Agent Standup is a local developer tool that turns Claude Code, OpenAI Codex, and Mistral Vibe session files into a live, animated team-call visualization. It shows which agents are working, waiting, stalled, or need attention, along with sub-agent hierarchy, prompts, tool calls, command output, files, diffs, token counts, voices, and screen sharing. It can also replay past sessions, generate recaps, and export calls as video clips. The project reads the session files these coding agents already write, merges leads and sub-agents into a timeline, and incrementally parses new lines for live updates. It runs as a VS Code extension, a Node.js terminal command, a macOS desktop app, or a standalone local server; it requires Node 18+ for the command-line and standalone modes. The server listens on localhost and sends nothing by default, while translation, jokes, music, usage counting, and optional natural voices are separate features. Agent Standup is an unofficial community project, distributed under the MIT license.
Anatomy is an open-source agent skill for Claude Code, Antigravity, OpenCode, and Codex that explains technical ideas by constructing interactive isometric illustrations of physical machines, instruments, tools, or other mechanisms. It uses a workflow that collects the relevant numbers, invents a physical object for the concept, creates a parts list and geometry, renders parts back to front, adds interaction, and verifies the result with geometry and browser checks. The skill can produce standalone HTML pages, embedded SVG files, or React/Next.js components. Its SVG kit provides isometric solids, round parts, pipes, cables, four-tone shading, animation, cards, and live readouts; optional WebGL layers add effects such as fire, exhaust, water, caustics, light, and glow, while a beta 3D mode re-projects and re-shades moving parts each frame. The repository includes Node build scripts, complete examples, React components, reference guides, and headless Chrome tools for screenshots, scripted interaction, line-end checks, and geometry audits. It uses Node.js built-in modules without npm dependencies for the kit and examples, with React 18+ required for React output and Chrome or Chromium required for browser checks. The project is distributed under the MIT license. Its stated limits include not being a chart library, not targeting people or exact maps, and requiring WebGL for shader effects.
BetterWispr is a free, open-source, local-first voice-dictation app for macOS. It uses Apple on-device speech recognition, NVIDIA Parakeet, or Whisper to transcribe speech locally and paste the result into the active application; built-in models work offline after download. It supports hold-to-talk or toggle shortcuts, vocabulary and phrase replacements, voice commands, correction learning, and locally stored transcript history. Its meeting mode provides live transcripts, titles, summaries, and action items using Apple Intelligence or a local Ollama model. Optional Sarvam AI, Smallest AI, and OpenAI-compatible endpoints can be added, in which case audio is sent to the selected endpoint rather than handled entirely on-device. The project requires macOS 14 or later, recommends Apple Silicon, and is distributed under the Apache-2.0 license.
Chrome Relay is an open-source macOS tool comprising a Chrome extension, a local Go relay, and a command-line interface that lets coding agents such as Claude Code, Codex, Cursor, and Paseo operate an existing Chrome profile in background tabs using its signed-in sessions. The extension opens agent-owned tabs in a collapsed group and controls them through Chrome's debugger protocol, while the relay exposes a per-agent DevTools endpoint and supports either step-by-step agent-browser commands or the `chrome-relay do` command for completing a stated goal. The `do` workflow fills named fields from agent-supplied data, replays previously successful site flows, or uses a decision model such as TypeSafe's Jev or Claude Haiku to choose among page controls; it checks elements before and after each action and returns control when a decision is uncertain or potentially consequential. Actions are recorded in a local JSONL audit trail that includes pages, clicks, scripts, screenshots, uploads, popups, sign-in actions, and blocked pages while omitting typed values and URL query strings. Optional features include 1Password-based sign-in, Secure Enclave passkeys, iPhone approvals, notifications, and a faster decision model. The project is currently built from source and loaded as an unpacked extension; it requires macOS, Google Chrome, Go, and agent-browser, and is released under the MIT license.
Claude Show Your Work is a Claude Code plugin for displaying images and playing videos in the terminal, including web pages, designs, charts, screenshots, and Playwright recordings. It uses the terminal's graphics support to render files locally; images remain on the user's screen and are not sent to Claude, while videos are decoded with ffmpeg in short segments. For visual QA, the plugin can play recordings of the same test on old and new versions side by side and in sync, list changes beside the video, and associate test results with timestamps. It also supports design comparisons, responsive-width checks, video history, and the /show-your-work command for settings and tours. The repository provides the show_image and show_video tools, requires macOS, Ghostty, and Claude Code 2.1.287 or later, and is released under the MIT license.
DeepSeek Bot (DS Bot) is an Apache-2.0-licensed plugin for DeepSeek Harness that turns its sessions into user-configured, long-lived bots. Each bot uses a selected DeepSeek Harness model and stores editable memory on the user's computer; a main bot can manage other bots and group chats, while bots can hand work to one another with @ mentions and return results asynchronously. Group-chat administrators route requests to other bots, and shared capabilities can provide MCP servers, skills, and packs, with approval required for local MCP commands. Usage can be viewed by time, bot, and model, and data remains in the DeepSeek Harness home directory. It is an independent community project, not an official DeepSeek product.
DigUp is a free, open-source macOS menu bar app for searching files by their contents, including images, PDFs, documents, audio, video, screenshots, and optionally source code. It runs on Apple Silicon Macs with macOS 14 or later and keeps indexed data on the device. DigUp uses Google DeepMind's multimodal EmbeddingGemma 2 through llama.cpp and Metal. It converts images, document passages, video frames, and overlapping audio segments into a shared vector space, then combines semantic nearest-vector search with exact-word matching. Results can open a video or recording at the relevant moment, a PDF at the matching page, or code at the relevant line in a configured editor. The app reads only selected folders, excludes sensitive and ignored file types, and does not upload indexed files. It downloads the model once and checks for updates, with telemetry and accounts not required. DigUp can be installed from a DMG or Homebrew, includes a command-line tool, and is licensed under the MIT License.
Doory is an open-source, self-hosted design editor that runs in a web browser. It supports posters, social posts, stories, slides, invitations, resumes, thumbnails, business cards and other multi-page designs, starting from templates or a blank page, with layers, groups, masks, filters, curved text, snapping, rulers, animation, audio, presentation mode, comments and real-time co-editing. Guests can begin without an account, with their designs saved in the browser; signed-in designs are stored on the server running Doory. The editor performs background removal in the browser and exports designs to PNG, JPG, SVG, PDF, MP4, WebM and animated GIF. Self-hosting is provided through Docker Compose, which runs the web app, API, database and live-collaboration services; uploads can use local storage or S3-compatible storage. The repository states that Doory is released under the Attribution Assurance License, which permits commercial use provided the required credit, license and NOTICE are retained.
effortless is an MIT-licensed Claude Code plugin that automatically selects the reasoning effort for each prompt using a judge model, so simple prompts can use lower effort while harder tasks use higher effort. It uses Haiku 5.5 by default, with optional Jev judging through TypeSafe, and keeps short follow-ups at the previously selected effort. A status bar shows the selected effort and model, context usage, prompt-cache time, judge explanation, usage warnings, and controls for compaction or handing a session off to a fresh chat. Handoff creates a summary and continues in a new chat; compaction uses Haiku 5.5 by default and falls back to Claude Code's normal compaction if needed. An experimental setting can route simple prompts to a cheaper model, but it is disabled by default because switching models can require rebuilding the prompt cache. The plugin runs locally as Claude Code hooks, provides commands for settings, benchmarking, diagnostics, statistics, handoff, and updates, and can be installed through Claude Code's plugin marketplace.
ESP32 iPhone Agent is an open-source AI automation tool that lets a vision model control a real iPhone without installing an app, using a computer, AirPlay mirroring, and an ESP32-S3 board acting as a USB mouse and keyboard. UxPlay receives the AirPlay stream, a GStreamer pipeline converts it into JPEG frames, and the computer sends relative pointer, typing, scrolling, and key commands to the board over Wi-Fi. A planner model creates a short plan from the task and current screen, while an actor model examines each subsequent screenshot and selects actions such as tapping, swiping, drawing, typing, pressing keys, going home, or finishing. The server tracks and recalibrates the relative pointer, checks whether the screen changed after each action, and can retry or re-plan after missed actions, repeated actions, dropped links, or invalid model responses. The project provides a Python/FastAPI server and browser-based interface with live phone control, logs, screenshots with action overlays, manual controls, calibration, and an emergency stop. It runs on macOS or Linux with an OpenAI-compatible vision model, either remote or local; the phone requires AssistiveTouch but no Mac, developer account, jailbreak, or installed app. The repository is licensed under MIT.
Finder Menu is a Swift macOS Finder extension that adds New File, Open Terminal Here, Open Folder With, Copy Path, and Go to Parent Folder actions to Finder's context menu. New File accepts arbitrary extensions and can use saved templates; the terminal and editor actions detect installed applications or allow a custom choice. A sandboxed Finder Sync extension reads the clicked folder and settings, while a small unsandboxed helper app performs file creation, application launching, and Finder window navigation through validated findermenu:// actions. It supports local, external, disk-image, and network volumes, has configurable menu items, terminals, icons, and templates, and contains no network code or telemetry. It can be built locally from source or installed through Homebrew, and is released under the MIT license.
ftop is a lightweight, open-source system monitor for Apple Silicon Macs, developed as a native macOS app in Swift. It displays current CPU usage and per-core frequency, memory and pressure, network throughput, GPU usage and frequency, power draw, temperature, and the busiest applications in a resizable floating glass panel. Process rows fold helper processes into their parent app, and selecting a process shows its CPU, memory, and helper-process details. The panel adapts its layout to its size, uses animated readings and a stable process ranking, and stops most monitoring when hidden or fully covered. It provides menu-bar controls, a JSON5 configuration file, command-line commands, optional access to system processes through a privileged helper, and automatic release updates. It requires an Apple Silicon Mac running macOS 15 or later, is distributed outside the Mac App Store because it uses private macOS interfaces for some readings, and is licensed under the MIT license.
gbos-vm runs a real Googlebook OS recovery image in a GPU-accelerated virtual machine on Apple Silicon Macs. Its install script downloads and verifies Google's ARM64 recovery image, replaces hardware-specific components with virtual-device components from Android Cuttlefish, and boots the result under QEMU with Vulkan translated to Metal through MoltenVK; the project also builds patched virglrenderer, Mesa, a macOS viewer, and an Android pointer helper. It packages the VM as a macOS app with clipboard integration, configurable display, memory, CPU, networking, audio, Bluetooth, and pointer modes. The project is intended as a development VM rather than a verified Google or Dell product; it requires an Apple Silicon Mac and substantial local storage for downloads and build artifacts, and is licensed MIT for its own code, with one Apache-2.0 patch.
Husk is an open-source iPhone app for running Android apps and games from APKs and split APK bundles. It uses a translation layer for compatible 64-bit Android software: the game's native code runs directly on the iPhone while Husk provides Android library calls, Java interfaces, graphics through ANGLE or MoltenVK, sound, touch, and controller support. For software the translation layer cannot run, Husk boots a full LineageOS system under QEMU, with substantially slower processor emulation. The app supports engines and frameworks including Unity, Unreal Engine 4, cocos2d-x, Godot, SDL2 and SDL3, GameActivity, NativeActivity, and Rockstar's engine used by GTA: San Andreas. Both modes require JIT on iOS; Husk supports sideloading through SideStore, AltStore, TrollStore, or other methods and requires iOS 16 or later. It is distributed as an IPA and licensed under GPL-2.0-or-later.
lintapple is an npm-based UI linting tool that checks a running web page against rules for fonts, text sizes, contrast, control size, screen bounds, spacing, alignment, radii, and related layout properties. It runs from the CLI against an application URL and reports each finding with the rule, measured value, and suggested fix; it can also wait for an application-ready selector and emit a JSON report. The package provides a Playwright helper, `expectUI`, which reads the project's `ui-policy.json`, fails the test when findings are detected, and attaches a JSON report. Additional layout checks use attributes such as `data-ui-stack`, `data-ui-actions`, `data-ui-container`, `data-ui-content`, and alignment or radius markers. It requires Node.js 22 or later and is distributed under the MIT license.
Mailcheap is a self-hosted newsletter application built with Next.js that uses Amazon SES for email delivery. It provides an editor with live email preview, Markdown and raw HTML support, signup forms, unsubscribe handling, a public archive, campaign statistics, and an API for contacts and campaign drafts. Before sending, it checks requirements such as a subject, preview text, unsubscribe link, postal address, verified domain, and Gmail's message-size limit. Campaigns can be tested, scheduled, and sent in batches through Upstash QStash, with engaged readers sent first. SES delivery, bounce, and complaint events are received through Amazon SNS; bounced or complaining addresses are suppressed across newsletters, and a campaign pauses when its bounce or complaint thresholds are exceeded. The application uses MongoDB for contacts, campaigns, and statistics, and can optionally use Upstash Redis for shared rate limits and Vercel Blob for editor image uploads. It supports multiple newsletters in one installation and can expose archives on custom domains. The repository is an MIT-licensed, one-time open-source release with no support, disabled issues, and no planned updates.
ManimGX is an open-source Python animation engine by Academa for producing mathematical and 3D videos through an API compatible with Manim Community Edition. Its Rust and wgpu renderer exports scenes to MP4, while the same system can run in a browser using Pyodide in a Web Worker and WebGPU. It supports 2D and 3D scenes, surfaces, meshes, moving cameras, text, equations, plots, LaTeX or Typst input, synchronized speech through self.say(), and an inspect command for storyboards and 2D layout checks intended to provide feedback to coding agents. It can be installed with pip on macOS, Linux, or Windows with Python 3.13+, or used from npm for browser playback. ManimGX's own code is MIT-licensed; native wheels include components with separate licensing terms, including GPL-licensed x264.
motionmaxxing is an open-source agent skill for Claude Code and other agents that read a SKILL.md folder. It generates motion-graphics films from plain HTML, using GSAP and a vendored Three.js runtime for frame-by-frame animation, 3D hero objects, typography, native-style product UI, and deterministic Chrome/FFmpeg rendering. Its ten-step process covers brand capture, concept and storyboard creation, scripting and voice, building the hardest beat first, rendering, scripted inspection, sound-to-picture, and delivery. The skill uses measured motion precedents and quality gates for render validity, empty frames, end-card duration, composition, and page-chrome patterns; it produces files including final.mp4, editable index.html, STORYBOARD.md, and NOTE.md. Installation checks Node.js, Python, FFmpeg, and Google Chrome without installing system-wide dependencies. The repository is licensed Apache-2.0, while bundled third-party components retain their own licenses, and the project describes itself as an early release with limitations including visual checks that remain manual and limited WebGL rendering speed.
open-appleremote is a Python tool for using first-generation Apple Siri Remotes (A1513/A1962) on Linux. It unlocks the Bluetooth HID device, decodes its touch surface, buttons, gestures, and tilt reports, and emits JSON events or Linux key presses; an optional virtual keyboard can control applications such as Kodi, browsers, and emulators. It also provides a local web lab, capture replay, and an experimental voice path that reconstructs Opus audio from Bluetooth HCI-monitor traffic and writes Siri presses as WAV files. The package uses Python 3.8 or newer with no dependencies for its main functions, and requires Linux with BlueZ, hidraw, and appropriate udev permissions for live use. Voice decoding additionally requires libopus and CAP_NET_RAW or its supplied system-service setup. The project is an early release, tested on an A1962 remote; second- and third-generation Siri Remotes are not supported, and the virtual keyboard and live voice path are identified as not fully hardware-tested. It is distributed under the MIT license.
OpenFilm is an open-source AI video agent and editor that lets coding agents such as Codex and Claude Code create films as HTML-based web pages. A film is a folder containing a film.html whose sections define timeline tracks of web pages, video, images, and audio; each animated page renders a frame at time t through a frame(t) function using browser technologies such as CSS, SVG, Canvas, WebGL, Three.js, GSAP, or video. Its browser-based Studio lets users move, scale, rotate, restyle, trim, and cut elements and clips, while saving edits back to the film files so agents can preserve them. The CLI can open films, inspect frames or ranges with look, render MP4 and other outputs, and obtain media through configured external providers. The look command checks frames in shuffled orders and reports changes, page errors, missing fonts or media, and text clipped by the frame. The CLI, Studio, desktop app, and format are MIT-licensed; local use requires Node, FFmpeg, and Git, while the desktop app bundles FFmpeg and a headless browser.
PS2 Web Recomp is an open-source project for running PlayStation 2 games in a browser through static recompilation. It analyzes a game's Emotion Engine executable, translates its machine code into C++, and compiles the result to WebAssembly so the main CPU is not interpreted. A runtime emulates the remaining console hardware and operating-system behavior, including the I/O processor, vector units, SPU2 audio, MPEG-2 movies, memory cards, and the Graphics Synthesizer; the graphics renderer uses WebGPU compute shaders, with a software rasterizer as a reference. The porting scripts accept an ISO or an archive containing one, iteratively discover code reached through indirect control flow, compile uploaded vector-unit microprograms, and produce a browser page and optional native runner. The repository currently documents ports of Gran Turismo 4 and God of War II, but includes no game data; users supply disc images of games they own. It requires a cross-origin-isolated server for the browser build and is licensed under the MIT License, with a retained BSD 2-Clause notice for adapted MPEG-decoder tables.
robot-heads is a React component library of animated 3D robot heads that visualize an AI agent's state through LED-matrix faces. It supports nine states, including idle, thinking, searching, listening, speaking, working, happy, error, and sleeping; each state has its own face, motion, and antenna light. Four head shapes, configurable colors and animation speed, pointer interaction, reduced-motion behavior, accessibility labels, and paused rendering are provided through component props. The heads are rendered on a 2D canvas without WebGL or runtime dependencies, share an animation loop, and reuse baked textures for heads with the same shape and size. It requires React 18 or newer and is distributed as an npm package under the MIT license.
RSI-Anything is a domain-general framework by Gen-Verse for recursive self-improvement using world-model predictions and interaction feedback. Its shared workflow proposes candidate actions, predicts their consequences with a JEPA-based world model, selects actions, evaluates them externally, learns from the resulting experience, and retains improvements for later rounds. The repository provides a Python command-line workflow and task adapters for code repair, GUI interaction, laboratory planning, reaction optimization, and immune-science workflows, with domain-specific states, actions, prediction targets, and feedback. It can be installed from the repository with Python 3.10 or newer and includes configuration-driven planning, execution, reporting, run resumption, tests, and task guides. The project-owned code is licensed under Apache-2.0.
RustGIS is a free, open-source GIS application written in Rust for mapping, analyzing, editing, and sharing geospatial data. It runs natively on Windows, macOS, and Linux and in the browser through WebAssembly, using a GPU-rendered ribbon interface and keeping local files on the user's machine unless a network operation is requested. It supports vector data such as GeoJSON, Shapefile, GeoPackage, GeoParquet, FlatGeobuf, KML/KMZ, and many other formats, as well as GeoTIFF and cloud-optimized GeoTIFF rasters. Its mapping and analysis functions include projections, 3D globes, symbology, labeling, editing with undo, raster processing, geoprocessing, ModelBuilder, and more than 1,000 GeoLibre and Whitebox tools. Every ribbon action is represented as a command that can be run through the desktop application, rustgis-cli, Python, notebooks, or an MCP server. The project also provides sandboxed WebAssembly plugins, web services and tile support, layouts and reports, export and sharing tools, and an optional AI Assistant that can invoke RustGIS commands through selected model providers. The repository is licensed under Apache License 2.0; desktop builds, a web application, command-line tools, package-manager installations, and source builds are provided.
SimRun is a macOS tool that runs iOS Simulator builds directly in ordinary Mac windows without booting a simulated iPhone. It launches the unchanged app built for the iOS Simulator, maps Mac mouse, trackpad, keyboard, clipboard, and motion input to iPhone interactions, and provides a launcher for managing apps, profiles, data containers, keychain items, logs, and runtime or device choices. Its command-line tool can launch, list, quit, screenshot, inspect, and stream logs from apps, run them headlessly with separate profiles, feed microphone input, and experimentally drive interfaces through a JavaScript API. It can also convert some development or ad hoc device builds for Simulator execution. SimRun requires an Apple-silicon Mac, Xcode with an iOS Simulator runtime, and disabled System Integrity Protection; it is an experimental project and does not replace testing in the Simulator or on a physical device. It is distributed under the MIT License.
Solid Glass is an open-source design system for AI applications, implemented as a single lightweight CSS stylesheet for shadcn/ui components. It targets shadcn/ui components through their existing data-slot attributes and classes, so the components generally require no changes, and supports light and dark modes with a material-inspired treatment based on edge highlights, borders, shadows, and three glass thicknesses. The repository includes a Create page with controls for thickness, tint, chart color, type, radius, rim, and blur; it previews each selection and outputs the corresponding CSS. It also provides shadcn-installable components for agent interfaces, including status indicators, composer fields, tool steps, diff views, context rings, task notices, and reasoning displays, along with blocks, examples, documentation, and a component registry. Solid Glass is distributed under the MIT License. The published site and registry are hosted through GitHub Pages, and the repository includes scripts for local development, building, type checking, documentation generation, and registry generation.
sqlite-multiwriter is a VFS extension for SQLite that allows multiple threads or processes to write to the same database concurrently without changing SQLite, the SQL, the schema, or the ordinary SQLite database format. Each writer uses a private write-ahead log and a snapshot; at commit, the engine validates the pages read and written, orders the commits, publishes them through a commit log, and compacts the results into the database file. Writers that conflict on the same page can receive SQLITE_BUSY_SNAPSHOT and must retry the whole transaction; the optional rebase mode can replay supported row-level INSERT, UPDATE, and DELETE operations that lost only because they shared pages. It supports extension loading or static linking, thread and multi-process modes, and platform packages for Linux, macOS, iOS, Android, and Windows. It requires WAL mode and a local filesystem, and is distributed under the Apache License 2.0; the vendored SQLite code is public domain.
Statecraft: Israel is an independent, open-source, turn-based grand-strategy game about governing Israel from 15 May 1948 through 1951. The player serves as prime minister, appoints real ministers within a coalition, allocates a limited budget among security, immigrant absorption, transit camps and austerity, manages security through an escalation ladder, and handles historically dated events such as the First Truce, the Altalena affair and the Bernadotte assassination. A deterministic C# simulation produces repeatable outcomes from the same choices; saved games contain the simulation seed and command list. Unity renders a terrain map built from elevation and climate data, with sourced locations, military tokens, historical events and a “Why?” panel explaining outcomes. The current build is an early prototype with cabinet, budget, defence, map, autosave/load and browser-build functionality; planned features such as full unit control, cities, construction, diplomacy and combat are not yet implemented. The project is built for local use with Unity and .NET, with no hosted build yet. Its code is licensed under MIT, original art, documentation and data under CC BY 4.0, and game content text under CC BY-SA 4.0. Historical facts and events are accompanied by cited sources, while the project explicitly identifies its framing as Israeli and says it is not affiliated with or endorsed by Firaxis, 2K or the State of Israel.
Talorys is a free, open-source personal AI assistant that runs in the user's own Cloudflare account. It provides chat with streaming responses, inspectable and editable memories, tasks, notes, projects, and one-time or recurring reminders and routines; it is designed for one user, with no signup or developer-operated account. A React frontend on Cloudflare Pages sends API requests through a Pages Function and service binding to a private Cloudflare Worker. The Worker hosts a TalorysAgent Durable Object with SQLite-backed conversations, memories, tasks, notes, projects, sessions, settings, and automations. Workers AI provides streaming chat and tool calling, while Durable Object alarms run scheduled reminders and routines. Relevant memories can be selected for each prompt, with optional semantic recall and memory history for superseded facts. The npm installer, `npx create-talorys@latest`, deploys the application and configures authentication, secrets, resource names, and the service binding. The project reports no telemetry, analytics, tracking, or advertising code; data remains in the user's Cloudflare account, although Cloudflare processes chat data for its services. It is licensed under the MIT License and is designed for Cloudflare's free plan, subject to Cloudflare quotas.
USD BRep is NVIDIA's standalone C++ and Python library for NURBS-based boundary-representation solid modeling in OpenUSD. It includes BRep modeling libraries, an omniSolid schema plugin, runnable examples, tests, and a Python GUI. Its healing utility, `usd.heal_file`, repairs BRepArray definitions in the layers used by a USD composition and writes a new scene while preserving layers, references, native instancing, transforms, and materials; material-subset opinions outside the BRep's defining prim specification are rejected rather than discarded. The repository uses Packman to obtain dependencies such as OpenUSD, Python 3.12, pybind11, and Premake, and provides build scripts for Linux, macOS, and Windows. USD BRep is licensed under Apache License 2.0; selected repository tooling is under the MIT License.
vegaml is a Python package for producing typed decisions from frozen Qwen language models without generating text. Given a state and questions with predefined answer types—choice, score, or boolean—it pools features from one model pass and feeds them into a 64-dimensional decision-space physics engine, where candidate answers form Gaussian potential wells and a damped particle simulation determines the result. Each answer includes probabilities, a conformal prediction set, an abstain flag, and an indicator for decisions that settle outside the model's expected range. The package supports 0.8B and 4B backbones, image inputs, local inference, and long contexts up to 73,728 tokens; it also supports fitting gated adapters while keeping the language model frozen. It is distributed through PyPI and uses Apache 2.0 licensing.
W1700K SDR is an open-source software-defined radio project that repurposes compatible W1700K-series Wi-Fi 7 routers as wideband receivers. It replaces the router firmware with custom U-Boot, Linux, kernel-driver, and MT7996 radio firmware components; the firmware initializes the radio, controls tuning, gain, filtering, and capture, and streams raw complex IQ samples over the router’s 10 GbE port. On the Linux host, w1700k-stream distributes the IQ stream through shared memory, w1700k-spectrum serves a browser-based spectrum and waterfall with live controls, and sdr-recv records captures. The project supports the router’s 2.4, 5, and 6 GHz radio bands and sample rates up to 320 Msps, with optional correction for decimation roll-off. Installation requires compatible router hardware, a serial adapter, and a 10 GbE-connected Linux host; it replaces the stock operating system but provides a restore procedure. The project is receive-only and is licensed under a combination of MIT, GPL, and BSD-derived licenses as documented in the repository.
Yingxia is a native SwiftUI TVBox player for Apple TV, iPhone, iPad, and Mac. It imports TVBox JSON subscriptions by URL and M3U or TXT playlists, presenting live channels in a guide, on-demand catalogs by source, and unified search across subscriptions. Playback uses AVPlayer with HLS support; the Mac version also provides a separate Picture in Picture player window. The app runs JSON and JavaScript sources and JAR sources, with Python support on Mac, using on-device plugin hosts and lightweight ports with the original plugin as a fallback. It has no account or analytics, keeps subscriptions and logs on the device, and retains the last working subscription copy when a refresh fails. The project is distributed as a signed and notarized free Mac download for macOS 15 and Apple silicon. Apple TV and iPhone/iPad versions are preview builds installed by sideloading, and the apps are not yet on the App Store. Yingxia's original code is licensed under GPL-3.0-only.
Searchable transcript of GitHub Trending Weekly #53: anatomy, talorys, DeepSeek-Bot, mailcheap, manimgx, Husk, doory, DigUp — Github Awesome (15:13). Search for a phrase, then click its timestamp to jump straight to that moment in the video.
Captions sourced from the original video on YouTube, published by Github Awesome. The video, its captions and all related intellectual property remain the property of their respective owners; AINotes claims no ownership. Provided for research, accessibility and search — see the Transcript Notice and Copyright Policy.
00:00 Welcome back to GitHub awesome. This is GitHub trending weekly number 53. 35 trending open source projects on GitHub right now. Let's go. Anatomy is an agent skill that turns technical ideas into interactive isometric illustrations. Ask it to explain a rate limiter and it builds a little machine with moving parts and live readouts. The result can be a standalone SVG, an HTML page, or a React component.
00:27 For effects like water, light, or exhaust, it can add WebGL to the drawing and check the geometry for mistakes. Tois gives you a personal AI assistant in your own Cloudflare account. Chat with it, save notes and tasks, organize projects, and schedule reminders that run even when your browser is closed. It remembers useful details, and you can inspect, edit, or delete those memories.
00:53 One command deploys the app with no server run by the toor's developers and no telemetry. DSBot is a plugin that turns Deep Seek Harness into a team of your own AI bots. Each bot is a longunning assistant with its own model and its own memory kept as markdown files you can open and edit. A main bot called Chief manages the rest and walks you through setup.
01:14 drop bots into a group chat and an admin bot answers by default, pulling in specialists when it needs them, or you can mention one directly. Mailchip is a self-hosted newsletter app that sends through Amazon SCES for about 10 cents per thousand emails. So, 10,000 readers getting four issues a month runs about $4. You write in an editor with live preview, and it checks for an unsubscribe link and a verified domain before anything goes out.
01:46 sends go out in batches and it pauses itself if bounces or complaints climb too high. Manomm GX lets you make math explainers and 3D animations in Python. It follows Manom's API with a Rust GPU renderer built for fast video output. Equations and plots can share a scene with meshes, surfaces, and moving cameras. You can also run scenes in a browser with web GPU, while storyboard and layout checks help coding agents see what their code actually drew.
02:17 Want to try an Android game on your iPhone? Husk offers two ways. For supported 64-bit games, it runs the game's Android code on the iPhone's processor while translating the graphics, sound, and input around it. For other apps, it can boot a full Android system through QU, though that's slower. Add an APK and launch it full screen. Setup requires sideloading Husk and enabling JIT.
02:42 ESP32 iPhone agent lets an AI model use a real iPhone with a $7 ESP32S3 board as its hands. The phone mirrors its screen to a computer over AirPlay so the model can see it. The board plugs into the phone as a regular USB mouse and keyboard and with assist of touch on every click becomes a tap. A planner model writes the steps and an actor model takes each action from the latest screenshot.
03:09 Dory is a canvas style design editor you can run on your own server. You can start without an account with guest design saved in your browser. Background removal runs there too alongside animation and video export. It has layers, masks, curved text, 27 entrance and exit animations, and background removal that runs right in the browser. Teams can edit together in real time.
03:30 An AI generated UI can look fine until a button is too small or text spills out of a card. Lintapple checks a running page against your UI rules and reports the element, measured value, and suggested fix. It covers fonts, contrast, control size, and screen bounds with optional checks for spacing and alignment. Run it from the command line or inside playright where findings fail the test.
03:58 Robot heads is a React component library that draws animated robot heads, each with a TV style body and an LED face to show what an AI agent is doing. It has nine states, including thinking, searching, speaking, and error. Each one gets its own face, motion, and light color, and a short glitch plays when the state changes. Everything renders on a plain 2D canvas with no WebGL and no runtime dependencies.
04:24 Finder menu is a small swift tool that adds the right-click options Mac OS Finder never shipped with. Rightclick empty space in a folder and you can create a new file with any extension, open a terminal right there, or open the folder in VS Code, cursor, zed or Xcode. It can also copy the folder's path or move up to the parent folder. It runs as a sandboxed finder extension plus a tiny helper app.
04:49 Digup is a Mac OS menu bar apps by what's inside them built on Google DeepMind's embedding Gemma 2 running through llama.cpp. Describe what you want in plain words and it finds the matching photo, PDF page, document, or code file. It can even jump to the right moment in a video or podcast. Text, images, audio, and video all share one index, and nothing leaves your Mac.
05:16 Openilm lets a coding agent build a video, then gives you an editor to change it. The agent makes shots as web pages and mixes them with footage, images, and sound on an HTML timeline. In Studio, you can move elements, change text and colors, trim clips, and export the result. The agent can also inspect frames for missing media, broken fonts, and text cut off at the edges.
05:37 Yinxia is a native Apple player for TV box subscriptions. Add your own JSON subscription or M3U playlist, and it organizes live channels and on demand cataloges with one search across your sources. It uses Swift UI and AV kit with a Mac app and preview builds for iPhone, iPad, and Apple TV. It supplies no media itself. What you can watch depends on the sources you add.
06:04 Schoolite multiwriter lets many threads or processes write to one SQLite database at the same time without changing SQLite itself. It loads as a VFS layer and gives each writer its own write ahead log. At commit, writes that touch different pages go through without waiting. If two writes touch the same page, the second one is told to retry. The file stays a normal SQLite database.
06:29 Chrome Relay lets coding agents like Claude Code Codeex use your everyday Chrome signed in as you. A Chrome extension and a small local relay connect the agent to your browser. Each agent works in its own background tabs tucked into a collapsed group, so your window never jumps. Agents can only touch the tabs they opened, and every page and click goes into a local audit log.
06:52 What would you prioritize if you were governing Israel in May 1948? Statecraft: Israel puts you in the prime minister's chair, balancing a tight budget, cabinet appointments, and security decisions as historical events unfold. The game cites sources for its facts and lets you inspect why events happened. It's an early Unity prototype covering 1948 through 1951 with an explicitly Israeli historical framing.
07:23 F-Top is a lightweight system monitor for Apple Silicon Max written in Swift. It puts everything in one floating glass panel that you can resize to any shape. You get one column per core with performance and efficiency core split plus memory per GPU network speed power draw in watts and the busiest apps by icon. It skips history graphs and only shows what's happening right now.
07:48 Vega returns type decisions from a small language model without generating any text built on frozen Quen 3.5 models at point8b and 4B. You give it a state like an email and a question with fixed answers. The model reads everything once and a small physics engine drops a particle into a decision space where each answer is a well. Wherever the particle settles becomes the answer with a calibrated confidence and an option to abstain.
08:17 Motion maxing is an agent skill for clawed code that makes AI generated motion videos look like a motion designer made them. It builds each film in plain HTML with GSAP and 3JS, then renders it locally. The skill works through 10 steps, starting with capturing the brand from a URL and ending with sound. Along the way, scripted checks catch dead frames and end cards that run too long.
08:40 Open Apple Remote makes the first generation Siri remote, the one with the touch surface from the 2015 and 2017 Apple TVs, work on Linux. Linux pairs with the remote, but it stays silent until it's unlocked. This tool unlocks it and decodes the touch and button reports in plain Python with no dependencies. You get swipes, clicks, and tilts as JSON events or map to key presses.
09:07 Better whisper is a free local voice dictation app for Mac OS that runs Whisper, Nvidia Parakeet, or Apple's own speech recognition right on your Mac. Hold option and space, speak, let go, and the text lands in whatever app you are using. It learns from your corrections, so a word it got wrong gets spelled right next time. It also handles meetings, writing up summaries and action items with a local model.
09:35 PS2 Web Recomp translates PS2 game code into C++, then compiles it to web assembly with web GPU graphics. The maintainers report Gran Turismo 4 running at 50 frames per second in Chrome on integrated graphics at twice the original resolution. Browser memory cards keep your saves. The catch, only two games have been brought up so far. You supply your own disk image and the generated builds stay on your machine.
10:03 Solid glass is an open-source design system for AI apps built as one stylesheet on top of shad/ UI. The look is a thick smoky glass that stays readable because the face is plain gray and doesn't pick up color from what's behind it. The glass effect lives only at the edges in a thin rim of light and a soft shadow. A create page lets you tune thickness, tint, and blur, then copy the result as a few lines of CSS.
10:30 RustGis is an open- source desktop GIS written entirely in Rust with a ribbon interface built on EGUE and rendered on the GPU. It loads shape files, geojson, and cloud optimized geoiffs. It handles symbology, labeling, and editing and covers analysis like buffers, clips, and dissolves. The same codebase runs on Windows, Mac OS, Linux, and in the browser through web assembly.
10:54 USD Bre is Nvidia's open-source solid modeling library for open USD with C++ and Python interfaces. It's a kernel for nerbs based boundary representations, the precise surfaces CAD tools use to define solid parts. It comes with examples for each API layer and a small guey you can run from Python. A healing tool repairs broken brep data in USD files and keeps their layers, references, and materials intact.
11:20 Agent Standup turns your claw code and codec sessions into an animated team call using the logs already on your machine. Agents speak, share actual command output, and diffs and show who's waiting on whom. The useful bit is spotting stalled work without digging through transcripts. It runs locally and the default visualization adds no model calls. RSI anything is a framework for recursive self-improvement built on Gen versus Jeppa anything world models.
11:46 Each round the system proposes actions and has a world model predict what they'll do. Then it picks the best ones, tests them against real feedback, and keeps whatever improved. The same loop works for code repair, browser tasks, lab planning, and chemical reaction tuning. GBOSVM runs Google Book OS, the Androidbased system made for the Dell Google Book in a GPU accelerated virtual machine on Apple Silicon Max.
12:14 The install script downloads Google's own recovery image, checks it against a pinned hash, and swaps in virtual hardware parts from Google's Cuttlefish project. It all boots under Cumu with Vulcan translated to Metal and comes packaged as a normal Mac app. Claude Show Your Work is a Claude code plugin that shows images and plays videos right in your terminal using the Kitty graphics protocol.
12:39 Claude can pull up a screenshot, a chart, or a Playwright recording without being asked. For QA, it puts before and after runs side by side and lines up each step, even when one version loads slower. Below the clip, you get a timestamped list of what changed and a short pass or fail summary. W700K SDR turns a Geekch U700K Wi-Fi 7 router into a wideband softwaredefined radio receiver.
13:08 It replaces the radio's firmware with the project's own code which was decoded register by register. The radio captures raw samples at up to 320 million per second across the 2.4 5 and 6 GHz bands. Those samples stream over 10 GB Ethernet to a Linux machine which shows a live spectrum and waterfall in the browser. The allword skill builds a polished website from a single line and it runs in claude codecs or any agent that reads skill.md files.
13:40 Ask for a portfolio with your name and the agent pitches three concepts, scores them and picks one. Then it writes a design system, builds the site with GSAP and WebGL motion, and checks it in a real browser on desktop and mobile. Effortless is a clawed code mod that picks the reasoning effort for every prompt, so you don't have to touch the effort control.
14:01 A fast judge model reads each prompt in about a second, gives quick questions low effort and hard tasks high effort, and tells you why. A bar above the prompt shows how full your context is and how long until the prompt cache goes cold. Iris 3B is an open textto image model with three billion parameters that draws every pixel itself with no VAE in between.
14:26 It's a diffusion transformer that reads prompts through a frozen quen 3VL text encoder. A small pixel transformer at the end decodes each patch straight back into pixels. The repo includes the weights, the full training code, and a Gradio demo, all under Apache 2.0. Simun runs iPhone apps built for the iOS simulator directly on a Mac without booting a simulated iPhone.
14:50 Every app gets its own regular Mac window that you can resize and minimize. The app still runs on the iPhone's own UI kit and Swift UI, while your trackpad scroll, pinch, and rotate become touch input. A command line tool launches apps, takes screenshots, and streams their logs to the terminal.