GitHub Trending Today #55: free-my-tv, pele, SPACE, ascii-sharks, harp, DeepInfer, ElyOffice, fx0 Transcript

00:00 Welcome back to GitHub awesome. This is GitHub trending today number 55. 35 trending open source projects on GitHub right now. Let's get into it. Free my TV is an ad-ree home screen for Google TV and Android TV built as a plain Android app of about 2 megabytes. Your streaming apps sit on a glass style shelf. The home screen shows three trending picks each hour taken only from apps you already have along with an hourly forecast and a big clock screen saver.

00:31 The theme switches between day and night at your city's sunset. Pelle is a small library from the Obsidian team that turns mermaid diagram text into SVG with no runtime dependencies. It covers every diagram type in Mermaid 12.1, but it uses its own layout engine and visual style. Call render and you get the SVG back right away with no waiting. Colors and fonts come from CSS variables, so a theme change updates without redrawing anything.

01:02 Space is Perplexity's open-source platform for running agent code safely written in Rust on Firecracker MicroVMs. Your app makes an HTTP call to create a sandbox from a template, and each one gets its own VM and file system. From there, you run commands, stream the output back, and pause, suspend, or wake the sandbox whenever you need it. It works with the E2B Python SDK, so existing code can move over.

01:25 Ashie Sharks puts swimming sharks and fish on a browser canvas drawn entirely with text characters. You choose the shark species and sizes, then let it TypeScript engine handle animation and resizing. Clicking creates a splash. There are no runtime dependencies and it pauses when the canvas leaves the screen. It also respects reduced motion settings by showing a still frame.

01:50 HARP is an unofficial fork of the open source warp terminal made for running a lot of clawed code and codeex agents at once. It swaps the tab bar for a sidebar where each row shows whether that agent is working, waiting for you, or done. You can sort tabs into named groups, and with T-Mox turned on, they survive a restart. A footer shows your 5-hour and weekly usage limits as rings.

02:17 Eagle 4 is the newest version of the Eagle speculative decoding method for speeding up large language models. A small draft model guesses the next several tokens, and the big model checks them all at once, so the output stays the same as before. It plugs into SG lang and VLM through engine hooks and sizes its guest tree from measured costs on your hardware.

02:41 Ask Magic Art lets you browse animated text art right in your terminal. Then copy a prompt that helps your coding agent add it to a website. You can switch the artwork into dots, binary, or your own characters and embed it with a web component. Each piece is also an editable JavaScript module. Web animations pause offcreen and show a still frame when visitors prefer reduced motion.

03:06 Elie Office wraps established document engines in desktop apps built with Rust and GPUi. Word processing, spreadsheets, and presentations use Libri Office Kit, while its PDF app adds editing and OCR through MUP PDF and Tesseract. Your files stay in ordinary folders with no account or cloud service required. There are markdown and git apps, too. Deep Infer is Deepseek's open- source inference framework, the same one that runs its production services built for Deepseek v4.1 Flash.

03:34 It started out built on VLM and runs on Deepseek's own kernels like Flash MLA, Deep Gem, and DPP. It supports expert parallelism for mixture of experts models, speculative decoding, and structured JSON output on both Nvidia and Huawei Ascend chips. Zeroach is a readonly scanner that checks your computer for traces of the adfection spyware campaign which researchers have tied to Candiru.

04:01 It's plain Python with no dependencies and it blocks its own network access while it runs. It compares your browser history, caches, and certain files against 338 published indicators. Every finding gets graded by how strong the evidence is. Editani is a skill for clawed code and codecs that edits video locally when you describe what you want. Open your agent in a folder of footage and explain the edit.

04:30 It transcribes the speech on your own machine, picks complete takes, and checks each cut against the waveform. It also removes dead air, breaths, and repeated lines. You review a short sample before it renders the final MP4 with hyperframes. pairs an AI tutor with a pen and paper you're already using. Describe what you want to learn, work through its exercises, then submit your handwritten work for feedback through a camera or photo upload.

04:58 It saves progress and reflections locally to help adapt later lessons. The app runs on your computer, but submitted work and relevant learning context go to OpenAI for tutoring. TV Detox speeds up an old slow Android TV by having clawed code or codecs clean it up over Wi-Fi. Paste its prompt into your agent and it lists the TV's apps and asks before switching anything off.

05:24 Nothing gets deleted and it keeps an undo list as it goes. Core parts like Google Play services and settings are on a protected list it never touches. Then it installs a tiny home screen that uses about 30 MGB of RAM. FX0 is Versell Lab's open- source design agent built on its lib FX agent library. Describe a page in chat and the agent writes it as one self-contained HTML file.

05:50 Each design gets version history so you can copy or download any earlier version. The agent can search the web for references and work in one of 10 bundled styles or one you upload. What stands out is that every step is saved as it runs. Panda Code puts the coding agents you already use into one desktop workspace alongside git diffs, a terminal, and a browser.

06:13 Choose a different agent and model per task. Track work on a board and review each turn's changes before committing. Separate Git work trees let tasks use their own checkouts. Conversations stay local while prompts go to your chosen provider. You still need its CLI and account. Scroll lets you show an AI agent a problem instead of typing it out. Built in Rust with Swift helpers for Mac OS.

06:40 Press a shortcut, then talk and draw red marks on your screen while it records. Press it again and scroll transcribes your voice on the device and picks key frames from moments like here or a click. It packs a contact sheet, full frames, and a timeline transcript into one bundle for your agent. and compiles mermaid diagrams into SVG before your page reaches the browser.

07:04 Its Rust compiler plugs into Vite or Markdown builds so visitors receive finished diagrams without downloading the parser and layout engine. You can also render files from the command line or generate SVG on a server. Colors follow CSS variables, so the same compiled diagram can switch between light and dark themes without another build. The trader skill is a set of 11 clawed skills for working with public market data built as plain Python with no dependencies.

07:32 One reads an investor's quarterly SEC filings and shows what changed in their portfolio. Others track news, congressional trade disclosures and earnings, size positions to your budget, and check each order against your risk limits. Rajget is an experiment that hot reloads Rust code in milliseconds using Rust Analyzer as the front end and Cranalift for the machine code with no LLVM and no linker.

07:59 When you save an edit, it recompiles only the functions that actually changed and swaps them in live and your program state stays where it was. The demo is Conway's game of life where changing the rules shows up on the very next frame. Twin links two computers on the same network, so they work like one desk across Mac, Windows, and Linux. It's built on Electron with a small native helper for each system, and WebRTC for the screen video.

08:27 One mouse and keyboard move across both machines. You can extend or mirror a screen, share the clipboard, drag files from one to the other, and open the other computer's apps in their own windows. Superite is a workflow plugin for claude code that takes a spec and comes back with finished tested work built on the ideas of the superpowers plugin. It plans the build and splits it into tasks that run in parallel waves with up to 18 agents at once.

08:55 Simple work goes to Haiku implementation to Sonnet and the final review to Opus. A fresh reviewer then reports problems without editing anything. Gram is a free writing assistant for Chrome that checks spelling and grammar as you type. English and German checks run inside your browser with bundled dictionaries, so you don't need an account or an API key.

09:18 It underlines problems where you're writing, and you click to accept or dismiss each fix. It also catches mixed up words that are spelled correctly but used wrong. Ruvio is a key value server written in Rust that speaks the Reddus protocol. So Reus colli and your existing Reddus clients connect without changes. Each key belongs to one shard which handles its requests without locks or thread switches.

09:45 Every right goes to a log before it's applied. Then come the built-in commands like leases with fencing tokens, semaphors, capped counters, versioned rights, and delayed Q entries. MEMT 1.0 0 is a memory model from memensor that builds lasting memory right into the model built on Quen 3.6 and Deepseek V4 flash. A memory block sits next to the model's attention and keeps a memory state that updates after every turn.

10:14 On the next turn, the model reads from that state instead of rereading the whole conversation. The memory stays the same size no matter how long the history gets. Dumb TV strips a supported Fire TV or Android TV back to your apps and HDMI inputs using ADB without root. Its own launcher has no internet permission, while cleanup scripts disable tracking packages and unwanted services.

10:38 It snapshots your settings first, so a restore script can put them back. Support varies by model, and streaming apps you choose to uninstall need reinstalling from the store. Capitza is a fast engine for running Quen and Neotron models locally on Apple Silicon Max. It reads your prompt on both the GPU and the neural engine. Then it writes the answer with speculative decoding where a small draft guesses ahead and the main model checks the guesses.

11:08 One command starts a chat page plus open AI and anthropic compatible APIs and it can launch clawed code or codecs against your local model. PaperLite turns AI research papers into coding exercises you can run on your own CPU. Read the method, fill in its core Python or PyTorch functions, then check your implementation against mathematical properties and gradient tests.

11:33 Interactive labs show how your changes affect the results with deliberately buggy examples to explain common mistakes. SP32 Thinks turns a small ESP32S3 gadget, the Sensecap Watcher with a round screen and a knob into a voice assistant built on Meta's open-source Muse Gadget SDK. Say, "Hi, ESP" or press the knob and ask a question. The device trims your recording, sends it off for transcription, and streams the reply back as speech, one sentence at a time.

12:04 Speechtoext, chat, and text to speech can each come from a different provider. KivyDB is a Reddus compatible in-memory database written in Rust. So your existing Reddus clients connect without changes. It splits data across shards, saves snapshots and an optional appendon log and uses IO uring on Linux when the kernel allows it. It also includes Reddus stack style search with full text tag, numeric and vector indexes.

12:30 Clustering, replication, and access control come built in. Sphinx pauses a git push to quiz you on what your code changes actually do. A separate reviewer model reads the committed diff without seeing your coding conversation, then asks about behavior and failure cases. Miss something and it explains the issue before giving you another scenario. You can answer with the code open using your own model provider.

12:57 Omiccast is a raycast style app launcher for Omari, the Linux setup built on Hyperland. It runs as a plugin inside Omar's own shell with no electron and no extra background services. Press alt and space and one list searches your apps about 400 Omar menu actions and the web results get ranked by how often and how recently you use things and it learns which words you type for each one.

13:23 Nanoi is a framework for building desktop apps in Zigg. Each app either draws its own native UI on the GPU or shows a web page in the system web view. It ports the core of Egoist's MIGO and borrows ideas from Zed's GPUi like chain styling and a virtualized list. The native views rebuild from your app state on each event. On Mac OS, they're drawn with metal and an idle app uses no CPU at all.

13:47 NSH lets you write interfaces in Rust, TSX, or both, and renders them through each platform's own UI system. TSX screens can reuse your Rust components with hot updates that preserve application state. You can embed it inside an existing native app or package an app on its own, including a static website. Hosts can also load signed TSX packages at runtime with explicitly configured sandbox capabilities.

14:15 Drag shift is a Windows utility for converting and editing files in the middle of a drag. Start dragging a file. Press shift and a radial menu pops up. Drop the file on a slice to convert it like an image to PDF or a PDF to Word. Press shift again for tools such as compress, resize, crop, redact, or strip GPS data. The result is saved next to the original, which stays untouched.

14:39 Solid CV is a computer vision library written entirely in Rust aiming to replace Open CV without the C++ build chain. It covers the familiar pieces like filters, edge detection, or features, stereo matching, and optical flow with kernels tuned for Apple silicon and Nvidia GPUs. It also adds tools for vision language models like skipping image patches that haven't changed between frames.