GitHub Trending Today #59: raea, exact, phonebox, cracked, line-art-animation, cleanyourtv, sakutto Transcript, AI Summary & Key Points

00:00 Welcome back to GitHub awesome. This is GitHub trending today number 59. 35 trending open- source projects on GitHub right now. Let's get into it. Modern web guidance is a collection of coding agent skills maintained by browser teams and the web community. It gives your agent targeted guides for modern browser features with compatibility data and advice on accessibility and performance.

00:24 that covers things like native dialogues, container queries, and view transitions. Its command line tool can also scan your codebase for older patterns and find guidance for updating them. Rhea is a rendering engine built with Rust and Bevy that turns React and TypeScript into videos, motion graphics, and interactive 3D scenes. You write a composition as code, and it renders on your own GPU.

00:51 It also gives coding agents tools to inspect individual frames and check the output for problems like clip text or low contrast. The same composition can render to video or run live. Exact is a cross-platform app runtime from the expo team with a Rust compiler and native hosts. You describe the interface in contract, a small declarative language, and write data logic in Typescript or Rust.

01:15 The compiler turns that into a plan that runs on the web, Mac OS, iOS, and Linux with no web view, and no JavaScript bridge in the UI. It's built so agents can build, run, and test apps, too. Phone box turns an old Android phone into a desk panel for Windows, Linux, or Mac OS. An app on the phone connects to your computer over USB or paired Wi-Fi, and you arrange widgets in a visual editor.

01:43 It can show system usage, control music, and track clawed or codec sessions. The phone also buzzes when a coding session finishes or needs your attention. Cracked is a collection of agent skills that turns your coding assistant into an interview coach. It runs mock interviews tailored to a company, executes your code against hidden tests, and gives feedback on how you explain your approach.

02:07 You can also practice system design or get your resume reviewed. Each session saves a scorecard locally so your next prep plan can focus on the weaknesses you've actually shown. Line art animation is a clawed skill and TypeScript library for building animated SVG scenes with an ink on paper look. It includes walking motion, wind, and a pallet that changes from day to night.

02:33 Characters react to clicks and scenes respond to your pointer. The walking system calculates where feet should land so they stay planted through each step. You can also capture the animation frame by frame for video. Clean your TV is an Android phone app that helps you tidy up an Android TV over Wi-Fi. It connects through network debugging, scans the installed apps, and shows you a cleanup plan before making changes.

02:58 On supported models, it can also install a simpler home screen. It creates a restore point first, so you can bring back an individual app or undo the whole cleanup from your phone. Super Mimic is a motion imitation training project from Hybrid Robotics built on MJ Lab. It implements Super Track to train world models and control policies from motion data.

03:21 The included G1 tasks let you work with either generalized or maximal coordinates. You can train a model on a selected GPU, then use the playback script to inspect the motion produced by a saved training run. Chrome capture is a command line tool and library that records Chrome tabs directly to AV1MPP4, including tab audio. It uses Chrome's native recording API to encode frames as the page runs.

03:49 You can record a URL or integrate it with Playright and Puppeteer to capture specific interactions. It requires Chrome 153 or newer and its Playright fixture can attach recordings directly to your test reports. Son Realtime is a family of voice models that can listen and speak at the same time. It uses tone and emphasis alongside your words and can add laughter or size to its replies.

04:13 When you interrupt, it stops and keeps only what you've actually heard in the conversation history. Its tag and flash versions run through VLM Omni on a single GPU with 80 GB of memory. Secudo is a Chrome and Firefox extension that translates web pages using a language model running on your device. It runs PLMO 3 translate through double ALMA with GPU or CPU processing.

04:38 You can translate a whole page, select a passage, or preview a translation of something you're typing. A bilingual view keeps the original text alongside the translation and your page text stays on your computer. Fly is an inference engine for Jev decision models built for workloads like email triage. It processes shared context once then forks the model state for each question.

05:04 Each branch gives the same answer the model would give if you ask that question on its own. In the maintainers demo, one H200 answered eight questions per email across 10,000 emails in 28 seconds. RAPVM gives AI agents an XFCE desktop inside a virtual machine. It comes with a persistent Chrome session that agent browser can attach to, plus agent desktop for interacting with Linux apps through the accessibility API.

05:32 WAPVM creates a separate machine for each project directory and you can open the desktop viewer in your own browser to watch what the agent is doing. ZTD charts is a React library for animated SVG charts. Each chart comes with keyboard navigation and a data table for screen readers. You can highlight one series and put that the takeaway right below the chart so readers know what to look for.

05:58 The same chart can be defined in JSON which lets a coding agent validate it and render a preview through the MCP server. Website bench is a research benchmark from Tiger AI lab that measures how well AI agents understand a website just by using it. An agent explores a site in a real browser, then has to build a full stack version that looks and behaves the same way.

06:23 Tasks run in Docker through the Harbor Framework and the results feed a public leaderboard with full agent trajectories published on hugging face. Image to sculpt is an agent skill that turns a single image into a 3D sculpt built in Python with Blender for the finish. It reads depth normals and an outline from the photo, generates side views, and fits a watertight surface that's checked against every angle.

06:49 Each build step is recorded so Blender can replay the whole sculpt as a time lapse. Icopy is a Windows tray app that downloads every original from your iCloud photos library to a local folder or external drive. Built in Rust with a small Python bridge, it checks each file's size, structure, and SHA 256 hash and picks up where it left off after any interruption.

07:14 It can even free up space in iCloud, but only deletes items that are already copied, verified, and confirmed by you. Yummy is a pixel art desktop environment for the Nerie Wayland compositor. It brings its own bar, launcher, lock screen, notifications, and doc, plus a file manager, image viewer, and media player, all sharing one accent pallet. Desktop widgets show the clock, calendar, and live CPU and network graphs, and a privacy curtain hides your screen while streaming.

07:47 Tailet do gives Cloudflare workers away onto your private Tailscale network. It runs Tailscale's own user space client compiled from Go to Web Assembly inside a durable object. So that object shows up on your tailet with its own address like any other device. Your workers can then call internal services, reach machines behind your firewall, or serve pages only your own devices can see all over RPC.

08:13 T80 humanizer is an agent skill that rewrites stiff AI generated text so people can read it easily. It works in two passes. First, it strips 26 common AI writing tells Wikipedia's catalog. Then it rebuilds sentences most of the way towards simplified technical English, the controlled style used in aerospace maintenance manuals. It can match your own voice from a sample and flags anything that changes the facts.

08:41 Dalib is a lightweight desktop ebook manager that reads and writes caliber libraries built in Go on the native myo framework with no web view. It uses caliber's own database and folder layout so your existing library opens as it is. A getbook sidebar pulls free titles from Project Gutenberg, and AI agents can search the library, edit metadata, and convert formats through a built-in MCP server.

09:08 Co-work OS is a workspace for people and AI agents to carry projects across multiple sessions. A main agent plans the work and delegates tasks, while the runtime tracks progress and keeps execution records. It connects with document editors, CAD and Blender, so you can inspect and revise the same files the agents are working on. Longer jobs retain their status and logs, allowing work to continue after an interruption.

09:35 Cinematic skill teaches coding agents to build narrated films with Python and Skia. It moves one continuous camera across photos, maps, and text with animations timed to individual words in the narration. The workflow covers research, image licensing, and sound effects, then renders the finished video with captions. If you replace the voice recording, the film adjusts its timing to match, so narration changes carry through to the visuals.

10:03 Persona is a self-hosted super agent that ties all your devices together. You run a hub on an always machine and it reaches the computers where agents like Claude Code and Codeex do the work. From one chat, the assistant pulls in your context, breaks a job into steps, and hands it off while you watch agents live and approve actions from your phone. Terminal phone puts an interactive iPhone right inside your terminal in Ghosty, Kitty, or iTerm 2 and shares it with your AI agents.

10:33 Built in Swift, it streams an iOS simulator or a paired real iPhone at 30 frames a second, and you control it with your mouse and keyboard. Agents drive the same phone through a JSON command line or an MCP server, reading the screen as accessibility elements. Vinyl is a browserbased record player that simulates a turntable from the groove up. Instead of playing an audio file straight, it cuts the sound into a spiral groove, then reads it back through a simulated stylus, cantaliever, and tone arm.

11:05 It models platter inertia, friction, and arm kinematics, and adds surface hiss, crackle, dust, and warp. You can cut your own file into a record, pick a stylish shape, and watch it all render in WebGL. Num is a skill and bash tool that organizes AI coding agents into a small company running inside the herder terminal workspace. You talk to a chief agent in plain language.

11:30 It routes each request to a manager for the right project and managers spin up workers in separate git work trees so tasks run in parallel. It can pull work from linear, sentry or GitHub issues, and nothing gets merged or pushed without your approval. Lot map is a cloud code plugin that shows a live map of your agents work beside the chat. Tool calls send animated pulses to stations for files, memory, the browser, and other tools with errors marked in red.

12:03 Sub agents appear around their own station while a central ring tracks context usage as the session continues. Frequently used connections grow brighter and thicker, showing where the work went. Git manual turns a repository link into a setup guide for your computer. It reads the project's readme, dependency files, and release details, then lists the tools you need and commands for Windows, Mac OS, or Linux.

12:28 It also checks the project's stated support against your system and chip. To open a manual, replace github.com with gitmanual.com in the link. You can download the setup steps as a script, too. Recalc is a desktop spreadsheet you share with your AI agents. A cell can hold a value, a formula, a terminal command that reruns when its inputs change, or a clawed code or codeex agent working on a task.

12:52 Answers come back with citations, and anything you've verified stays locked so agents can't overwrite it. Everything stays local with no account or server, and boards export to Excel. Shoji bar 4 is a desktop shell for shogw written in Rust with material 3 expressive styling. It puts workspaces, media controls, and system status into floating sections across the top with a dock that appears when your pointer reaches the bottom edge.

13:21 It also includes quick settings and clipboard history. Colors come from your wallpaper or a chosen seed and animate together across the shell and window frames. Aimac turns your iPhone into a pointing controller for your Mac over Bluetooth. Aim the phone to move the cursor. Tap to click and hold to drag. A receiver runs on the Mac and translates your movements into cursor input.

13:46 The repository also includes a browser table tennis demo where aiming moves your paddle and swinging faster makes the shot harder. iPhone Mirror for Windows turns a PC into an AirPlay receiver, so your iPhone's screen shows up in a floating iPhone-shaped window with sound and nothing to install on the phone. It's written in Python with QT and FFmpeg, and it handles rotation, screenshots, and resizing.

14:12 Plug in over USB and you can control the phone from your PC with taps, swipes, typing, and the home button. Flash states is a caching system for serving hybrid language models with VLM. It saves the model's state at the exact token where an agent's turn ends, so the next turn can pick up there and spend less time reprocessing old context. In a benchmark replaying 48 concurrent coding agent sessions with GLM5 through flash, the team reports an 88% drop in P90 time to first token compared with VLM.

14:48 If you expose services through FRP, brute force bots eventually find your tunnels. FRP ban is a plugin for the FRP server that keeps one shared ban list across all your machines. Small clients watch your log files with regx rules, spot things like SSH brute forcing, and send ban requests over TLS to a central server, which then refuses those IPs at login and on every new connection.