Searchable transcript of Trending Open-Source Github Projects : Macro, holaOS, Switchyard, Axolotl, Octop & MiniMax H3 #284 — ManuAGI - AutoGPT Tutorials (16:29). 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 ManuAGI - AutoGPT Tutorials. 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 Every week, developers release powerful open-source GitHub projects, and staying on top of them can be hard. This is your weekly GitHub project update video, covering the top trending open-source GitHub projects this week, from local first agent workspaces to fine-tuning frameworks and whiteboard SDKs. Without wasting time, let's get started. >> Before we jump into today's project updates, here's a quick announcement for everyone.
00:24 We've launched a brand new YouTube channel called AI Agent Studio dedicated entirely to AI Agent projects, tutorials, and tools. So, if you're interested in staying up to date with the latest AI agent open source projects, learning how to build your own agents, or exploring cuttingedge agent frameworks, make sure to check it out. Subscribe now to get weekly videos, in-depth guides, and real time project breakdowns.
00:50 The link is right there in the description. Don't miss it. All right, let's get into today's video. >> Project number one, Macro. Open-source unified workspace with shared team memory. Macro is an open-source unified workspace for teams that brings email, messages, tasks, docs, calls, CRM, and agents into one keyboard first app. Everything connects through birectional at linking.
01:15 So, a mention links two items both ways. A shared team memory lets agents recall work across all of it. and one inbox splits into Signal and Noise built-in Rust and SolidJS. It replaces Slack, Linear, Notion, and Superhum, and connects to Claude Code over MCP. Sign up and unify your team. Project number two, Neolite. A browser you and your agents share.
01:38 Neolite is a free browser where you and your AI agents work in parallel. Each agent runs its tasks in its own space, working in the background while your tabs stay yours, reaching your login through a helper, ego browser. It exposes the page as JavaScript tools, snapshot, click, fill. So any agent CLI, claude code, codecs, or cursor, writes one snippet instead of step-by-step calls.
02:03 It inherits your Chrome login and runs on Mac OS. Download it and browse alongside your agents. Project number three, Hola OS local first workspace to run any AI agent. Hola OS is an open-source local first workspace where you and your AI agents share one computer. Run any agent, claude code, codecs, or the built-in one side by side, sharing one memory, tools, and apps.
02:29 That memory lives on your machine as plain files. So, an agent resumes where you left off. Frontier models are built in or bring your own keys. Apps open as surfaces beside the agent and 100 plus integrations and MCP servers connect instantly. Install it and share your computer. Project number four, Switchyard. Rust proxy that routes and translates LLM traffic.
02:52 Switchyard is an open-source Rust proxy and library for LLM traffic from Nvidia Nemo. It translates between OpenAI chat, anthropic messages, and OpenAI responses. So clawed code or codeex keeps its native API while requests are served by VLLM, NVIDIA NIM, a Lama or any OpenAI compatible endpoint. It routes with classifier, stage and escalation strategies or your own and records Prometheus metrics.
03:18 Run it as a launcher, standalone server or Rust library. It's pre-alpha. Install it and route your traffic. Project number five, Booktoskill. Turn any book into an ondemand agent skill. Booktoskill is an open-source agent skill that turns any technical book, folder, or documents into a skill your agent loads on demand. You point it at a PDF, EPUB, DOCX, or folder, and it distills it into frameworks, decision rules, antiatterns, and per chapter files, structure, not a summary.
03:50 Your agent reads the relevant chapter and answers from real content so it doesn't hallucinate. It runs locally and works in clawed code, copilot CLI, and AMP. Install it and study any book. Project number six, Open Work. Open-source alternative to Claude Co-work for sharing workflows. Open Work is a free open-source desktop app for sharing AI workflows.
04:12 An alternative to Claude Co-work that runs on Mac OS, Windows, and Linux. You add one open work mcp to codeex clawed code cursor or another compatible agent and reuse the same skills mcps and connected services like Google workspace and Microsoft 365 across your tools teammates and machines. The desktop app is optional and an admin console lets organizations publish capabilities and manage access.
04:41 Install it and share your first work. Project number seven, dossier liver. Lightweight desktop GIS running maps in the browser. Geo Liber is a desktop GIS application built with Tari React and Maple Liber GLJS that lets people view and edit geospatial data without heavy installed software. It runs a maple liver map locally, loads gson, geoaret, geo package and shape file layers and uses duct dbw for spatial processing in the browser or desktop shell.
05:09 A plug-in system adds base maps, street view, lidar, and editing tools with an optional Python sidecar for heavier tasks. It suits developers and analysts who want a fast, extensible mapping tool. Explore the repository to try it locally. Project number eight, Axelottle, open-source framework for fine-tuning large language models. Axelottle is an open-source framework that streamlines post-training and fine-tuning for large language models.
05:38 It supports many models from the hugging face hub [clears throat] plus multimodal vision and audio models using methods like full fine-tuning Laura Qura quantization aware training preference tuning and reinforcement learning. A single YAML configuration file drives data set pre-processing training evaluation and inference. It runs on NVIDIA or AMD GPUs locally or through Docker images and cloud platforms.
06:02 It suits researchers and engineers customizing LLMs efficiently. Start your first fine-tune with the quick guide. Project number nine, PNDFCN. Ready to use PDF components for React apps. PDFCN is a component library that gives React developers ready-made building blocks for generating PDFs so they don't have to design PDF layouts from scratch. It works through Tacumi and Forme as rendering engines and follows the Shad CN/ISOI registry format, letting developers copy components directly into their codebase using the
06:35 same CLI workflow. It includes tables, forms, charts, invoices, and report blocks styled to match shared themes. This matters for developers building document heavy apps quickly. Add it to your project and start composing documents. Project number 10, Loom Feed, self-hosted community platform built for AI agents and humans. Loomfeed is a self-hosted social platform similar to Reddit that treats AI agents as full participants alongside people.
07:02 It runs on a Go backend. Next JS front end, Postgress and Reddus deployed together through Docker Compose. Agents connect over REST, MCP, or the A2A protocol, then post, comment, and vote using API keys and reputation scores. Every agent generated post carries provenence data, source citations, and epistemic status labels while only humans can verify content.
07:29 This suits developers building communities where agents and humans research together. Clone the repository and self-host your own instance. Project number 11. Via Snapshane native screenshot and annotation tool for Whan desktops. ESNAP is a native screenshot and annotation editor built for Whan desktops running Hyperlind with EA as its main target. It is a desktop tool that captures the focused monitor through Wayland image copy protocols then opens an overlay editor where annotations exist as movable resizable vector
08:01 layers. It offers region, window, and full screen capture, arrows, text, freehand strokes, OCR text recognition, and undo history, all processed locally. Output copies to the clipboard or saves as PNG files. It suits Linux users on Hyperlind who want fast native screenshot editing. Clone the repository and install it with the EA installer. Project number 12, Paristoric Animal Museum, a calm 3D dinosaur museum for children and parents.
08:30 Preistoric Animal Museum is a web application that lets children explore 18 preistoric animals in a calm, interactive 3D space. It is a browser-based museum where a child can turn a model with a finger or mouse, zoom in, and listen to short English or Mandarin narration that never plays automatically. Each exhibit includes a parent guide covering when the animal lived, fossil regions, size, diet, and classification.
08:55 It runs entirely with static assets and makes no calls to AI, advertising, or analytic services and needs no signin. This suits parents and young children who want gentle screenlight learning together. Visit the museum and look closely at one animal at a time. Project number 13, Scene Card Studio turns personal photos into structured visual narratives.
09:17 Scene Card Studio is a Python command line tool that turns personal photos into editable narrative decisions instead of applying a simple style filter. It separates what a photo shows from how it should be interpreted, recording observations, tenative themes, and story roles as editable scene cards. A compiler then converts these cards, a chosen narrative system, and an expression profile into a versioned prompt contract for image generation with retries and reviews tracked through a hash chain.
09:46 Cloud image generation only runs after explicit consent. Otherwise, everything stays local. It suits people building structured personal photo stories. Try it on your own photos with the quick start commands. Project number 14, Kirro Crew, a persistent AI agent workspace that keeps working across sessions. Kirk Crew is an open- source development workspace that runs an AI agent continuously on hardware you control instead of ending when a chat window closes.
10:15 It works through a gateway process that connects to a desktop app, web dashboard, CLI, or messaging tools like Slack and Discord, driving Kira, CLI sessions that persist memory, schedules, and task checkpoints across restarts. It handles longunning tasks, scheduled jobs, and parallel sub aents locally or on a remote host you choose with tool approvals and sandboxing at the runtime level.
10:39 This suits developers who want an agent that keeps context and improves over repeated work. Set it up and give it a task to run unattended. Project number 15. Quick Draw. Free infinite canvas whiteboard SDK for apps. Quickdraw is a whiteboard SDK that drops a complete drawing surface into React, React Native or plain JavaScript apps with no license fee for commercial use.
11:03 It is a library, not a hosted service. So it runs directly inside a developer's own app and stores its document as a flat map of records with every change emitted as a JSON diff for sync or persistence. It includes pressure sensitive pen strokes, shapes, arrows, sticky notes, image handling, and an infinite pannable canvas, all with zero runtime dependencies.
11:25 Developers bring their own transport for real-time collaboration. Since diffs are sync ready, but no server is bundled. This suits developers who want a whiteboard feature without paying a licensing fee. Install the package and drop the board into your app. Project number 16, MiniAX H3. Openweight model that generates video with synced audio. MiniAX H3 is an openweight omnimodal generative model that produces video together with native stereo audio from text, images, video or audio inputs.
11:56 It is a 33 billion parameter transformer that encodes each modality with dedicated encoders and jointly predicts video and audio latence in one pass so sound stays synchronized with motion. The release checkpoints run locally through frameworks like SG Lang, VLLM diffusers or Comfy UI producing clips up to 15 seconds at resolutions up to 768p with a hosted API available for 2K upscaling and prompt prep-processing.
12:24 It supports text to video, image to video, and multi-reference generation guided by images, video clips, or audio. This suits developers building video generation tools who want an open model to run or fine-tune themselves. Download the checkpoints and try it with your preferred inference framework. Project number 17, Gemma Translator. Offline voice translator built for a physical twoperson kiosk.
12:49 Gemma translator is a fully offline voice translation appliance that runs entirely on local hardware like a Raspberry Pi. It combines lighter TLM running the Gemma model. Moonshine for speechtoext and texttospech, a Python API server, and a React front end built for small kiosk screens. Two people facing each other each get a lane with a rotating language selector.
13:12 Pressing a key records speech, which is transcribed, translated, and spoken back in the other person's language, all without an internet connection. It ships with 3D printable case files and a one-comand Raspberry Pi deployment script. This suits builders creating physical translation devices for face-to-face conversations. Set it up on a Raspberry Pi and start translating deviceto.
13:34 Project number 18, a claude code skill that directs cinematic websites. AUR is an agent skill for Claude code that guides an AI coding agent through building a website the way a film director plans a shoot. It scouts award-winning sites for real fonts, pallets, and motion budgets. Locks a single art direction into a written commit sheet before any code is written, then generates imagery and geometry through local CLIs like Blender or image generation tools.
14:02 The build uses one WebGL context with scroll-driven motion, and nothing ships until it passes a zero dependency llinter and a playright motion test that catches dropped frames, generic AI gradients, and accessibility gaps. It works with Claude Code and other agents that read a skill.md file. This suits developers who want their AI built sites to avoid generic templated design.
14:26 Install the skill and ask your agent to build a cinematic landing page with it. Project number 19, Malin's Codeex router. Lets Codeex use Kimmy and DeSseek models alongside GPT. Codex router is a local proxy that lets Codex's normal model picker offer Kimmy and Deepseek models without disturbing native GPT models or an existing chat GPT signin. It runs as a background service on local host, translating requests between codeex's responses API and the chat completions APIs that Kimmy and DeepSeek expose.
14:57 So streaming tool calls and compaction traffic work correctly. Setup is guided, detects existing credentials, takes a roll back snapshot before touching configuration, and only adds a small marked block to Codeex's config file. API keys are entered through a hidden terminal prompt and never printed. This suits Codeex users who want access to alternative models through the same interface.
15:19 Run the guided installer and pick your providers from the model picker. Project number 20, Octtop. Self-hosted, multi-user, multi-agent AI assistant. Octtop is a self-hosted AI assistant platform that runs a single local process serving a web dashboard, a CLI, and IM channels like FishU, Ding Talk, QQQ, Discord, and we come. Each user gets multiple agents drawn from a built-in expert library or given one of 16 MBTI style personas with their own workspace, providers, and cron jobs.
15:50 It stores all conversations, credentials, and agent memory in a local SQLite database under the user's home directory with pluggable storage backends like local disk, Docker or S3. It supports OpenAI compatible APIs, Dash, Scope, and Alma as model providers, plus OOTH and MCP connectors and birectional integration with coding agents like Claude Code and Codeex.
16:15 This suits households or small teams wanting a private self-hosted assistant instead of a cloud service. Install it and run the setup wizard to create your first agent. Thanks for watching.