📦 vshulcz/deja-vu

Deja Vu

Search months of local coding conversations across multiple AI assistants in milliseconds. This tool indexes your local chat history from tools like Claude Code, Cursor, and aider, turning past conversations into a searchable memory layer without sending anything to the cloud. It automatically scrubs API keys and secrets, meaning your indexed data remains secure. You can query your past solutions from the command line or let your AI agents automatically pull in historical context before starting a new task. It is a simple way to make sure your agents never waste time re-solving a problem they already fixed.

📰 https://news.ycombinator.com/item?id=48923111

🆔 @hackernewsgithubprojects

📦 munch2u-a11y/helix-agi

Helix AGI

Run a continuous, self-improving AI agent locally without burning through your wallet on massive context windows. Helix-AGI replaces expensive semantic database lookups with an elegant, physics-based memory system. Instead of scanning massive text blocks using costly API calls, it uses a lightweight math simulation to pull relevant thoughts and memories based on concept gravity, recency, and importance. This keeps the prompt incredibly small, cutting down context usage by over ninety percent. The agent runs on a continuous background pulse, learning from its own tool failures and updating its own code documentation automatically while you sleep.

🆔 @hackernewsgithubprojects

📦 avadhesh18/at4k

AT4K Launcher

Your Android TV is likely cluttered with annoying ads and messy layouts, but a clever project called AT4K fixes this by bringing the clean, distraction-free Apple TV aesthetic to your television. It is a lightweight, privacy-focused home screen replacement that completely strips out trackers and corporate advertisements. Instead of fighting a chaotic interface, you get a simple grid of your favorite apps, including sideloaded ones, that you can drag and drop to reorganize. It works on almost any modern TV box without needing to root your device. It is the perfect way to finally reclaim your living room screen from pushy advertisers.

🆔 @hackernewsgithubprojects

📦 run-llama/liteparse-server

liteparse-server

Turn messy PDFs, images, and unstructured documents into clean, structured data on your own machine without sending your data to external services. Running a locally hosted engine can be painful, but liteparse-server wraps everything up into a simple Express or gRPC server. You just send a document to its local API, and it spits back structured JSON, clean plain text, or formatted markdown. It can even take page screenshots or run a quick scan to tell you if a document is complex enough to need heavy-duty OCR. It is the perfect, lightweight backend for self-hosted document parsing.

🆔 @hackernewsgithubprojects

📦 techjarves/openclaude-portable

OpenClaude Portable

Carry a full-featured AI coding assistant in your pocket and run it on any computer directly from a USB drive without installing a single thing. This clever project, called openclaude-portable, bundles a self-contained programming environment and web dashboard into a single folder that leaves absolutely zero footprint on the host machine. It even features a built-in speed proxy that slashes massive system prompts by ninety percent, letting you run local offline models on basic hardware without frustrating lag. Just plug your drive into any Windows, Mac, or Linux machine, launch the script, and start building with your favorite AI model anywhere you go.

🆔 @hackernewsgithubprojects

📦 adaptivemotorcontrollab/fmpose3d

fmpose3d

The 3D pose estimator that reconstructs life in three dimensions from flat, single-camera images, fmpose3d is a clever library built for both humans and animals. Instead of relying on expensive multi-camera setups or complex depth sensors, it takes a single standard 2D image and uses a mathematical process called flow matching to generate multiple highly realistic 3D poses in just a few steps. It then smartly glues those possibilities together into one incredibly accurate final skeleton. It is a massive win for tracking everything from human movement to running dogs without the usual hardware headache. Try it out and see your images leap into 3D.

🆔 @hackernewsgithubprojects

📦 nv-tlabs/ardy

Ardy

You can now control 3D virtual characters in real time using nothing but a web browser and plain English text prompts. A clever project called ardy combines autoregressive diffusion with a physics-informed model to generate incredibly fluid, responsive human and humanoid motion on the fly. Instead of just creating pre-baked animations, it lets you steer characters with live text, keyboard controls, or even by clicking on a map to drop waypoints for them to walk toward. It smoothly handles complex constraints like full-body keyframes and joint positions while keeping the movement look natural. It is basically like playing a video game where your text prompts and clicks are translated instantly...

🆔 @hackernewsgithubprojects

📦 tandpfun/wardrobe

Wardrobe

Your computer can now scan your physical clothes, isolate them from the background, and instantly generate realistic photos of you modeling them in different outfits. This is Wardrobe, a clever local-first application that uses vision models to catalog and organize your clothes automatically. By simply pointing it at your photo library or dropping in a snapshot, it extracts crisp, clean garment cutouts, detects the exact color palettes, and catalogs them locally. The most magical part is how it uses a reference photo of you to generate modeled previews of your clothes, letting you style and preview complete looks before you even put them on. It keeps your entire database, original...

🆔 @hackernewsgithubprojects

📦 wieslawsoltes/progpu

progpu

Build crazy fast, lightweight user interfaces in.NET by sending your graphics directly to the graphics card. Traditional app frameworks can be sluggish and resource-heavy, but progpu fixes this by routing all vector graphics, text layout, and visuals straight to native WebGPU draw pipelines. It basically bypasses the old, bloated rendering paths to give you lightning-fast performance across different platforms, including the browser via WebAssembly. It even integrates with popular tools like Avalonia and WinUI. If you want to build highly responsive, hardware-accelerated interfaces without the usual overhead, you should definitely check this out.

📰 https://news.ycombinator.com/item?id=48800949

🆔 @hackernewsgithubprojects

📦 tastyeffectco/sandboxd

Sandboxd

Sandboxd is the open-source engine that lets you self-host your own AI app-building platform. Instead of relying on expensive third-party services to generate and host web applications, this clever tool runs directly on your own server. When you give it a prompt, a built-in coding agent immediately spins up a secure, isolated container, writes the code, and serves the working application at its own live preview URL. It even supports running over eighty popular open-source tools with a single click. It is the ultimate way to safely host and test AI-generated projects on your own infrastructure.

📰 https://news.ycombinator.com/item?id=48939246

🆔 @hackernewsgithubprojects

📦 karna14314/pdf_tools

pdf_tools

Manage your documents securely on the go by converting, merging, splitting, and signing your files with pdf_tools. This Android application handles all your document tasks entirely on your device, meaning your sensitive files never upload to a cloud server or leave your phone. By operating completely offline without requiring internet permissions, it eliminates the privacy risks of traditional online document editors. You get a fully private workspace to scribble notes, fill out forms, and clean up images on the fly. Download it today to keep your personal data exactly where it belongs, right in your pocket.

🆔 @hackernewsgithubprojects

📦 dimillian/pokeswift

PokeSwift

Developers are actually rebuilding the entire Pokemon Red and Blue game engine from scratch using native Swift and SwiftUI. Instead of running a lazy emulator with a ROM file, this project reads the original Game Boy assembly code and assets to extract the raw maps, sprites, and rules into deterministic files. A native macOS app then reads that data to run the game natively with modern macOS UI elements, smooth SwiftUI rendering, and built-in telemetry tools. It already runs perfectly from the opening title screens all the way through Mt. Moon and Misty's gym in Cerulean City.

🆔 @hackernewsgithubprojects

📦 sysprog21/elfuse

elfuse

Run Arm64 and x86 Linux binaries directly from your macOS terminal with no virtual machine images, no Docker, and zero background daemons. elfuse makes this happen by launching each Linux program inside its own lightweight process-level hypervisor, translating Linux system calls into native macOS actions on the fly. It automatically detects the binary type, running Arm64 natively and x86 via Apple's Rosetta. This approach bypasses the slow boot times of traditional virtual machines, making it perfect for rapid testing. Download elfuse today to experience instant, lightweight Linux execution on your Apple Silicon Mac.

📰 https://news.ycombinator.com/item?id=48940084

🆔 @hackernewsgithubprojects

📦 signalpilot-labs/autofyn

autofyn

autofyn is the AI execution loop that puts language models into self-correcting cycles to hit measurable software and logic goals. Instead of asking an AI for code and hoping for the best, you give this tool a specific target, like fixing security flaws, writing comprehensive tests, or even solving complex math olympiad proofs. It automatically spins up specialized subagents to analyze the project, write solutions, review the work, and debug failures. If it gets stuck, it recalibrates and tries a new angle until the task succeeds. It is a fascinating glimpse into how we will soon delegate entire engineering tasks to autonomous, self-improving networks.

🆔 @hackernewsgithubprojects

📦 microsoft/haste

Haste

Haste is the open-source emergency framework that turns satellite data into instant disaster maps when every second counts. Built with AI-driven analysis, it takes massive remote-sensing imagery and quickly flags damaged buildings, tracks flood boundaries, and maps out destruction. What makes it brilliant is the interactive labeling tool: you can simply drag a box over structures on a map to instantly train a localized machine learning model, predicting damage across the entire region on the fly. It is a vital tool for disaster response that puts advanced geospatial intelligence right into the hands of rescue teams.

🆔 @hackernewsgithubprojects

📦 heardlabs/heard

Heard

Your AI coding agent can now talk directly to you while you step away from your keyboard. Heard is a clever macOS voice companion that narrates what your AI developer is doing in real time, meaning you can grab coffee or pace around the room while your code compiles. Instead of just reading back boring text, it dynamically decides what is actually worth saying based on your active files and tools. It even handles multiple agents running at once, assigning each one a distinct voice so you can easily tell them apart. It is like having a pair-programmer with a dry wit keeping you in the loop.

🆔 @hackernewsgithubprojects

📦 spacecell/minarrow

Minarrow

The Rust library minarrow lets you pass one million rows of data between Rust and Python in under two hundred and twenty nanoseconds. It achieves this incredible speed by serving as a minimal, ultra-low-latency implementation of the Apache Arrow data format, built from the ground up for high-performance computing. Rather than relying on heavy dependencies, it features an incredibly fast compile time and automatically aligns your data to sixty-four bytes for native parallel processing. By prioritizing simplicity and clean compile-time type safety over complex dynamic operations, it gives you a lightning-fast data bridge that feels completely native to both Rust and Python.

🆔 @hackernewsgithubprojects

📦 heyputer/firefox-wasm

firefox-wasm

Run the entire Firefox browser engine directly inside other web browsers by compiling the core Gecko engine into WebAssembly. This impressive experiment utilizes emscripten alongside a custom networking protocol to bring Firefox to any browser environment. The repository includes a built-in monorepo containing the main engine, a developer demo, and a unified benchmarking harness designed to test a JavaScript-to-WebAssembly compiler under the hood. You can test and benchmark everything from simple loops to intensive rendering scripts, opening up wild possibilities for web-based emulation, secure sandboxing, and browser-in-browser applications.

📰 https://news.ycombinator.com/item?id=48930819

🆔 @hackernewsgithubprojects

📦 tracebit-com/context-bombs

Context Bombs

Stop hostile AI agents in their tracks by turning their own safety guardrails against them. By placing specific, highly sensitive text strings inside decoy files and secrets, you can instantly shut down a machine-learning attacker. When an AI scans your network and reads these strategically placed files, its internal safety filters kick in and force it to immediately stop what it is doing. During testing against major AI models, putting a single one of these strings in an environment dropped successful admin hacks from over half down to just five percent. It is a brilliant way to turn safety rules into a defensive kill switch.

🆔 @hackernewsgithubprojects

📦 alxndrtl/little-book-rl

The Little Book of Reinforcement Learning

The Little Book of Reinforcement Learning is a concise guidebook that cuts through the dense mathematics of teaching computers to learn by trial and error. While most AI textbooks are massive, intimidating volumes, this project offers a highly accessible, short introduction paired with actual working PyTorch code for classic training methods. By stripping away unnecessary academic bloat, it lets you see exactly how algorithms make decisions step-by-step. It even includes rigorous handwritten mathematical proofs for those who want to understand the foundational logic. It is the perfect, friendly companion for any developer wanting to grasp AI decision-making without getting lost in jargon.

📰 https://news.ycombinator.com/item?id=48941104

🆔 @hackernewsgithubprojects