This media is not supported in your browser
VIEW IN TELEGRAM
π¦ aaryanverma/inkrail
Inkrail β a typewriter that actually feels like one
Inkrail is the quiet, offline typewriter app that makes writing feel like paper, ink, and a little room to breathe. It turns your Mac or Windows computer into a warm desk where letters appear on textured paper instead of a sterile screen. What makes it clever is the moving machine: as you type, the capsule and rail glide along below your words, like a real typewriter. You can choose paper, ink, fonts, lamps, sounds, and desk scenes, then switch to Focus Mode for a wider page and a calmer view. It saves locally, works offline, and is completely free, no accounts or subscriptions.
π @hackernewsgithubprojects
Inkrail β a typewriter that actually feels like one
Inkrail is the quiet, offline typewriter app that makes writing feel like paper, ink, and a little room to breathe. It turns your Mac or Windows computer into a warm desk where letters appear on textured paper instead of a sterile screen. What makes it clever is the moving machine: as you type, the capsule and rail glide along below your words, like a real typewriter. You can choose paper, ink, fonts, lamps, sounds, and desk scenes, then switch to Focus Mode for a wider page and a calmer view. It saves locally, works offline, and is completely free, no accounts or subscriptions.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ t21dev/turbo-reader
Turbo Reader: A Tiny RSS Reader That Actually Works
Turbo Reader is a desktop RSS reader with an installer under five megabytes that opens in under half a second and holds twenty thousand articles in a fourteen megabyte core process. Most feed readers open like a junk drawer, but Turbo Reader opens on what happened since you last looked, with counts for today, this week and this month, pinned feeds first, and a section per category. It ranks articles locally with no account and no API key, using freshness, what you read, which feeds you actually open, and your own interest and mute lists.
π @hackernewsgithubprojects
Turbo Reader: A Tiny RSS Reader That Actually Works
Turbo Reader is a desktop RSS reader with an installer under five megabytes that opens in under half a second and holds twenty thousand articles in a fourteen megabyte core process. Most feed readers open like a junk drawer, but Turbo Reader opens on what happened since you last looked, with counts for today, this week and this month, pinned feeds first, and a section per category. It ranks articles locally with no account and no API key, using freshness, what you read, which feeds you actually open, and your own interest and mute lists.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ tabulai/mlx-learn
mlxlearn
mlxlearn runs scikit-learn-style classical machine learning on Apple silicon with the Metal GPU, and it can make clustering and nearest-neighbor search dramatically faster. The idea is simple: normal machine-learning code stays familiar, but the heavy work moves onto a GPU built into an M-series Mac. It is useful for people who want quicker fits without rewriting their project or learning a new modeling language. The concrete win is visible in clustering: k-means++ seeding for 8,192 centers on MNIST takes about 5.4 seconds instead of 217 seconds, while still matching scikit-learn's chosen starting centers for the same random seed.
π @hackernewsgithubprojects
mlxlearn
mlxlearn runs scikit-learn-style classical machine learning on Apple silicon with the Metal GPU, and it can make clustering and nearest-neighbor search dramatically faster. The idea is simple: normal machine-learning code stays familiar, but the heavy work moves onto a GPU built into an M-series Mac. It is useful for people who want quicker fits without rewriting their project or learning a new modeling language. The concrete win is visible in clustering: k-means++ seeding for 8,192 centers on MNIST takes about 5.4 seconds instead of 217 seconds, while still matching scikit-learn's chosen starting centers for the same random seed.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ kevinpbuckley/vibe3d
Vibe3D
Vibe3D turns Unreal Engine's Modeling Mode into an API an AI agent can drive. Instead of clicking through a cluttered viewport, a connected agent can create a mesh, shape it, unwrap it, bake it, rig it, and save it as a static or skeletal mesh without ever touching a gizmo. It is useful because it lets a computer build real 3D assets the way a designer would, just faster and repeatable, with a built-in agent skill teaching the agent exactly how the whole workflow fits together. The cool part is that it feels like handing an AI the keys to a full 3D workshop inside Unreal.
π @hackernewsgithubprojects
Vibe3D
Vibe3D turns Unreal Engine's Modeling Mode into an API an AI agent can drive. Instead of clicking through a cluttered viewport, a connected agent can create a mesh, shape it, unwrap it, bake it, rig it, and save it as a static or skeletal mesh without ever touching a gizmo. It is useful because it lets a computer build real 3D assets the way a designer would, just faster and repeatable, with a built-in agent skill teaching the agent exactly how the whole workflow fits together. The cool part is that it feels like handing an AI the keys to a full 3D workshop inside Unreal.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ taucetiproject/taucetiroadmap
Tau Ceti Roadmap: The Map for Human-Checked Math
Tau Ceti Roadmap is a human-controlled map of math goals for Tau Ceti, an AI-welcoming Lean library built on Mathlib. Its one clear capability is turning big proof projects into tracked folders with status and progress notes, so anyone can see what is planned, done, or still open. That matters because machine-written math still needs human review, and this makes that review visible. Instead of guessing whether a library is finished, you can open the roadmap and read the trail. The takeaway: this is not just a list of math topics, it is a public checkpoint system for keeping formal proofs honest.
π @hackernewsgithubprojects
Tau Ceti Roadmap: The Map for Human-Checked Math
Tau Ceti Roadmap is a human-controlled map of math goals for Tau Ceti, an AI-welcoming Lean library built on Mathlib. Its one clear capability is turning big proof projects into tracked folders with status and progress notes, so anyone can see what is planned, done, or still open. That matters because machine-written math still needs human review, and this makes that review visible. Instead of guessing whether a library is finished, you can open the roadmap and read the trail. The takeaway: this is not just a list of math topics, it is a public checkpoint system for keeping formal proofs honest.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ dnotitia/dna-factory
DNA Factory
DNA Factory is the one-YAML toolkit that trains language models without turning your weekend into a config maze. It handles the usual post-training work β teaching a model from examples, tuning it on good versus bad answers, and rewarding what it does well β all with the same simple setup. The neat trick is that every method shares one training flow, so you write one config and change the method, instead of juggling five separate pipelines. It even checks the results while training is still running, so you can watch progress without stopping the show.
π @hackernewsgithubprojects
DNA Factory
DNA Factory is the one-YAML toolkit that trains language models without turning your weekend into a config maze. It handles the usual post-training work β teaching a model from examples, tuning it on good versus bad answers, and rewarding what it does well β all with the same simple setup. The neat trick is that every method shares one training flow, so you write one config and change the method, instead of juggling five separate pipelines. It even checks the results while training is still running, so you can watch progress without stopping the show.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ jasnell/node-skills
Node Skills: Teach Coding Agents New Node.js
Node Skills gives coding agents a cheat sheet for brand new Node.js features that they were never trained on. It covers the latest built-in modules, APIs and flags, changed defaults and removals since Node.js 24 LTS, so your assistant won't guess wrong or reach for an old npm package. Every skill is written from pinned official docs and verified by runnable examples, and it works in tools like Claude Code and opencode. Think of it as a tiny, trustworthy field guide for the newest JavaScript runtime. Install it as a plugin and your agent actually knows the present, not the past.
π @hackernewsgithubprojects
Node Skills: Teach Coding Agents New Node.js
Node Skills gives coding agents a cheat sheet for brand new Node.js features that they were never trained on. It covers the latest built-in modules, APIs and flags, changed defaults and removals since Node.js 24 LTS, so your assistant won't guess wrong or reach for an old npm package. Every skill is written from pinned official docs and verified by runnable examples, and it works in tools like Claude Code and opencode. Think of it as a tiny, trustworthy field guide for the newest JavaScript runtime. Install it as a plugin and your agent actually knows the present, not the past.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ ruoyiqiao/mjlab_hand
mjlab_hand: Teach Hands to Juggle
Watch a simulated hand pick up a ball and start juggling. mjlab_hand is a training playground for dexterous robot hands, where you can teach virtual fingers to grasp, twist, throw, spin a plate, or juggle three balls at once. It brings together several hand models, including Allegro, LEAP, Shadow, Sharpa, Wuji, and XHand, so you can test the same moves across different designs. The cool part is that it is built to learn through trial and error, rewarding the hand for making new contact with objects instead of just hitting the target. If you are into robotics, you will see how messy physics turns into a neat skill.
π @hackernewsgithubprojects
mjlab_hand: Teach Hands to Juggle
Watch a simulated hand pick up a ball and start juggling. mjlab_hand is a training playground for dexterous robot hands, where you can teach virtual fingers to grasp, twist, throw, spin a plate, or juggle three balls at once. It brings together several hand models, including Allegro, LEAP, Shadow, Sharpa, Wuji, and XHand, so you can test the same moves across different designs. The cool part is that it is built to learn through trial and error, rewarding the hand for making new contact with objects instead of just hitting the target. If you are into robotics, you will see how messy physics turns into a neat skill.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ illumina/cellularai
CellularAI: Predicting Single-Cell Responses to Genetic Perturbations
CellularAI predicts how cells respond to genetic changes, turning a set of control cells in a new setting into realistic single cells for each perturbation. It takes control cells and a list of genetic perturbations, estimates how those cells react, and generates integer-count cells for every target. The method uses precomputed training summaries from the Illumina Billion Cell Atlas and public CRISPRi Perturb-seq datasets, then pools effects weighted by how similar training controls are to the new context, decides which genes respond, refines those effects, and synthesizes the cells.
π @hackernewsgithubprojects
CellularAI: Predicting Single-Cell Responses to Genetic Perturbations
CellularAI predicts how cells respond to genetic changes, turning a set of control cells in a new setting into realistic single cells for each perturbation. It takes control cells and a list of genetic perturbations, estimates how those cells react, and generates integer-count cells for every target. The method uses precomputed training summaries from the Illumina Billion Cell Atlas and public CRISPRi Perturb-seq datasets, then pools effects weighted by how similar training controls are to the new context, decides which genes respond, refines those effects, and synthesizes the cells.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ pierridotite/mobile-agent-testers
Mobile Agent Testers
Mobile Agent Testers sends 20 AI agents to test an app on 20 different phones in parallel, for free. Each agent plays a persona like an iPhone SE in Swedish or a Pixel 7 at 200 percent text, uses the app like a new user, and writes a bug report with screenshots. In one real run, 20 agents found 46 distinct bugs in a popular Flutter app, including a message dialog that overflowed by 149 pixels and hid the confirm button. No test scripts to write. It works with Flutter, React Native, Expo, and more, and the reports are encrypted so the logs stay private.
π @hackernewsgithubprojects
Mobile Agent Testers
Mobile Agent Testers sends 20 AI agents to test an app on 20 different phones in parallel, for free. Each agent plays a persona like an iPhone SE in Swedish or a Pixel 7 at 200 percent text, uses the app like a new user, and writes a bug report with screenshots. In one real run, 20 agents found 46 distinct bugs in a popular Flutter app, including a message dialog that overflowed by 149 pixels and hid the confirm button. No test scripts to write. It works with Flutter, React Native, Expo, and more, and the reports are encrypted so the logs stay private.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ souldewapp/souldew
souldew
souldew turns PokΓ©mon SoulLinks into a live, no-spreadsheet session for two to four players on Mac, Windows, Linux or Steam Deck. Instead of shouting catches into a chat, you play your own legally obtained games in a built-in emulator while the app reads your team from memory and syncs everyone in real time over encrypted connections. A SoulLink means what you catch is tied to what your friends catch, and if one faints, the group goes down together. The cool part is the bookkeeping disappears: catches, deaths, routes and wipes show up automatically, with solo Nuzlocke mode, randomizers, ROM-hack patching, fast-forward with normal music, and OBS overlays for streamers.
π @hackernewsgithubprojects
souldew
souldew turns PokΓ©mon SoulLinks into a live, no-spreadsheet session for two to four players on Mac, Windows, Linux or Steam Deck. Instead of shouting catches into a chat, you play your own legally obtained games in a built-in emulator while the app reads your team from memory and syncs everyone in real time over encrypted connections. A SoulLink means what you catch is tied to what your friends catch, and if one faints, the group goes down together. The cool part is the bookkeeping disappears: catches, deaths, routes and wipes show up automatically, with solo Nuzlocke mode, randomizers, ROM-hack patching, fast-forward with normal music, and OBS overlays for streamers.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ mitaku/andon
Andon: One Person, A Whole Software Factory
Andon is the one-person software factory that keeps AI coding agents moving from idea to merge without dragging you in for every tiny step. Inside a single container cell, a coordinator agent reads a plain-file issue ledger, launches one worker agent per issue in its own git worktree, hands the pull request to a reviewer from a different model family, and lets a rule-based gate decide what can merge on its own. You only get called in when it is truly your call: what to build, how users should feel, promises to the outside world, irreversible risks, or changes to the line itself.
π @hackernewsgithubprojects
Andon: One Person, A Whole Software Factory
Andon is the one-person software factory that keeps AI coding agents moving from idea to merge without dragging you in for every tiny step. Inside a single container cell, a coordinator agent reads a plain-file issue ledger, launches one worker agent per issue in its own git worktree, hands the pull request to a reviewer from a different model family, and lets a rule-based gate decide what can merge on its own. You only get called in when it is truly your call: what to build, how users should feel, promises to the outside world, irreversible risks, or changes to the line itself.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ sohzm/winedroid
WineDroid: Android Apps on Your Desktop
WineDroid runs Android apps on web, linux, macOS and windows, like Wine for Android. Built in Rust, it executes DEX bytecode, implements Android framework behavior, and renders interfaces with Skia. Desktop builds use a Cranelift JIT, while the browser version uses the interpreter. Right now, apps like Weather Overview and Candle run with Gold and Silver ratings, though compatibility is still early and missing pieces are expected. It is an experimental compatibility layer, not a full virtual machine, and it shows missing APIs clearly instead of pretending they work. If you want to see how far it can go, try a browser demo or test a few APKs yourself.
π @hackernewsgithubprojects
WineDroid: Android Apps on Your Desktop
WineDroid runs Android apps on web, linux, macOS and windows, like Wine for Android. Built in Rust, it executes DEX bytecode, implements Android framework behavior, and renders interfaces with Skia. Desktop builds use a Cranelift JIT, while the browser version uses the interpreter. Right now, apps like Weather Overview and Candle run with Gold and Silver ratings, though compatibility is still early and missing pieces are expected. It is an experimental compatibility layer, not a full virtual machine, and it shows missing APIs clearly instead of pretending they work. If you want to see how far it can go, try a browser demo or test a few APKs yourself.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ alesha-pro/maya-inkling-mini
Maya Inkling Mini
Maya Inkling Mini lets you run a huge, capable AI model on just one gaming GPU, giving you fast private chat with images and tool use in about 128 tokens per second. It works by keeping the most-used parts of the model in your graphics memory and swapping the rest from RAM or disk as needed. That means you can talk to a massive model on your own machine without sending data anywhere, all while it stays quick and responsive. You get a local, private assistant that feels fast, private, and practical, and it makes a big model feel like it belongs on your desk.
π @hackernewsgithubprojects
Maya Inkling Mini
Maya Inkling Mini lets you run a huge, capable AI model on just one gaming GPU, giving you fast private chat with images and tool use in about 128 tokens per second. It works by keeping the most-used parts of the model in your graphics memory and swapping the rest from RAM or disk as needed. That means you can talk to a massive model on your own machine without sending data anywhere, all while it stays quick and responsive. You get a local, private assistant that feels fast, private, and practical, and it makes a big model feel like it belongs on your desk.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ wrong-nebula/hypbr
hypbr turns 3D textures into seamless materials
hypbr turns plain textures into perfectly tileable 3D materials, and it does it fast. In one command, it builds a full set of physical maps for metal, stone, or fabric that wrap around a surface without visible seams. It is useful for game art, 3D scenes, and quick previews because you can tune weathering, blend two finishes, and export for tools like Blender, Unreal, or Unity in minutes. The main idea is simple: real materials should look convincing when repeated, not only when viewed once. If you want believable surfaces without hand-painting every repeat, hypbr is the tool to try.
π @hackernewsgithubprojects
hypbr turns 3D textures into seamless materials
hypbr turns plain textures into perfectly tileable 3D materials, and it does it fast. In one command, it builds a full set of physical maps for metal, stone, or fabric that wrap around a surface without visible seams. It is useful for game art, 3D scenes, and quick previews because you can tune weathering, blend two finishes, and export for tools like Blender, Unreal, or Unity in minutes. The main idea is simple: real materials should look convincing when repeated, not only when viewed once. If you want believable surfaces without hand-painting every repeat, hypbr is the tool to try.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ sshkeda/minecraft-server-on-orange-cloud
Minecraft Server on Orange Cloud
Deploy a Minecraft server on Cloudflare with no VPS or port forwarding. Minecraft Server on Orange Cloud runs a Paper Java server on the orange cloud, so players join at a hostname on your domain. Spectrum takes the player's connection, a Worker passes it to a Durable Object, and that object starts the Paper container. After inactivity the server saves the world and sleeps; the next player wakes it in about 30 seconds. While it is awake, the container saves to R2 every five minutes. It is not free: roughly twenty five to thirty dollars a month plus container time while awake.
π @hackernewsgithubprojects
Minecraft Server on Orange Cloud
Deploy a Minecraft server on Cloudflare with no VPS or port forwarding. Minecraft Server on Orange Cloud runs a Paper Java server on the orange cloud, so players join at a hostname on your domain. Spectrum takes the player's connection, a Worker passes it to a Durable Object, and that object starts the Paper container. After inactivity the server saves the world and sleeps; the next player wakes it in about 30 seconds. While it is awake, the container saves to R2 every five minutes. It is not free: roughly twenty five to thirty dollars a month plus container time while awake.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ shinmegamiboson/lookatmyquant
LookAtMyQuant
LookAtMyQuant is the compression tool that shrinks large AI models without losing their smarts, turning a big model into a small one that stays fast on your computer. It measures exactly how much each piece of a model hurts the output when squeezed, then picks the best code for every piece so the whole thing fits a tight size budget with the least damage. On a Qwen 3.5 4 billion parameter model, it reaches 1.23 gigabytes at 2.3 bits per weight while still solving most math problems, and a small helper model can make it generate up to twice as fast on Apple silicon.
π° https://news.ycombinator.com/item?id=50591853
π @hackernewsgithubprojects
LookAtMyQuant
LookAtMyQuant is the compression tool that shrinks large AI models without losing their smarts, turning a big model into a small one that stays fast on your computer. It measures exactly how much each piece of a model hurts the output when squeezed, then picks the best code for every piece so the whole thing fits a tight size budget with the least damage. On a Qwen 3.5 4 billion parameter model, it reaches 1.23 gigabytes at 2.3 bits per weight while still solving most math problems, and a small helper model can make it generate up to twice as fast on Apple silicon.
π° https://news.ycombinator.com/item?id=50591853
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ 5m0k3/cache-maxxing
cache-maxxing
cache-maxxing helps you protect your Claude subscription by showing a live bar above your chat that tells you exactly when your prompt cache expires and what your next message will cost. It works locally, sends nothing anywhere, and warns you in plain colors before you send a message into an expired cache. Instead of guessing, you can see a countdown, a cost estimate, and a one-line tip. It even offers a handoff command that summarizes your progress into a fresh chat, saving your quota. The bar shows in the desktop app, VS Code, and terminal, and it keeps your flow by not blocking your messages.
π @hackernewsgithubprojects
cache-maxxing
cache-maxxing helps you protect your Claude subscription by showing a live bar above your chat that tells you exactly when your prompt cache expires and what your next message will cost. It works locally, sends nothing anywhere, and warns you in plain colors before you send a message into an expired cache. Instead of guessing, you can see a countdown, a cost estimate, and a one-line tip. It even offers a handoff command that summarizes your progress into a fresh chat, saving your quota. The bar shows in the desktop app, VS Code, and terminal, and it keeps your flow by not blocking your messages.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ fkubota/understand-mode
Understand Mode: The Post-it Board for Code
Understand Mode turns Claude Code's output, pull requests, and code into a post-it board you can wander through, not a dense wall of text. Instead of forcing you to read everything at once, it maps the big picture and lets you click through details as you go, like taking a shortcut that actually explains itself. You get the overall shape first, then the tangled parts are waiting to be poked at whenever you need them, which is a lot less boring than staring at a raw diff.
π @hackernewsgithubprojects
Understand Mode: The Post-it Board for Code
Understand Mode turns Claude Code's output, pull requests, and code into a post-it board you can wander through, not a dense wall of text. Instead of forcing you to read everything at once, it maps the big picture and lets you click through details as you go, like taking a shortcut that actually explains itself. You get the overall shape first, then the tangled parts are waiting to be poked at whenever you need them, which is a lot less boring than staring at a raw diff.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ m1stermo/gta2-browser
Grand Theft Auto 2 in Your Browser
Grand Theft Auto 2 in Your Browser β a 1999 game now plays right in a web tab. No installs, no plugins β just click play and steal a Z-Type in Downtown Anywhere City. The engine is a real decompilation of the original code, compiled to WebAssembly, so missions, gangs, cops, radio and all three districts work exactly like they did in 1999. On top of that, it adds widescreen with no edge pop-in, a calmer camera, a real map that fills in as you explore, a radar, smooth 60-to-165 Hz frame rates, and up to six players in co-op, Gang Wars and free-for-all.
π @hackernewsgithubprojects
Grand Theft Auto 2 in Your Browser
Grand Theft Auto 2 in Your Browser β a 1999 game now plays right in a web tab. No installs, no plugins β just click play and steal a Z-Type in Downtown Anywhere City. The engine is a real decompilation of the original code, compiled to WebAssembly, so missions, gangs, cops, radio and all three districts work exactly like they did in 1999. On top of that, it adds widescreen with no edge pop-in, a calmer camera, a real map that fills in as you explore, a radar, smooth 60-to-165 Hz frame rates, and up to six players in co-op, Gang Wars and free-for-all.
π @hackernewsgithubprojects