📦 mindlab-research/longstraw

LongStraw

Scale up reinforcement learning post-training to multi-million-token contexts without needing a massive GPU budget. Longstraw makes this possible by introducing a smart memory-saving technique called resident-prefix training. Normally, computing reward-based updates requires holding huge, complex autograd graphs for both the prompt and the response in memory at the same time. Instead, this system evaluates the shared prompt exactly once, locks that resident state in place, and then replays only the short trainable responses one at a time. It successfully processed a massive 2.1-million-token prompt workload, proving that we can train smarter on the hardware we already own.

🆔 @hackernewsgithubprojects

📦 kunkka19xx/look

look

This lightweight launcher runs searches in under one millisecond while keeping your data completely local. The project, called look, is an open-source, keyboard-first alternative to Spotlight and Raycast built for macOS, Windows, and Linux. It scans your local files, folders, and clipboard history instantly without sending any data to the cloud. You can also calculate math on the fly, kill system processes by port, and manage daily tasks directly from the search bar. Because it is built as a single native app with no background daemons, it consumes minimal system memory. It is the perfect, privacy-first productivity tool that stays out of your way.

🆔 @hackernewsgithubprojects

📦 openbmb/minicpm-robot

MiniCPM-Robot

Control your physical robots with an ultra-fast, local AI brain that fits on a single tiny chip. The openbmb/minicpm-robot repository provides a family of lightweight vision-language-action models designed specifically for real-world robotics. Its standout feature is an incredibly compact point-nine-billion parameter tracking model that runs fully on-device. It processes visual inputs and natural language commands locally on a quadruped robot at over five frames per second with just one hundred eighty milliseconds of latency. This means your robot dog can track targets and navigate complex environments dynamically without needing any external servers or internet connection. It is the perfect starting point for building responsive, intelligent machines.

🆔 @hackernewsgithubprojects

📦 oblien/openship

Openship

Openship is the self-hosted deployment platform that finally makes shipping code to your own servers feel like a premium cloud experience. You just point it at your repository, and it automatically detects your stack, builds your application, and handles the hosting. It completely eliminates the headache of complex configuration files, custom pipelines, and messy server management. The most exciting part is that it bundles your databases, domains, SSL certificates, backups, and even a fully featured mail server into one cohesive system. It gives you absolute control over your applications without sacrificing the convenience of modern push-to-deploy workflows.

🆔 @hackernewsgithubprojects

📦 microsoft/ontology-playground

Ontology Playground

Design and explore complex data relationships visually right in your browser without dealing with any messy database setup. This tool gives you an interactive, drag-and-drop canvas where you can build node-and-edge diagrams to map out how different concepts connect. You can load pre-built templates for industries like healthcare or finance, edit them on the fly, and instantly export your creations as standard data models. It is a brilliant way to wrap your head around structured data mapping. Grab a template, connect some nodes, and see how simple visualizing complex schemas can actually be.

🆔 @hackernewsgithubprojects

📦 praneybehl/llm-wiki-plugin

LLM Wiki Plugin

The second brain for AI agents that finally gives your digital assistants a shared, persistent memory. LLM Wiki Plugin is an open-source tool that turns your PDFs, articles, and meeting notes into a self-maintaining markdown knowledge base. Instead of starting every conversation from scratch, your coding agents can search, cite, and update this shared wiki directly inside your project directory. The absolute best part is that it requires no complex databases or embedding services, relying entirely on plain text files that you can easily read, edit, and track in git. Try it out to keep your projects organized.

🆔 @hackernewsgithubprojects

📦 manuxd32/ai-doomsday-toolbox

AI Doomsday Toolbox

Turn your pile of old, dusty Android phones into a fully functioning, offline AI supercomputer. With the ai-doomsday-toolbox, you can connect multiple phones over your home network to share the heavy lifting of running large language models, transcribing audio, and generating images completely offline. It is the ultimate way to rescue old hardware from the junk drawer and build a private edge-compute cluster right on your desk. You can even generate custom datasets or run an AI-powered virtual pet. Grab your old devices and start building your own independent, pocket-sized server farm today.

🆔 @hackernewsgithubprojects

📦 alexeygrigorev/ai-engineering-field-guide

ai-engineering-field-guide

Seven out of ten generative AI projects never actually make it past the pilot phase. To bridge this gap, the ai-engineering-field-guide provides a data-driven resource built entirely on real-world hiring and engineering data. Instead of relying on generic advice, this repository analyzes nearly five thousand actual job descriptions and over one hundred real interview experiences. It maps out exact learning paths for transitions from backend or data roles, lists real take-home assignments, and catalogs system design case studies from major tech companies. It is an incredibly practical blueprint for anyone trying to understand what modern companies actually look for when hiring.

🆔 @hackernewsgithubprojects

📦 0xk3vin/megamemory

MegaMemory

MegaMemory is the persistent knowledge graph tool that finally gives your AI coding agents a reliable long-term memory across development sessions. Instead of wasting time parsing raw code or generating noisy structural trees, this clever system lets your AI agent act as the indexer, writing down key architectural concepts, patterns, and design decisions in plain language. It runs a local SQLite database and generates in-process embeddings entirely on your machine without needing external API keys. Your agent simply queries its own memory to orient itself instantly, and can even merge conflicting memories when switching branches. You can even spin up a local web explorer to visually inspect your project's concept...

🆔 @hackernewsgithubprojects

📦 korroai/mue-x

Mue-x

Mue-x is the self-evolving AI assistant that actually rewrites its own source code in real-time as you watch. Instead of just running prompts, this autonomous assistant constantly reads its own Python files, generates and validates code mutations, and applies them directly to its own system. It even scans GitHub to clone relevant projects, autonomously absorbing new patterns and skills based on what you are talking about. With built-in safeguards to prevent errors and roll back bad code, it is a fascinating look at truly self-modifying software. It is a wild experiment showing how programs might write themselves in the future.

🆔 @hackernewsgithubprojects

📦 sneg55/pingcrm

PingCRM

Keep your personal and professional relationships from going cold by automatically pulling all your interactions into one private, self-hosted dashboard. This clever tool syncs with your Gmail, Telegram, Twitter, and WhatsApp, merging the same person across different networks into a single, unified timeline. It calculates a simple score to warn you when someone is sliding out of touch and even drafts personalized follow-up messages using AI so you know exactly what to say next. It is the perfect, private way to stay thoughtful without the manual tracking headache, leaving you with a perfectly organized digital Rolodex that practically runs itself.

📰 https://news.ycombinator.com/item?id=47423835

🆔 @hackernewsgithubprojects

📦 initializ/forge

Forge

Deploy portable, hardened AI agents right on your own infrastructure with Forge. This open-source runtime lets you run self-contained agents with custom skills next to your existing services, whether you are running on Kubernetes, on-premise, or even in air-gapped environments. What makes it incredibly interesting is how it secures these agents by default, using outbound-only connections, strict network controls, and detailed audit logging to keep everything safe. You can define specific agent skills in a simple markdown file, and the system compiles them into secure containers. It is the perfect, straightforward way to run reliable local agents without relying on third-party cloud connections.

📰 https://news.ycombinator.com/item?id=47323134

🆔 @hackernewsgithubprojects

📦 aws-samples/sample-geospatial-agent-on-aws

sample-geospatial-agent-on-aws

sample-geospatial-agent-on-aws is the AI assistant that turns simple English questions into fully rendered satellite map analysis. Instead of wrestling with complex satellite data portals, you just ask it to show things like vegetation health in Central Park or wildfire damage in Los Angeles. The app instantly grabs real Sentinel-two satellite imagery, calculates environmental indexes behind the scenes, and plots the results on an interactive map. It makes heavy-duty earth observation data feel as simple as sending a chat message. If you want to build an intelligent map interface that actually understands natural language, this is the perfect blueprint to explore.

🆔 @hackernewsgithubprojects

📦 egocs-400k/dataset

Dataset

Dataset is the gaming repository that finally bridges first-person gameplay video with precise controller inputs. It packs over ten thousand hours of Counter-Strike footage, but instead of just raw video, it aligns every single frame with per-tick keyboard and mouse movements, player health, and in-game actions. Using smart math, it clips the footage into perfect, model-ready moments without cutting off a crucial reload or jump mid-action. It is the ultimate playground for training AI to actually understand and play games like a human. Check it out to see how raw gameplay becomes structured data.

🆔 @hackernewsgithubprojects

📦 arcboxlabs/arcbox

arcbox

You can now run full, isolated Linux microVMs on macOS with a boot time under two hundred milliseconds. A brand new open-source container runtime called arcbox is built entirely from scratch in Rust to replace closed-source virtualization tools on Apple Silicon. Instead of relying on bloated virtual machines, it uses a custom hypervisor backend to spin up lightweight sandboxes for running AI agents, untrusted code, or local testing. It even exposes a docker-compatible socket, meaning your existing compose files and tools just work. It is a incredibly fast, local-first playground that lets you build secure sandboxes on your own machine.

📰 https://news.ycombinator.com/item?id=48980217

🆔 @hackernewsgithubprojects

📦 rasinmuhammed/misata

Misata Synthetic Data Engine

Misata is the open-source synthetic data engine that finally lets you declare your exact analytical outcomes instead of trying to mimic real databases. While other tools struggle to copy existing datasets, Misata lets you write down the exact answers you want, like aiming for a specific revenue curve or a precise three-percent fraud rate, and then mathematically builds individual rows from scratch to match that target to the penny. It uses clever closed-form math instead of heavy machine learning models, ensuring you get perfect relational integrity across your tables with zero data leaks. You can run it directly from a single plain-English sentence, connect it to your database, or use...

🆔 @hackernewsgithubprojects

📦 gakonst/nanocodex

Nanocodex

Nanocodex is the embeddable Rust library that finally brings fast, stateful AI coding agents directly into your local process without the bloat of an application server. Instead of dealing with massive, slow runtimes, this clever SDK connects straight to OpenAI's real-time WebSockets to run a persistent coding loop. You get lightning-fast workspace edits, file patching, and shell execution that run up to four times faster by resuming directly from memory checkpoints rather than re-sending your entire history. It is brilliant for building custom command-line tools or embedding smart agents into your own apps. Check it out and build faster.

📰 https://news.ycombinator.com/item?id=48970300

🆔 @hackernewsgithubprojects

📦 timskap/box3d-unity

Box3D Unity

Erin Catto, the creator of the legendary Box2D engine, has developed a new high-performance 3D physics engine, and it now integrates directly with Unity. Called box3d-unity, this project bridges a portable C17 physical simulation backend into a familiar component workflow. Instead of writing custom bindings, you get ready-to-use rigid bodies, colliders, joints, and continuous collision detection backed by the native library via P/Invoke. It is incredibly useful because you get high-speed, cross-platform performance without having to manage complex native compilation. Drop the package into your project, add a physics world component to your scene, and your game objects are ready to roll.

🆔 @hackernewsgithubprojects

📦 dondai1234/master-fetch

Master Fetch

Master-fetch is the model context protocol server that finally gives your local AI agent complete, unrestricted access to the web without requiring any subscriptions or API keys. Instead of getting blocked by paywalls or complex bot detection, this clever tool runs directly on your machine to let your agent search, crawl, and fetch live web pages. It seamlessly combines ten parallel search engines with a smart local browser that bypasses typical web blocks, and it can even run character recognition on scanned documents. It gives your AI helper a powerful, self-hosted set of eyes for deep online research at zero cost.

🆔 @hackernewsgithubprojects

📦 frank-zy-dou/dynamics-modeling

dynamics-modeling

The dynamics-modeling project is the robotics framework that finally brings accurate force perception to low-cost robot arms without expensive physical sensors. By training a smart transformer model through a differentiable simulator, it maps basic motor signals, like positions and currents, directly to real-world force estimates. It essentially turns cheap servos into virtual force sensors, allowing a budget arm to accurately feel when it is picking up payloads or hitting an obstacle. This award-winning MIT research makes complex force-aware robot control accessible to everyday developers. Grab the open-source code and pre-trained models to upgrade your hardware with instant physical awareness.

🆔 @hackernewsgithubprojects