GitHub Trending Weekly #55: owly, ai-job-search, tsrs, dev-null.sh, laya, reylo, openterminal, onkey Transcript, AI Summary & Key Points

3D Model Anything is an open-source iPhone LiDAR capture app, Three.js viewer, and agent skill for reconstructing rooms as editable Three.js scenes. The app records room geometry and photographs; the scan-to-threejs skill guides an AI coding agent through measurement, modeling, and matched-camera visual comparisons rather than performing a one-click conversion. The repository includes an editable office demo, browser views such as dollhouse, walking, overhead, wireframe, and knolling modes, and workflows for importing user captures. LiDAR-equipped iPhones or iPads are required for scanning, while non-LiDAR devices can import and review scans. The reusable source code and skill are MIT licensed; the example's images and textures have separate terms, and the repository does not distribute other rooms or original scan bundles.

AI Job Search is an open-source job-application framework that runs locally through Claude Code. It builds a candidate profile from documents or an interactive interview, searches supported job portals, evaluates postings against fit criteria, and uses a drafter-reviewer workflow to produce tailored CVs and cover letters in LaTeX. The workflow also supports interview preparation, application tracking, skill-gap analysis, follow-up drafts, PDF layout and ATS text-layer checks, and optional synchronization with Notion or Gmail. Its core commands include /setup, /scrape, /rank, /apply, /interview, and /outcome; the job-portal skills are extensible through generated or copied integrations. The repository is MIT-licensed and requires Claude Code plus local Python, Bun, and a LaTeX distribution for the standard workflow. It is an independent project and is not affiliated with Anthropic.

Bambu Bar is a native macOS menu-bar application for monitoring and controlling Bambu Lab 3D printers. It displays print progress, time remaining, stages, temperatures, filament and AMS spool status, errors, and printer information, with controls for pausing, resuming, stopping, changing print speed, and switching the chamber light. It supports multiple printers, desktop widgets, notifications, printer discovery, and access-code import from Bambu Studio or OrcaSlicer. The app communicates directly with printers on the local network using their TLS-protected MQTT service and UDP discovery, without an account, cloud service, analytics, or third-party Swift packages. It stores printer data locally and keeps access codes in the macOS keychain; print controls require LAN Only mode with Developer Mode enabled on the printer. The project is built with Swift and SwiftUI, supports macOS 26 Tahoe or later on Apple silicon or Intel, and is released under the MIT License.

BitKindle is a Windows x64 C compiler toolchain that bootstraps itself from a 3,584-byte machine-code seed. Its source tree includes custom A0/A1 assemblers, the transitional C0 compiler, a self-hosted C1 compiler written in C, a runtime library, and a custom PE linker. PowerShell 5.1 decodes the seed and coordinates staged builds; each stage generates the machine-code binaries, object files, and executables needed by later stages, culminating in self-compiled generations whose outputs are checked for byte identity. The toolchain supports C89 and many C99 features, produces Windows x64 PE executables and its own .c1o object format, and includes interfaces for multiple source files, Windows DLL imports, and console or GUI subsystems. The repository also documents example programs including a command-line program, a WinHTTP-based minicurl client, and a C-written Mario-style game. Building requires Windows x64, Windows PowerShell 5.1, a writable source directory, and approximately 5 GiB for the full test run. It does not require MSVC, GCC, Clang, the Windows SDK, an external linker, or downloaded dependencies. The project does not claim complete C99 conformance and does not target general Windows SDK headers, COFF object or library compatibility, or external-toolchain interoperability. The source code and documentation use the MIT License.

Bursar is a self-hosted operations and governance platform for shared Linux GPU servers. It preserves users' normal SSH workflow while a central controller provides node visibility, usage accounting, points and quotas, account mapping, access policies, and security-event logging. The system consists of a Go controller and policy engine, a Go node agent, a Vue administrator and user interface, and a PostgreSQL database. Node agents report GPU, CPU, memory, disk, process, session, service, and health data over authenticated HTTPS; the controller applies accounting and policies for resources such as GPU visibility, CPU and memory limits, disk use, exclusive access, SSH access, and remote process actions. It also supports registration and account management, provisioning, temporary users, delegated administrators, per-node pricing overrides, carryover and overdraft policies, audit events, TOTP two-factor authentication, encrypted backups, and optional controller high availability. Bursar is intended for teams that already operate Linux GPU hosts and want governance without replacing SSH. It is not a scheduler, Kubernetes operator, hosted control plane, or replacement for network segmentation, operating-system hardening, monitoring, and tested backups. The controller can be run with Docker Compose or from source, while the node agent runs directly on each compute host. Because it can execute privileged actions on compute nodes, the project recommends reviewing its security model, beginning with dry-run mode, and completing the go-live checklist before production use. It is licensed under Apache License 2.0.

dev-null.sh is a social network for developers. It uses GitHub accounts as identity, lets users follow developers and share commits, and presents posts and timelines as logs that can be accessed from the terminal. The project is open source and community-built. Its source repository is implemented for PHP 8.5 or later with Composer, Bun, and PostgreSQL 17 or later using pgvector and pg_trgm, and is licensed under the MIT License.

dquickts is a MIT-licensed fork of Fabrice Bellard's QuickJS embeddable JavaScript engine. It adds durable execution: running JavaScript can be suspended at safe points, including in the middle of a loop, with the heap, closures, pending jobs, modules, and suspended call stack serialized to a snapshot and later restored in another process. Full snapshots and incremental base-and-delta snapshots are supported, as are host calls that remain suspended until the host supplies a result. The engine runs TypeScript directly by erasing types during parsing and compiling runtime constructs such as enums, namespaces, parameter properties, and decorators to bytecode; it does not perform type checking. Its qjs executable includes a durable REPL whose functions, variables, objects, and closures survive restarts. The repository also provides Rust bindings derived from rquickjs, exposing suspension, snapshotting, durable host calls, and TypeScript evaluation. Snapshots require compatible engine builds and matching native registries, and cannot include several forms of host or shared state, including SharedArrayBuffer memory, unsupported opaque host objects, or code running on the C stack. The repository includes C and Rust APIs, command-line tools, tests, and documentation.

ELYTerminal is an open-source terminal application for macOS, Windows, and Linux, written in Rust with GPUI and Ely components. It uses alacritty_terminal for terminal emulation and arranges sessions on the left, the terminal in the center, and Git or Files on the right. It supports 256 colors and truecolor, shell integration for several shells including zsh, bash, fish, PowerShell, Git Bash, and WSL shells, scrollback search, links, notifications, prompt and command status reporting, and Git and file browsing panels. Releases provide platform-specific builds; the repository documents macOS, Windows, and Linux installation and build procedures. The project is licensed under MIT, with separately documented licenses for bundled Ghostty-derived code, themes, and fonts.

HatchZip is a Rust archive extractor with desktop and command-line interfaces that extracts archives in place while reclaiming disk space. It saves the archive index separately, extracts independent pieces from the end of ZIP, 7z, tar, and RAR archives, and truncates or deletes each completed portion so the freed space can hold subsequent output; split archives are consumed volume by volume, while tar.gz streams are processed from the front with filesystem hole punching where supported. Before changing anything, it checks the required space, preserves a configurable free-space margin, validates paths and archive integrity, and writes and flushes each output file before removing its corresponding archive data. Interrupted or crashed extractions can resume, and unfinished extractions can restore the remaining archive. It supports ZIP, 7z, tar, tar.gz, RAR volume sets, and split archives, with optional archive preservation through a normal extraction mode. HatchZip is distributed for Windows, macOS, and Linux, and is dual-licensed under Apache-2.0 or MIT; RAR support uses the separately licensed UnRAR source.

Hi3D v1.0 is an AI tool for refining a coarse 3D mesh into higher-resolution geometry conditioned on a reference image of the same object. Built on the Sparc3D representation and refinement architecture, it extracts sparse coordinates from the coarse geometry, conditions sparse flow with image features, progressively samples sparse latents, decodes them with a self-pruning sparse VAE, extracts a mesh, and revoxelizes that mesh to provide support for the next refinement pass. It supports nominal resolutions of 512, 1024, and 1536, corresponding to one, two, or three passes, and exposes Python and command-line inference interfaces that accept an image and coarse mesh and export a geometry-only mesh. The project provides inference code, model artifacts, and documentation, with model files hosted on Hugging Face. It requires Linux, Python 3.10, and an NVIDIA GPU with compute capability 8.0 or newer using a CUDA 12.0–12.5-compatible PyTorch environment; there is no CPU inference path. The repository does not include training code, the evaluation harness, or benchmark data, and the output does not include textures, materials, or vertex colors. It is distributed under the Hi3D Open License Agreement for Hi3D-owned materials, with separate Apache-2.0, MIT, and third-party terms applying to other components.

img2 Interior is an agent-operated interior reconstruction skill for Claude Code, Codex and OpenCode, powered by the img2 harness. Given a room reference image, the agent inspects the image, authors scene-specific editable geometry and physically based materials in Blender, renders and compares the result through an iterative visual-review workflow, then produces a packed editable Blender master, a runtime GLB with a Three.js browser viewer, and a separate portable GLB. The viewer can include camera views, wireframe mode, material or curtain controls, and downloads; the handoff also includes the scene specification, Python builder, source assumptions, resource notices, verification evidence and known limitations. It is distributed as the public npm package @img2threejs/img2interior and as a GitHub project. The project states that it is not a pretrained image-to-mesh model or a one-click fidelity guarantee: single-image dimensions, hidden geometry, camera properties and materials remain estimates, and lighting, interactions and performance vary by scene. Use is free, including commercial project work, under the source-available img2 Interior Free Use, Restricted Commercialization License 1.0; selling or rebranding the tool, bundling it into commercial software, or offering paid hosted/API access requires prior written permission.

Laya is an Apache-2.0 Python package and encoder-model decision engine developed by Convai Innovations. It accepts text or JSON state and evaluates typed questions in a single forward pass, returning categorical choices with probabilities and confidence, ordinal scores, or calibrated Boolean probabilities rather than generated text. Its stated uses include routing, ticket triage, prompt-injection detection, moderation, phishing detection, and churn-risk classification. Laya includes English, multilingual, and typed-decisions checkpoints, plus a router that selects a checkpoint using explicit settings, detected script, and best-effort language detection. The checkpoints use reinforcement learning against strictly proper scoring rules (RLCD); the project also provides fine-tuning and temperature-calibration workflows. The multilingual checkpoint is intended for more than 100 languages, while the repository cautions that the base models can perform near chance on some zero-shot typed-decision tasks and that large choice sets are constrained by the option-token budget.

Ming Voice is a local, real-time speech-to-speech voice assistant for AMD Radeon GPUs, with experimental NVIDIA support, running on Linux or Windows with WSL2. It uses inclusionAI's Ming-3.0-audio-tiny model, Silero VAD, Smart Turn v3 turn detection, speculative generation with rollback, echo-aware interruption handling, and streaming speech synthesis. The server communicates with a browser or client over WebSockets, accepts 16 kHz PCM audio, and returns 44.1 kHz audio; fused Triton and HIP-graph paths are used to reduce inference latency on supported GPUs. The assistant supports hands-free, push-to-talk, and wake-word operation, voice cloning from a short sample, conversation persistence, English and Chinese, and phone access over HTTPS on a home network. Optional tools provide web search, weather, calculation, timers, reminders, lists, memory, Home Assistant control, Jellyfin playback, and hand-off to a Hermes Agent. Risky smart-home actions require confirmation by default, and the project states that audio remains on the local machine while enabled search and weather tools send only the relevant query data to their external services. Ming Voice also exposes an OpenAI-compatible local API for text-to-speech, transcription, and chat with audio input and output. The repository provides installers for Linux and Windows, downloads the model separately, requires a GPU with at least 16 GB of VRAM, and is licensed under the MIT license.

Motion2Code is a skill for coding agents that rebuilds motion-design videos as clean, editable inline SVG and GSAP code. It uses video-analysis scripts to extract timed frames, hard cuts, palettes and loop information, then drives the result with a single seekable GSAP timeline. Its comparison and rendering tools generate frame-by-frame source/code diffs, check loop seams and deterministic seeking, and render proof videos or GIFs. For screen recordings of websites and apps, it can instead produce real interactive pages using GSAP ScrollTrigger. The skill supports Claude Code, Codex, Cursor, Copilot, Gemini and other agent environments; its scripts require Python 3.9+, ffmpeg and, for browser-based checks, Playwright. The repository is MIT-licensed.

new-term-radar is an open-source Python tool for finding emerging AI names—models, products, tools, and repositories—before they appear in conventional keyword tools. It reads signals from X posts, Hugging Face trending models and Spaces, GitHub repository growth, sitemaps, newly indexed sites, optional Google Search Console data, Reddit requests, payment-link changes, and revenue listings, then produces reports of candidate terms and their evidence. Its optional keyword-use-case pipeline extracts generic searches associated with each name, screens them with a model, estimates search demand and trends, checks page-one competition, and assigns decisions such as prioritize, watch, or skip. The tool also provides command-line reports, configurable daily and source-specific runs, scheduled GitHub Actions workflows, agent skills, and an MCP server with tools for recent leads, keyword tables, and radar runs. It uses Python 3.10 or newer and the standard library for its core runtime; external services such as AIsa and optionally GoAnyAPI supply paid signals and model or search-data calls. The repository is licensed under the MIT License.

Omnara is an open-source managed agent platform for running AI agents as a service on any compatible model and machine. It provides a REST API, TypeScript SDK, CLI, dashboard, and MCP server for launching agents, sending messages, streaming events, and handling approvals. The platform runs agent workers that build model context, call models, and execute tools, while PostgreSQL stores each agent's event log and durable state so workers can resume agents after crashes, restarts, or machine disconnections. Agents can use built-in tools, skills, MCP servers, custom tools, persistent memory, subagents, permissions, schedules, webhooks, and integrations with Slack, Discord, and GitHub. Machines connect through the Omnara daemon and may be local laptops, virtual machines, containers, or managed sandboxes; agents can run across multiple machines without inbound ports. Omnara Cloud is available at app.omnara.com, and the platform can be self-hosted with Docker Compose under the Apache License 2.0. Self-hosted deployments keep agent history in the operator's PostgreSQL database and can use any supported model provider or compatible self-hosted endpoint.

OnKey is a free, open-source singing trainer for macOS that provides live pitch feedback while the user sings along with music played in Spotify, Apple Music, or YouTube. After a first listen, it separates the lead vocal from the music, uses CREPE to track the singer's pitch, and uses the GAME note model to convert the melody into syllable-based note bars. On later plays it locates the song from its audio, tracks the user's voice live, displays the pitch against the notes, and produces a score with a line-by-line report. It runs locally on Apple Silicon Macs with macOS 14.2 or later, does not provide or download music, and stores learned songs and recordings on the Mac. OnKey is released under the MIT license.

OPAL (OPtimizer with an Agent in the Loop) is a PyTorch optimizer wrapper developed by Xingjian Bai and contributors. It turns a fixed optimizer into two editable programs: an updater that converts gradients into parameter updates, and an analysis program that reports training statistics. A coding agent, Claude Code by default, periodically reads the goal, losses, gradients, and analysis reports, then can rewrite learning rates and schedules, momentum, preconditioning, clipping, parameter-group update rules, and related optimizer logic while training continues. New programs are trialed for 20 steps and can be reverted on errors when a parameter-and-state snapshot is available; until a change is made, the wrapped run is bit-for-bit equivalent to the original optimizer run. OPAL supports CPU, single-GPU, and single-node DDP PyTorch training, with integrations for selected Hugging Face Trainer and Accelerate workflows, but does not support multi-node training, FSDP, or DeepSpeed. Agent runs require Claude Code and an authenticated account, while the keep backend provides an agent-free dry run. The project is distributed under the MIT license.

Open Math Harness is the open-source runtime of Open Math OS, an infrastructure platform for collaborative mathematical research. Developed in Rust, its Nucleus generation provides a capability-based microkernel, services, and agents that run open models with streaming replies, tool use, sub-agents, persistent project history, and Git-based project memory. Its Ultramath service coordinates multiple agents exploring different approaches, while proof search explores candidate proofs in parallel and accepts a proof only when Lean verifies it in a confined workspace. Agents can use files, shell, web and PDF tools, skills, MCP servers, and plugins; protected actions require human approval, and credentials remain with privileged providers rather than models or logs. The repository supports building the runtime and a self-contained capability-checking example on Linux with Rust and bubblewrap, but the full interactive runtime currently depends on the Open Math OS deployment environment and is not yet a standalone local setup. It is licensed under Apache-2.0.

Open Terminal is an open-source, AI-native trading terminal that runs locally and builds a workspace of live charts, tables, option chains, and other widgets from plain-English questions. It uses an agent with four tools—route search, queries, pandas-based transforms, and workspace replacement—to find data routes, execute recipes or connector endpoints, derive frames, and assemble a workspace whose widgets remain editable queries. Replies link claims to the widgets that display their source data, and the server checks reply numbers against the frames observed by the agent. The desktop application supports macOS, Windows, and Linux, while the same application can run in a browser through its local Node-based server. It includes OpenBB, Yahoo Finance, and optional Finnhub connectors; OpenBB providers are ported to TypeScript, and connectors expose searchable route descriptors that can be extended with additional data sources. The project supports Anthropic and OpenAI models through the Vercel AI SDK, stores configuration, workspaces, recipes, and cached frames locally, and does not require an account or hosted backend. It is MIT-licensed.

OpenVLHarness is an open-source framework for turning multimodal language models into visual agents without fine-tuning their backbone weights. The model selects tools and reads their evidence through nine task-level capabilities: visual grounding, OCR, depth estimation, camera-trajectory estimation, image zoom, text search, reverse image search, webpage visits, and a Python coding agent. Its multimodal memory keeps images, boxes, masks, depth maps, numerical data, and web references addressable across a question, allowing later tools or generated Python code to combine earlier results—for example, locating objects, estimating their spatial relationship, or searching for and verifying an unfamiliar object. The harness includes model presets, a configurable tool router, evaluation utilities, and support for local or API-served models; the main code is released under the MIT License, while some evaluation, visualization, model, service, and dataset components have separate licenses. The local Python coding executor runs with the agent host's permissions and network and filesystem access, so it should be isolated when handling untrusted workloads.

Owly is a Linux tool for transferring files between iPhones and Linux computers over Apple's AirDrop protocol, using supported non-Apple Wi-Fi hardware. It provides a command-line interface for sending, receiving, browsing nearby AirDrop receivers, checking hardware compatibility, and testing new cards; an optional desktop integration is available through Omdrop for Omarchy. Its owlyd system service takes control of the Wi-Fi card, runs Apple's AWDL protocol, and exposes the resulting awdl0 network interface to the user-level AirDrop application. Depending on the hardware, it uses shared mode while preserving the existing Wi-Fi connection, exclusive mode with Wi-Fi temporarily down, or active-monitor mode. The repository documents tested Wi-Fi chipsets, driver-specific behavior, channel requirements, configuration options, and known issues, and includes Rust crates for AWDL, AirDrop, and Linux radio support. Owly is distributed from its GitHub repository and through packages for Arch, NixOS, Debian/Ubuntu, and Fedora, as well as from source. Supported hardware varies by Wi-Fi card and driver; untested cards require explicit opt-in.

penmentor is an AI-assisted mathematics tutor for pen tablets and handwritten work in Xournal++. A keypress sends the page to a configured AI command-line agent such as Claude Code, Codex, or Antigravity; the agent marks the work on the page with colored annotations, circles the first mistake, ticks correct lines, and provides graduated hints rather than directly giving the answer. Users can select part of a problem for an interactive Manim animation, and the tool supports practice prompts such as find-the-error drills and technique identification. The project is designed for Linux on X11 and uses an Xournal++ plugin together with tools including rofi, slop, xdotool, and PyQt6. It is distributed under the MIT license.

Plume Launcher is a lightweight, ad-free home-screen launcher for Android TV and Google TV. It replaces the stock home screen with an apps row, Continue Watching items from Android TV's Watch Next data and supported app channels, TV input shortcuts, and optional weather, wallpapers, and screensavers. Users can rearrange or hide apps, assign press and hold actions to supported remote buttons, switch themes and icon styles, and lock the layout with a PIN. The launcher can intercept the Home button through an Android accessibility service, while holding Home still opens the Google TV dashboard. It is distributed as a sideloaded APK and supports Android 8.0 or newer; the repository also documents installation through Downloader, USB, or adb and provides Gradle build instructions. Plume reports no analytics, ads, account, or tracking, and is released as free software under GPL-3.0.

prompt2rig is an open-source browser application that turns a text motion prompt into a rigged 3D character performance. It sends the prompt to a video model on fal, uses MediaPipe for an immediate live preview, and uses a local GPU pipeline with GVHMR to recover 3D body motion and WiLoR to track wrists and fingers. The recovered motion is retargeted to Mixamo characters, with support for loading Mixamo FBX or GLB files, streaming chained performances, and exporting animated GLB files for Blender or Unity. The high-quality pass runs through a local worker built around GVHMR and produces SMPL rotations and a root path; WiLoR runs in parallel for hand and finger tracking, with MediaPipe Hands as a fallback. The repository also includes Blender rendering and off-screen browser-capture tools. The web application requires Node 18+ and a fal API key; the optional high-quality path is documented for Windows with an NVIDIA GPU, CUDA PyTorch, ffmpeg, and separately obtained SMPL-X and MANO model files. The project's own code is licensed under Apache-2.0, while several high-quality motion and hand-tracking components have educational, research, non-commercial, or other separate licensing restrictions.

Python is a programming language and the implementation environment maintained in the CPython repository. The videos describe its use for back-end applications, bots, scientific analysis, API scripts, file-operation primitives, database and vector-search result processing, coding-session dashboards, document-converter bindings, and compiled agent skills. CPython can be built on Unix-like systems and Windows, with optional profile-guided and link-time optimization; installable distributions and documentation are provided through python.org.

rec is a local inference engine developed by Recursive for running its Flint language models on Apple Silicon Macs. It starts a local model server and provides OpenAI- and Anthropic-compatible chat-completion APIs, allowing clients to connect through localhost; the default model is flint. It includes launch and execution commands for the pi and OpenCode harnesses, as well as commands for serving, downloading, updating, diagnosing, authenticating, and stopping the model server. The project is optimized for single-session usage and requires an Apple Silicon Mac with an M4 Pro, M5 Pro, or M5 Max, at least 48 GB of unified memory, approximately 30 GB of free disk space, and Node.js 22 with npm when using pi. It is distributed under the Apache 2.0 license.

Reylo is a TypeScript tool that imports .sql files as typed modules, using TypeScript content mappers to provide static types without generated code. It has the configured database prepare each query without executing it, then infers column and parameter types, nullability, cardinality, unions, and supported JSON shapes from the query and schema; imported queries expose methods such as .run(), .runSync(), .try(), .with(), .sql, .raw, and .source. It supports PostgreSQL, MySQL, and SQLite with user-provided drivers, and can be used with TypeScript 7.1 beta, Node 24 or later, and a running development database. Reylo is distributed under the MIT license.

rl-env-use is an open-source tool from Browser Use for building reinforcement-learning environments for arbitrary domains from real cases mined from the web by Browser Use agents. Given a one-line skill description, it finds source materials and later-published outcomes, computes gold answers in code, generates difficult tasks, and rejects tasks with leaked answers, memorized outcomes, or ambiguous prompts. It can hill-climb a model's system prompt using train/test splits, a noise floor, and paired confidence intervals, then export the resulting dataset for an RL trainer or serve it in code through an environment interface. The repository provides a command-line workflow, requires Python 3.11 or newer, can use Node.js 18 or newer for HTML reports, and is licensed under MIT.

An installable Claude Code skill for animating SVG drawings in React Native and SwiftUI. It guides an agent to use one clock for a scene, real pivots, delayed follow-through, persistent idle motion, smooth paths through key poses, seeded variation, and measured motion from reference video rather than swapping between static poses. The repository includes Motion engines for React Native and SwiftUI with loop and one-shot clocks and motion primitives such as flow, segment, wave, loop, bell, seeded values, and curves; live-face components for eyes and mouths; and scripts for video contact sheets, edge tracking, and color sampling. It also provides reusable templates and demo scenes for characters, plants, cooking, and depth-based card orbits, with an offline macOS renderer that outputs MP4 or GIF files.

Rubix is an experimental systems programming language whose compiler is written entirely in GNU x86-64 assembly. It compiles `.ix` source files through a lexer, recursive-descent/Pratt-style parser, symbol table, and native code generator that emits x86-64 assembly, which is assembled and linked into standalone executables. The language includes typed integers, booleans, characters and double-precision floats, functions, structs, loops, dynamic arrays, modules, embedded file/string/math routines, overflow traps, and DWARF debug information. The generated runtime avoids libc, C, LLVM and virtual machines. It uses direct Linux system calls for ELF64 binaries and includes a Windows PE64 target with a dynamic kernel32 resolver; the repository also describes multi-OS runtime dispatch. The project provides the `rubixc` compiler driver, examples, a test suite, and a bootstrap compiler prototype. It is free and open source under the MIT License.

security-audit is a Cloudflare coding-agent skill that orchestrates multi-phase security audits of codebases. It uses reconnaissance to map architecture, trust boundaries, input surfaces, prior evidence, and deterministic coverage; assigns isolated agents to coverage-led hunting; sends each unique candidate to a fresh verifier; and records confirmed, needs_validation, and rejected results in machine-readable findings.json validated against a report schema. Independent agents verify final source claims, after which the skill derives target-neutral reports and coverage records. It can be installed with the Skills CLI and requires a tool-using coding agent with parallel sub-agent support, Node.js for its validators, and an OS-enforced sandbox for target-controlled execution. The repository is licensed under MIT.

Sumi.css is a minimalist, classless-first CSS framework for styling semantic HTML without JavaScript, dependencies, or a build step. Its element-based styles cover typography, forms, tables, navigation, cards, accordions, modals, progress indicators, and other common interface elements; classes and ARIA or data attributes add layouts, utilities, validation states, tooltips, and theme controls when needed. It supports automatic light and dark modes, CSS custom properties for theming, zero-specificity element styles for easy overrides, visible focus indicators, reduced-motion support, and ARIA-driven states. The distribution includes commented source and a minified production file, and is licensed under the MIT License.

tsrs is a Rust-based drop-in replacement for the TypeScript 7 compiler and language server. It aims for full TypeScript conformance while using less memory and checking types faster, and can be run through its CLI with pnpm or used through Node APIs. The project also provides WebAssembly support through @tsrs/wasm for WASI, Node, and browser environments. It is licensed under Apache-2.0 and is not affiliated with or endorsed by Microsoft.

An installable set of Claude Code skills that teaches Claude to act as a first-pass video editor in DaVinci Resolve Studio. It guides a workflow that ingests footage, cuts retakes and dead air, preserves and places B-roll, applies vlog-pacing rules, prepares audio and music, creates one-line captions, adds Remotion-based motion overlays, and configures YouTube exports. The package contains skill instructions rather than application code; Claude builds the small scripts for the user's setup and uses Resolve's external scripting interface. It includes setup, rough-cut, B-roll, Resolve scripting, audio, captions, motion-overlay, Figma round-trip, and export skills, along with configuration and editing learnings. It requires DaVinci Resolve Studio, ffmpeg, Python 3.9 or later, and a speech-to-text API with word timestamps; Epidemic Sound MCP, Figma MCP, and Node with Remotion are optional. The repository is licensed CC BY-NC 4.0 and prohibits commercial use without the author's permission.

video-short is an AI-agent skill that turns a raw recording into a polished short video with edited speech, B-roll, motion graphics, captions, music, and multiple layouts. It transcribes the recording, removes retakes, false starts, and dead air, researches the topic, finds relevant footage from the source video and the web, and presents a scene-by-scene plan before rendering. Its review page lets users pause on individual frames and provide feedback for iterative revisions; each render is saved as a separate version, with a 720p copy for sharing. The skill can use Claude, ChatGPT, Gemini, or another compatible agent, and is installed through an agent instruction or npx. It uses FFmpeg, yt-dlp, Python, and Node.js with the open-source HyperFrames tool and Parakeet speech model; optional AI-generated images, video, and music require a Gemini API key with billing enabled. The repository is licensed under MIT.

A digital notebook and canvas for students to write mathematical problems by hand for AI grading.