GitHub Trending Weekly #59: 35 trending open-source projects on GitHub right now
Agentes.chat is an open-source, self-hosted team-chat platform in which people and AI agents work together in channels, threads and direct messages. Each agent has a role, owner, model and monthly budget, can use permitted MCP tools, and can operate its own computer and browser. A router selects agents for unassigned messages, while sensitive actions such as merges, deployments and rollbacks pause for approval in chat; agents can also run scheduled or event-triggered routines, retain workspace memory and work with Claude Code or Codex on a connected computer. The platform includes web, desktop and phone clients, a wiki, drive, tasks and notifications, and can be deployed with Docker Compose and PostgreSQL or used as the hosted agentes.chat Cloud service. The core application is licensed under AGPL-3.0-only; its maker, Codari Global, LLC, offers commercial licensing for closed or rebranded modifications.
AgentRun is an open-source platform from Parcha Labs for running and managing agents built on harnesses such as Claude Code, Codex, OpenCode and pi. It provides a server, runners and a shared SQLite or Postgres store for durable execution: runs, revisions, steps and journals persist, interrupted work can resume after a server restart, and additional runners can claim work after a lease expires. Its learning system gives each agent a git-backed learning folder. After a run, the agent can propose notes, skills and workflows; code checks and a grader filter the proposals before accepted items are committed for later runs. Typed workflows can call tools, models, decisions and other workflows, with branching, bounded loops and agent fallbacks. The platform also supplies run-level approvals, spending caps, questions for human input, triggers, MCP and Composio connections, templates, and computer environments including sandboxes, desktops, phones and browsers. AgentRun includes a browser console, CLI, API, TypeScript client and Python client. It can run locally with npx, in Docker, or as a server deployment, and is licensed under Apache-2.0.
creator-generalist is an MIT-licensed skill pack for AI coding agents that guides 2D animation, 3D production, video editing, sound design, and small-game development. It provides generalist, craft, tool, learning, and sound skills that can be installed as plugins in Claude Code, Codex, Cursor, Gemini CLI, or OpenCode, with support for tools including Blender, Three.js, Godot, Unity, Unreal, Remotion, Krita, Resolve, and FFmpeg. Its production skills use an eight-step workflow: brief, reference, script and beats, preparation and a look or play gate, assets, build and animation, sound, and quality assurance with a preview and final output. The 3D workflow presents style frames, uses real CC0 or CC BY assets, produces a 540p preview before final rendering, and checks issues such as foot slip, teleports, contact gaps, clearance, and caption synchronization. The game workflow starts with a playable greybox before adding art, then tests errors, win and lose paths, restart behavior, determinism, frame time, controls, and softlocks through scripted probes. The pack can create outputs such as animated 2D explainers, 3D scenes and assets, video clips, sound mixes, and playable web or engine-based game prototypes.
Folio, also called Files, is Omarchy's file manager, built with C++ and Qt Quick. It provides grid, list, and Finder-style column views, with keyboard navigation, on-demand folder reading, and thumbnails shared through the standard thumbnail cache. Space opens full-size previews for images, video, audio, PDFs, text, and Markdown; the sidebar includes home folders, pinned folders, Recents, Trash, drives, and tailnet machines accessed over SFTP. It supports copying, moving, renaming, compression, deletion, and sharing through Dropbox links, Taildrop, or LocalSend, and a split-pane mode moves items between folders. Plugins can add sidebar sections, sharing options, and views, including the Tailscale, Dropbox, and LocalSend integrations. The project is distributed under the MIT License.
github-cool-readme is a GitHub Action that generates animated SVG statistics cards for a GitHub profile README and publishes them to a cards branch of the profile repository. It can produce summary, language, weekly-rhythm, notable-contributions, streak, and about cards from GitHub contribution data, repository stars, and pull-request activity; its language card weights each repository's language mix by the user's commits, while the weekly and streak cards use the contribution calendar. Private and organization work can be included, but cards show only aggregate totals for private repositories and do not expose their names. Cards support GitHub-style classic rendering, round-LED dot-matrix rendering, and square-pixel OLED rendering, with configurable colors, backgrounds, animations, ordering, and displayed metrics. A setup script creates the workflow and configuration on the user's account; it requires a GitHub token with the documented repository, organization, and user scopes. The project is distributed under the MIT license and runs locally with Node.js 22.18 or newer.
gpucanvas is a TypeScript and React library for placing live DOM or React nodes on an infinite canvas rendered with WebGPU. It packs node snapshots into texture-array atlas pages, parks rendered pages with content-visibility, and draws visible nodes with instanced GPU quads while pan and zoom leave the DOM unchanged; the real elements continue to receive input, maintain state, and support CSS and accessibility. It includes an editor for selecting, drilling into, moving, resizing, and snapping nodes or elements; WGSL shader effects and transitions such as glass, frost, glitch, dissolve, wipe, and fade; snapshots, image diffs, and filmstrips; and marks including arrows, pins, notes, and cursors. Its agent-view API exposes the visible canvas as a JSON tree with element identifiers, roles, text, and rectangles, and can capture elements or regions as images. The package is installed from npm and provides a React API including GpuCanvas and CanvasNode. It uses WebGPU and HTML-in-Canvas where available and falls back to a CSS-transformed DOM container elsewhere; the repository describes the project as early and subject to API changes.
Haru is a low-latency, multilingual, full-duplex spoken-dialogue model that listens and speaks simultaneously. It processes one 133-millisecond audio frame at a time and produces an audio frame and text, combining user audio, model audio, and model text in a single language-model position so listening and speaking do not wait for each other. The model generates a 64-dimensional continuous audio latent per frame through a small MLP head using MeanFlow sampling, without a separate audio transformer or discrete audio tokens. A text character card controls the character, while a short reference recording controls the speaking voice. It supports English, Spanish, Portuguese, German, French, Chinese, Japanese, Russian, and Korean, and uses the causal VibeVoice acoustic tokenizer for 24 kHz audio. The repository provides PyTorch inference and a browser demo, an MLX implementation with a native Apple Silicon macOS app, and a vLLM-Omni plugin for serving concurrent conversations. Haru-4B is released as a pretrained base model for research under the Apache License 2.0; the README notes that it is intended for natural conversation rather than reliable factual answers or instruction following.
islandcn is an open-source Dynamic Island-style React component installed through shadcn. It provides an activity stack with idle, compact, split, and expanded presentations, supporting actions such as uploads, timers, music playback, incoming calls, rides or delivery ETAs, recordings, nearby transfers, and battery status. A React provider mounts the island, while components start, update, and end activities with separate compact and expanded content; the engine includes spring-based transitions, long-press interactions, alerts, and reduced-motion support. The project includes presets, a shared content kit, a playground for building activities and tuning motion, and documentation. Its repository is licensed under MIT. It recreates the look and motion of Apple's Dynamic Island without using Apple's assets and states that it is not affiliated with Apple.
klad is a 2D painting and animation app by Indicode BV for collaboration between a human and a coding agent. The human paints in the editor while an agent uses klad's command-line tools on the same drawing, brushes, layers and timeline; strokes retain their paths, pressure, brush and colour, and canvas notes and selections provide feedback to the agent. The app supports raster painting, editable vector shapes, pixel art, layers, masks, clipping, effects, animation timelines, onion skins, in-betweens, bones with inverse kinematics, and exports to PNG, sprite sheets, GIFs and Godot assets. Drawings are folders of plain files, and the documented format, CLI and agent skills are MIT-licensed so other programs can read and modify them. klad is source-available rather than open source. The repository can be built for macOS, Linux and Windows, although the current editor is primarily developed on macOS and Linux tablet pressure is not yet supported. There are no installers in the repository's current early release; users build it themselves, subject to the PolyForm Internal Use License terms.
Linger is an open-source personal-agent application from SimplexAI that runs AI models locally on consumer Macs and PCs. It provides a desktop app and a Docker Compose deployment, with support for local models and selected external runtimes such as Claude Code, Codex, or Pi. Its PolyStrata inference engine distributes mixture-of-experts model components across GPU memory, system RAM, and SSD storage. An Autopilot installer examines the computer's hardware, evaluates available model and inference configurations, selects a model and quantization that fit, installs it, and verifies the installation by asking a real question. Linger also includes a local model scheduler for multiple agents and background jobs. The application stores agent memory and searchable wiki pages on the user's device, allowing agents to share information, preferences, and decisions across conversations. Its Another Me agent can monitor an inbox, calendar, and Slack, act proactively, and learn operating rules from responses. Agents can operate desktop applications and browser sessions, form teams, delegate long-running work to background agents, and record meetings with live transcription. The repository is licensed under the Linger Open Source License, based on Apache 2.0 with additional conditions.
Mate is an open-source starter for a text-first shopping agent. Users can send a store or product link through web chat, the command line, or an SMS adapter and receive demo responses for cashback, discount codes, package tracking, and price-drop watches. The starter parses the message into an intent and routes it to demo cashback, code, tracking, or price-watching providers. It uses placeholder links, DEMO codes, scripted shipment statuses, and a local simulated price-watch job rather than live merchant data; an optional OpenRouter fallback handles unrecognized intents. The repository includes a small Python agent, web chat server, CLI, Twilio webhook stub, documentation, and original icons. It runs on Python 3.10 or newer with no required third-party runtime packages and is licensed under the MIT License. The repository is explicitly a starter and is not the production Mate service or connected to its live merchant network.
Obsidian Skills is a collection of agent skills that teaches compatible coding agents, including Claude Code, Codex, and OpenCode, to work with Obsidian vaults and their open file formats. Its skills create and edit Obsidian Flavored Markdown with wikilinks, embeds, callouts, and properties; create and edit Bases files with views, filters, formulas, and summaries; and create and edit JSON Canvas files with nodes, edges, groups, and connections. Additional skills provide interaction with vaults through the Obsidian CLI, command-line syncing and publishing with Obsidian Headless, clean Markdown extraction from web pages using Defuddle, and Markdown-template rendering from JSON or CSV data with Knap. The skills follow the Agent Skills specification and can be installed through supported agent marketplaces, npx skills, or by copying or cloning the repository into an agent's skills directory.
OpenOrange is an open-source runtime from Rivet for running existing Cloudflare Workers and Durable Objects on AWS, Google Cloud, Azure, or self-hosted infrastructure without changing the Worker code or Wrangler configuration. Its orange CLI reads Wrangler projects: `orange dev` runs Workers locally, while `orange deploy` builds and deploys them to a Rivet cluster. Each Durable Object runs as a Rivet Actor with its own SQLite database; KV, D1, rate limiting, queues, Hyperdrive, and other bindings are supported with varying levels of compatibility, and S3 backs Durable Object storage. The `orange migrate` command moves Durable Objects off Cloudflare while the application continues serving. OpenOrange implements the Workers API on Node.js rather than running Cloudflare's workerd, and is distributed under the Apache-2.0 license.
Open Shelf is an open-source app store for Android TV, Google TV, Fire TV, and other Android boxes. It lists TV-compatible apps from GitHub releases and installs them using only the remote, selecting an APK that matches the device's CPU. Users can browse 18 categories, search by app name or function, view installed apps and available catalog updates, and choose light or dark themes. The catalog is built from public GitHub data and packaged inside the APK. It includes repositories whose latest release contains an APK published within the past three years and whose manifest declares an Android TV launcher entry; non-fork repositories are preferred, with limited exceptions for maintained forks. Open Shelf checks an application's signing key before installing updates and holds an update when the key changes. IPTV players and debrid or Stremio streaming clients are hidden by default, and abandoned applications can link to maintained forks with newer TV builds. The app requires Android 9 or newer and is distributed through GitHub Releases for sideloading. The project is licensed under the GNU General Public License v3.0.
Orwell is an end-to-end encrypted chat and calling platform for communities, developed by a solo developer in Portugal. It provides text rooms, categories, roles, direct messages, account walls, voice and video calls, screen sharing, pinned messages, ticket rooms, device pairing, and self-hosted homeservers. It runs at orwell.pt, in a browser, and as desktop applications for Windows and macOS; a Linux version is in development. The platform uses its Enclave protocol, with MLS (RFC 9420) for rooms and direct messages, X-Wing post-quantum key exchange combining X25519 with ML-KEM-768, and SFrame-style encryption for call media. Messages, metadata such as names and room names, and call media are encrypted on users' devices; the server relays and stores encrypted data but cannot open it. Accounts use key-pair identities rather than email addresses or phone numbers, and channel addresses are opaque. Federation is supported in the code but is disabled by default and has not yet been used between two live servers. Orwell is free software under AGPL-3.0-only and includes the homeserver in the repository. It can be run with Node.js 22 or later and Docker Compose, or used through the hosted service; a commercial licence is available for deployments that cannot comply with the AGPL.
P&ID to Knowledge Graph is a Claude skill by Scott Duncan that reads digital or scanned piping and instrumentation diagrams (P&IDs) and process flow diagrams from PDF files, then produces a process knowledge graph. Claude identifies equipment, pipe runs, valves, instruments, signal lines, and off-page connectors; tested Python scripts prepare pages, build the graph, and perform quality checks. The graph represents upstream and downstream flow, instrument locations and measurements, and control loops from sensors through controllers to valves or drives, including cascades, interlocks, I/P relays, and VFDs. It outputs a readable plant summary, a QA report, JSON, and a DuckDB database for SQL and recursive flow queries, with optional Neo4j or GraphML export. The repository includes ISA-5.1 parsing, DEXPI class mapping, schemas, extraction references, scripts, and pytest tests. It requires Claude Code and Python 3.10 or newer; extraction quality depends on drawing legibility, and the project requires qualified-engineer review before use in design, operations, management of change, or safety studies. Licensed under MIT.
PINE is a reactive harness for AI agents from pmndrs. It lets agents subscribe to event sources such as folders, commands, websites, FTP endpoints, webhooks, schedules, Telegram, and MQTT, then react by performing tasks such as extracting information from new files. Agents can retain memory, use MCP servers and skills, browse the web, operate computers on macOS, write code, and delegate work to parallel sub-agents. Its terminal interface supports interactive controls, images, video, and drag-and-drop; the daemon can run as a background service so agents continue responding without an open terminal. PINE is distributed through npm as @pmndrs/pine and runs on macOS and Linux, with Docker support for other environments.
Prime Files is a Windows file explorer written in Rust and forked from Nohrs. It provides title-bar tabs, a two-pane split view, folder-specific view and zoom settings, and a viewer for images, GIFs, video, audio, PDFs, 3D models, text, Markdown, folders, and archives. It preserves Windows Explorer integration, including thumbnails, the right-click menu, drag and drop, and opening folders from other applications. The application includes whole-computer search, subfolder search, quick filtering, an integrated tabbed terminal, drive storage information, nine languages, and a compact mode that hides the interface until the pointer reaches an edge. It can optionally act as the default Windows file explorer and integrate with Listary. Prime Files is distributed as a per-user Windows installer for 64-bit Windows 10 version 1903 or newer. The installer and releases are provided through the repository, which also supplies automatic updates and SHA-256 checksums. The project is licensed under the MIT License.
Relax is an experimental Rust live compiler and JIT runtime from the Makepad project for development iteration. It runs Cargo packages directly from source without rustc, LLVM, or a linker: functions are type-checked and compiled when first called, hot functions are recompiled through interpreter, O1, and O3 tiers, and changed function bodies are rechecked, recompiled, and swapped into the running process while preserving state. It also provides live test reruns, panic and breakpoint inspection, profiling, flight recording, memory analysis, and a command-line runner. Relax replaces major parts of the usual Rust compilation pipeline with its own lexer, parser, type system, borrow checker, optimizer, register allocator, x86-64 and AArch64 backends, runtime, and cache. Its cache keys results by input hashes, and published machine code remains available while older versions may still be on a stack. It is intended for the edit-run loop rather than release builds; the repository reports slower optimized code, higher memory use, incomplete crate support, and live-patching limitations for changes such as struct layouts, imports, traits, impl headers, and macros. The project is licensed under MIT and has been tested mainly with Makepad, on Linux x86-64 and macOS arm64.
sand3d is a Python library for writing 3D objects as short programs designed for LLMs and human readers. Built over Blender's bpy, it represents profiles, solids, edits, frames, joints, groups, and assemblies as named parts; parts, edges, and faces are addressed by name rather than hand-computed coordinates, and the same program produces the same mesh deterministically. It can export meshes and articulated assemblies to GLB, URDF, OpenUSD, OBJ, and STL for web viewers, game engines, robotics and simulation tools, CAD workflows, and 3D printing. The library is distributed under the Apache License 2.0 and includes example programs, documentation, and datasets of procedural reconstructions.
ScanDot is an experimental Android app for capturing, saving, reopening, and viewing 3D point clouds on phones that support ARCore Depth. It combines ARCore camera tracking with depth estimates; its viewer supports rotation, panning, and zooming, while an optional full-depth mode interpolates missing depth and may distort surfaces. The repository also includes a separate workstation video-photogrammetry pipeline. FFmpeg and FFprobe prepare video frames, and COLMAP performs sparse or dense reconstruction; this processing does not run the ordinary video reconstruction on the Android device. ScanDot is an early prototype: scans can contain gaps, drift, and noise, and the project notes difficulties with reflective, transparent, and textureless subjects. The Android app requires JDK 17 and Android SDK 36 to build, an ARCore Depth-compatible device, and camera permission.
shorts-kit is an MIT-licensed toolkit for producing narrated vertical shorts, product promos, and launch films with a coding agent and Remotion. Its agent skills guide a project through fact verification with a fact ledger, selection of one of 20 storytelling flows, script and voice generation, storyboard and scene construction, rendering, cover creation, and caption preparation; the kit does not publish videos itself. It includes word-timed captions, kinetic typography, maps and a 3D globe, motion graphics, a shot library, optional text-to-speech or recorded-voice workflows, and scripts for rendering and platform-ready encoding. A local dashboard previews progress, narration with spoken words highlighted, storyboards, stills, renders, covers, and captions before approval. The repository supports Claude Code and agents that read AGENTS.md, and distributes CC0 sound effects alongside map and font assets with their stated licenses.
Silica is a lightweight native code editor for macOS, built with SwiftUI and AppKit. It opens projects as folders with file trees, fuzzy file search, project-wide text search, tabs, syntax highlighting for 26 languages, an optional minimap, and Markdown, CSV, and HTML previews; images, PDFs, and videos can open in their own tabs. Projects restore their open tabs, and the app integrates with Finder through Open With, drag-and-drop, and the macOS open command. Silica requires macOS 26 or later, is distributed as a DMG with daily update checks that can be disabled, and is released under the MIT License.
Steel is Skild AI’s GPU-accelerated rigid-body physics solver for scalable robot-manipulation simulation, developed in collaboration with NVIDIA and implemented inside Newton, the open-source physics engine hosted by the Linux Foundation. Its repository provides runnable scenes and solver presets for robot arms, grippers, object stacks, cloth, cables and closed-loop mechanisms, along with documentation, tutorials, an interactive browser implementation, live solver tuning, benchmarking tools and agent workflows for debugging, validating, extending and porting simulation scenes. The solver is distributed as part of Newton and is licensed under Apache License 2.0; the repository’s examples include Apache-2.0 code and BSD-3-Clause robot-model data.
ThreeRT is a real-time path-tracing renderer and game runtime from MeshyAI, with a Three.js-style TypeScript API. Its Rust and wgpu engine traces lighting on the GPU, including shadows, bounced light, reflections, and emissive surfaces; it uses a software BVH where hardware ray queries are unavailable. Scene code and game logic run together in the engine, while the same scene can also run as Three.js in a browser. The project provides the @threert/scene package for meshes, physically based materials, lights, glTF models, physics, and volumes; @threert/gameplay for state, handlers, frame scripts, input, HUDs, and deterministic replayable gameplay; and @threert/voxel for destructible voxel levels with explosions, cuts, and rigid-body debris. ThreeRT can be used through its macOS and iOS clients or built from source, and includes a command-line development workflow for bundling and serving games. The engine and repository are MIT-licensed, while the macOS and iOS clients are free to use and closed source.
TUFFUI is a free, MIT-licensed collection of React components and complete screens for AI chat and agent interfaces. It includes UI for streaming conversations, tool-call approvals, agent plans and traces, memory, usage meters, agent-data charts, and other parts of AI products that generic UI kits do not cover. The components are distributed through a shadcn registry: installing a component copies its source into the user's project, where it can be edited directly. Screens are composed from the components and previewed at desktop, tablet, and phone widths using container-query layouts. The repository includes the registry build, component documentation and demos, scripted mock runs for previews, and development commands for building, linting, type-checking, and formatting. The project contains 76 components according to its README, has no Pro tier or signup requirement, and is available for personal and commercial use under the MIT license.
UI SDK is a Vercel Labs library of reusable behaviors for building user interfaces, including keyboard navigation, focus recovery, selection, dismissal, pop-ups, drag and drop, and the ARIA attributes required for them. Components are defined from these behaviors as primitives and rendered through framework adapters for React, Vue, Svelte, Solid, and Web Components. The project also provides a skill that coding agents can use to build UI SDK components. It is currently in beta, with APIs, documentation, and behavior subject to change under Vercel's beta terms.
VideoHub is a local video review studio and render queue for films that Claude Code or OpenCode generates as code. It plays a film directly from its web page, allowing frame-accurate playback, frame annotations and review notes before rendering; the agent reads those notes through MCP tools, edits the film code, and updates the live preview. Its render queue processes one render at a time and can render films in two-second chunks, re-rendering only the sections touched by review notes. It also provides project and version management, sound controls, thumbnail workspaces, render protection, cleanup tools and hardware-encoder selection through FFmpeg. The software runs on Windows, Linux and macOS, requires Node.js 22 or later, FFmpeg and a native Chromium-based browser, and is distributed under the MIT license.
Waymark is a local-first, Git-friendly work tracker that stores each work item as a Markdown file with YAML frontmatter under a project's work/ directory, while configuration lives in .waymark/. It supports work-item hierarchies, dependencies, customizable fields, and Git-based history, so changes can be reviewed, diffed, and reverted alongside code. The CLI builds an in-memory projection from the files for collection queries, verifies it against content fingerprints, and refreshes it incrementally without re-parsing unchanged documents. Responses record their source, completeness, and freshness; incomplete projections are rejected rather than published. The included waymark-mcp server uses Model Context Protocol over stdio so MCP clients such as Claude Code and Cursor can read and write work items. An optional Web package provides a read-only browser view bound to loopback. Waymark requires no database, daemon, or runtime network access. Prebuilt basic and Web packages are available for Linux, macOS, and Windows, and the Rust crates can be installed from crates.io. The project is pre-1.0, and the CLI and document format may change between releases. It is dual-licensed under the MIT License or Apache License 2.0.
xj-scanner is a GPLv3 hobby diagnostic tool for reading 1991–95 Chrysler and Jeep OBD1 engine computers from a laptop through an Arduino Nano bridge. Its Python scanner reads fault codes and live engine data, logs drives, checks the computer's part number, and runs built-in actuator tests; a DEMO mode simulates a Jeep for testing the laptop-side software. Claude Code, Codex, or another coding agent can use the repository's step-by-step playbooks to guide software setup, breadboard wiring, connector testing, and the first reading. The project was built and tested on a 1994 Jeep Cherokee XJ with a 4.0L engine computer, and warns users to measure their own diagnostic connector before connecting hardware.
Yogurt is a free, open-source local AI meeting-notes application from Frigade for macOS, Windows, and Linux. It records microphone and system call audio without joining meetings as a bot, produces live and saved transcripts, combines transcripts with user notes to generate summaries, and answers questions across meetings with links to source material. It transcribes speech with whisper.cpp, or Apple SpeechAnalyzer for live words on supported Macs, and runs summaries and questions through local open-weight models using llama.cpp. Meeting recordings, transcripts, notes, settings, and models remain on the computer after setup; the application goes online only for model downloads and update checks. It includes folders and search, Markdown export, call-recording reminders, automatic stopping when a call ends, speaker identification for Zoom, Teams, and Meet on Mac, and an optional read-only local MCP server for assistant applications. Yogurt is distributed as an Electron application for Apple Silicon Macs running macOS 14.4 or newer, 64-bit Windows 10 version 2004 or newer, and 64-bit Linux systems such as Ubuntu 22.04 or newer. It requires at least 8 GB of memory, needs no account, and is licensed under the MIT License.
ZigZag is a free, open-source visual embroidery editor written in Rust for macOS, Windows, and Linux. It imports SVG artwork and images, converts vector shapes into embroidery stitches, and lets users edit paths, thread colors, stitch types, sewing order, machine settings, and hoop layouts. Its stitch engine supports tatami, satin, running, contour, blanket, appliqué, manual, meander, ripple, cross-stitch, and tartan techniques, with underlay, pull compensation, auto-splitting, routing, overlap removal, cap order, and digitizing checks for issues such as unsuitable widths, small lettering, gaps, and hoop fit. The editor can open and export common machine formats including PES, DST, EXP, JEF, VP3, and PEC, produce HTML worksheets, and send designs over Bluetooth to a Brother Skitch machine. It also includes a command-line tool for conversion and inspection. An MCP server and bridge allow Claude Code, Codex, Claude Desktop, and other MCP clients to create, inspect, modify, preview, check, save, export, and send designs; agent changes operate on the live document and can be grouped into atomic batches.
Searchable transcript of GitHub Trending Weekly #59: open-orange, obsidian-skills, sand3d, yogurt, agentes-chat, zigzag — Github Awesome (15:40). 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 59. 35 trending open source projects on GitHub right now. Let's go. Open orange is a runtime from Rivet for running Cloudflare workers on your own servers or another cloud. It reads your existing Wrangler config. So orange dev and orange deploy replace the usual commands. Each durable object becomes a rivet actor with its own SQLite database and KVD D1 and rate limiting work as dropin bindings.
00:30 The team says it can migrate durable objects off Cloudflare with no downtime. Obsidian skills teaches coding agents how to work with Obsidian vaults and their file formats. It covers markdown features like wiki links and callouts bases with filters and formulas and canvas files with connected nodes. Agents can also use the Obsidian CLI to interact with a vault or extract clean markdown from web pages with Defut.
00:56 The skills work with compatible agents including claude code and codecs. UI SDK is a library from Versell Labs for building interfaces from reusable behaviors. Its core handles keyboard navigation, focus, and selection along with interactions like pop-ups and drag and drop including the Arya attributes they need. You combine those behaviors into a component, then render it through an adapter for your framework.
01:23 It supports React, Vue, Spelvelt, Solid, and Web Components, and includes a skill for coding agents. Server OS panel is a web dashboard for managing Linux servers across hosting providers. Each server runs a small Damon that connects to the panel where you can deploy apps from Git, configure domains and HTTPS, and inspect resource usage. It also brings backups, uptime alerts, and service logs into one place.
01:53 Workflows can run on a schedule or a push, and an MCP server lets your assistants work with your servers. San 3D is a Python library for building 3D objects as short programs designed for language models to write and people to read. It runs on Blender's Python API with commands for shapes, cuts, and assemblies. Parts, edges, and faces can be referenced by name, and joints describe how components move.
02:16 You can export the same model as a GB for a viewer, a URDF for robotics, or an STL for printing. Yogurt is a meeting notes app for Mac, Windows, and Linux that records your microphone and call audio. It transcribes the conversation and combines it with your own notes to write a summary using models that run on your computer. After downloading those models, transcription and summaries work offline.
02:43 You can ask questions across past meetings, follow links back to the source, and export your notes as markdown. Agent's chat is a team chat app you can host yourself where AI agents join channels and direct messages alongside people. Each agent gets a role, an owner, and a budget with access to MCP tools and a computer for running tasks. You can schedule routines or trigger work from events.
03:09 Actions such as merges and deployments wait for approval in the chat where your team can review them. Zigzag is an embroidery editor written in Rust that turns SVG artwork into stitches for embroidery machines. You can adjust paths, thread colors, and stitch settings visually, while built-in checks flag problems like tiny lettering or a design that won't fit the hoop.
03:31 An MCP connection lets AI agents create and edit designs alongside you. Export machine formats such as PEES and DST or send directly to a Brother sketch over Bluetooth. Agent Run is a platform for managing agents built with tools like Clawude Code, Codeex, and Open Code. It tracks runs so work can resume after a server restart with approvals and spending caps in one console.
03:57 After a run, an agent can propose notes, skills, or reusable workflows for review and storage. Future runs start with that saved knowledge, and you can inspect what was kept in the learning tab. Scan DOT is an experimental Android app that captures 3D point clouds using our camera tracking and depth estimates. On a supported phone, you move around a subject, save the scan locally, and inspect it with rotation, pan, and zoom controls.
04:26 results can still have gaps, drift, and noise. The repository also includes a separate workstation workflow that extracts frames from video and reconstructs them with coal map. Island CN is a React component that brings a dynamic island style activity panel to your app. It can show upload progress, a timer, or music controls with compact and expanded views that animate as activities change.
04:50 You install it through Shad CN, then start, update, and end activities from your components. A playground lets you try presets, tune the motion, and copy the code for your own interface. Minia H3 image is Miniax's new model for both generating and editing images in one checkpoint. It outputs around 16 megapixels natively up to 4,96x 4096 without an upscaler.
05:16 The diffusion transformer has 33 billion parameters and is distilled down to just six sampling steps. That's a lot of pixels for six steps. Editing takes up to 16 reference images and you can serve it locally with diffusers or comfy UI. Pine is an agent runtime that reacts to events from folders, websites, and web hooks. You can ask it to watch a scans folder and extract details whenever a new document arrives.
05:44 Agents keep memory. use MCP tools and delegate work to other agents. It runs on Mac OS and Linux with a terminal interface for chat and coding, plus a background service that keeps agents responding when the terminal is closed. XJ Scanner connects a laptop to an older Jeep's engine computer through an Arduino Nano. It reads fault codes and live engine data with logging tools and playbooks that help a coding agent guide you through setup.
06:13 The project targets Chrysler and Jeep OBD1 computers from 1991 to 1995, though it's been tested on one 1994 Cherokee. A simulated Jeep mode lets you check the laptop software before connecting hardware. Haru is a multilingual voice conversation model that listens and speaks at the same time. A text card sets its character while a reference recording sets the voice.
06:40 The repository includes a browser demo, a PyTorch implementation, and an MLX version for Apple Silicon, plus serving through VLM Omni. It's a researchbased model, so its focus is natural conversation. Clad is a painting and animation app where you and a coding agent work on the same drawing. You paint in the editor while the agent uses command line tools with the same brushes and layers.
07:04 Every stroke keeps its path, pressure, and color, so you can reshape or recolor individual marks later. You can pin feedback directly on the canvas, then export finished animations as sprite sheets or into goat. Silica is a native code editor for Mac OS built with Swift UI and AppKit with a glass style interface. Open a project folder and you get a file tree, fuzzy file search, and syntax highlighting, plus a minim for navigating longer files.
07:36 Markdown has its own preview with highlighted code blocks. It requires Mac OS 26 or later and restores your project tabs when you return. Waymark is a work tracker that stores each task as a markdown file inside your repository. It supports task hierarchies, dependencies, and custom fields with Git handling the history so you can review or revert ticket changes alongside code.
07:59 You can manage work from the command line and an MCP server lets coding agents read and update the same tasks. An optional browser view gives you a readonly project overview. Waiting 3 minutes for a Rust build kills any feedback loop. Relax from the MakePad team runs a cargo package straight from source with no RSC c LLVM or linker involved. It compiles each function on first call, optimizes hot code in the background, and patches changed functions into the running program without losing state.
08:31 The team reports a large 3D game reaching its first frame in 11 seconds on an M4 MacBook. GitHub Cool ReadMe generates animated SVG stats cards for your profile using a scheduled GitHub action. It can include private and organization contributions while showing only totals for private work. Cards cover activity, language usage, and notable pull requests with classic dot matrix or OLED styles.
08:57 The generated images live on a separate branch in your profile repository. And you can adjust their colors and which statistics appear. OpenShelf is an Android TV app store that installs apps directly from their GitHub releases using your remote. You can browse categories or search by what an app does and it picks the APK that matches your TV's processor.
09:21 It also tracks updates for its apps in its catalog and checks their signing keys before updating. If a key changes, it holds the update for review. Linger is a personal agent app that can run models locally on supported Macs and PCs. It works across files and apps you authorize with searchable wiki pages that preserve preferences and decisions between conversations.
09:45 You can have it monitor your inbox, calendar, and Slack, then bring in other agents when a task needs them. In local mode, its memories stay on your device. Mate is an open-source starter for a text first shopping agent written in plain Python with no required dependencies. You text it a store or a product link over web chat, a command line or SMS, and it answers with cashback offers, discount codes, package tracking, or a price drop watch.
10:17 Everything here runs on labeled demo data. So, it's a template to build on rather than the live service. Creator generalist is a skill pack that guides coding agents through 2D animation, 3D scenes, and game development. It connects production steps with tools like Blender, 3JS, and go do covering assets, animation, and sound. Before building the final piece, the agent shows style frames and a preview for feedback.
10:43 For games, it starts with a playable prototype, then checks things like restart behavior, performance, and paths that could leave players stuck. PNID to knowledge graph is a claude skill that reads piping and instrumentation diagrams from PDFs and turns them into a process knowledge graph with tested Python scripts doing the prep and QA. It maps what flows upstream and downstream of each piece of equipment which instrument measures what and every control loop from sensor to valve.
11:16 Orwell is an end-to-end encrypted chat and calling app for communities built in Typescript with React and Electron by a solo developer in Portugal. It has text rooms, voice and video calls with screen sharing and direct messages and the server can't read your messages, names or call audio. Encryption runs on MLS calls use S-frame and key exchange is postquantum.
11:40 Tough UI is a collection of React components for building AI chat and agent interfaces. It covers streaming conversations, tool approval cues, and agent plans, plus views for memory and usage. Components install through a Shad CN registry, which copies the source into your project so you can edit it directly. The collection also includes complete screens assembled from those components with previews at desktop, tablet, and phone sizes.
12:08 Steel is a GPU physics solver from Skilled AI developed with NVIDIA for simulating robot manipulation. It runs inside the Newton physics engine with examples covering robot arms, grippers, and stacks of objects. This repository provides scenes, tutorials, and workflows for tuning simulations and investigating unstable behavior. You can open examples in an interactive browser, adjust solver settings live, and benchmark the scene before using it in a robot learning environment.
12:39 3RT is a rendering and game runtime from Meshy AI with a 3JS style TypeScript API. Its Rust and WGPU engine uses GPU ray tracing for shadows, reflections, and bounced light, including surfaces that glow and illuminate the scene. You write scene and game logic together with packages for physics and destructible voxal worlds. Shorts kit is a motion toolkit that guides coding agents through making narrated videos.
13:05 It covers factchecking scripts and voice timing, then build scenes with captions, maps, and motion graphics. You can use a generated voice or bring your own recording. A local dashboard lets you review the storyboard, listen to the narration, and inspect still frames before rendering with the finished video, cover, and caption prepared for upload. Prime Files is a Windows file explorer written in Rust with tabs and a split view for working across folders.
13:34 Its built-in viewer handles files, including PDFs, videos, and 3D models, while keeping Windows thumbnails and the familiar right-click menu. There's also a terminal with tabs and search across your computer. Compact mode hides the surrounding controls until your pointer reaches the edge, giving the file list more room. Watched shows which satellites passed over any place on Earth in the past week and whether any of them took a public photo of it.
14:03 It's built with React and 3JS, and the orbit math runs in your browser with SGP4. Search a city or your own location and it checks nearly 400 imaging, weather, and radar satellites, then matches Sentinel 2 and LANCAT photos to the past that captured them. VideoHub is a local review studio for videos built as code by Claw Code or Open Code. It plays the film directly from its web page, so you can step through frames and draw notes before rendering.
14:34 Your agent reads those notes through MCP, fixes the code, and updates the preview. A render queue processes one job at a time, and chunked projects can rerender just the sections you changed. GPU canvas is a TypeScript and React library that puts thousands of lived DOM nodes on an infinite canvas drawn by WebGPU in a single instance pass. Pan and zoom never touch the DOM, yet every node stays live for input, state, and accessibility.
15:03 You can run WGSL shader effects like glass, frost, or glitch on real components, and agents get a JSON view of everything on screen. Folio is Omar's file manager built with C++ and cutequ. It follows your desktop theme and supports keyboard navigation across grid, list, and finder style column views. Press space to preview files or split the window to move items between folders. Plugins add features like browsing machines on your tailscale network over SFTP and sharing through local send.