-world superpowers.

🆔 @hackernewsgithubprojects

📦 getclawe/clawe

Clawe

You can now run a full digital agency where the employees are all AI agents collaborating on a virtual Kanban board. An open-source project called clawe lets you deploy a team of specialized AI workers, like a content writer, a designer, and an SEO specialist, who coordinate in real-time. These agents wake up on their own scheduled heartbeats, check their shared inbox, and update a Trello-like board using their own CLI tools. You can watch them assign tasks, share files, and chat through a clean web dashboard. It is a fascinating peek into the future of autonomous, multi-agent teamwork.

🆔 @hackernewsgithubprojects

📦 prismml-eng/llama.cpp

llama.cpp

Run massive artificial intelligence models directly on your everyday computer without relying on expensive cloud servers. The llama.cpp project provides a lightweight, pure C and C plus plus implementation that runs language models locally with incredibly fast performance. Instead of requiring a massive graphics card, it uses smart math called quantization to shrink these heavy models so they fit onto normal laptops, even utilizing your regular processor alongside your graphics memory. It supports almost every major open model out there right now. You can host your own private, secure, and lightning-fast AI assistant entirely offline.

🆔 @hackernewsgithubprojects

📦 prismml-eng/bonsai-demo

Bonsai Demo

You can now run a massive twenty-seven billion parameter artificial intelligence model directly on a standard smartphone or a normal consumer laptop without running out of memory. The bonsai-demo project achieves this wizardry by compressing large language models down to a tiny one-bit or ternary footprint, packing complex weights so efficiently that they fit into a fraction of the space normally required. It gives you fully localized vision capabilities, deep reasoning, and tool calling with zero external API fees. It is basically the holy grail of offline, private intelligence that actually runs fast on everyday hardware you already own.

🆔 @hackernewsgithubprojects

📦 prettysmartdev/awman

awman

You can now run and coordinate multiple AI code agents simultaneously in container-isolated sandboxes right from your terminal. awman is a terminal interface and API that automates your entire development lifecycle from an open GitHub issue to a merged pull request. It runs different agents like Claude or Codex in parallel tabs, isolating their environments with Docker and Git worktrees so they do not conflict. By defining structured workflows in simple configuration files, you can coordinate step-by-step tasks like planning, coding, and testing. You can even run it in yolo mode to let agents autonomously fix bugs, check CI status, and push pull requests while you watch the progress update.

🆔 @hackernewsgithubprojects

📦 wassermanproductions/wassermans-filmmaker-suite

Wasserman's Filmmaker Suite

Choreograph 3D camera moves and actors on a virtual stage, then export exact motion files that AI video generators can actually understand. Wassermans-filmmaker-suite is a collection of four desktop applications built to bridge the gap between traditional pre-production and modern generative video. Instead of fighting with random text prompts, creators can block out scenes, extract motion data from real footage, and even split audio tracks into separate stems locally. An integrated server system even lets external AI assistants automate editing tasks right inside DaVinci Resolve. It turns chaotic video generation into a precise, controllable filmmaking workflow.

🆔 @hackernewsgithubprojects

📦 nv-tlabs/learning-convex-decomposition

learning-convex-decomposition

Break down complex 3D shapes into simpler, snug-fitting blocks to make physics simulations and collision detection run incredibly fast. The learning-convex-decomposition project introduces a clever feed-forward model that predicts a continuous feature field across a 3D object, which then clusters itself into neat convex parts using a purely geometric, self-supervised process. What makes this so exciting is that it is the first open-world method of its kind, meaning it generalizes beautifully to completely unseen objects, whether they are represented as standard meshes, CAD models, or even Gaussian splats, helping game developers and animators instantly optimize complex 3D scenes.

🆔 @hackernewsgithubprojects