An open-source application for running compressed Bonsai language models locally on modest hardware, with a built-in server or web UI.
Behind this: 3 things the video never answers.
A dataset product containing 1,324 exercises with structured metadata, bilingual instructions, images, and animation GIFs, plus browser tools for searching and generating database or API code.
Behind this: 3 things the video never answers.
A platform for creating, scheduling, coordinating, and receiving reports from teams of AI agents, with editable memory, model access, skills, and MCP-compatible plugins.
Behind this: 3 things the video never answers.
An open-source research environment that converts finance questions into analysis using market-data loaders, backtesting engines, reports, persistent memory, and multi-agent teams.
Behind this: 3 things the video never answers.
An open-source system that processes books, videos, podcasts, interviews, and courses into structured, executable agent skills rather than summaries.
Behind this: 6 build steps · 2 tools and how each is used · how to validate demand · 3 things the video never answers.
A reusable GitHub Actions component that places repository code into a workflow workspace and supports commit-only, full-history, sparse, partial, submodule, LFS, and multi-repository checkouts.
Behind this: 1 tool and how each is used · 2 things the video never answers.
An open-source rigid-body physics and collision-detection library supporting continuous collision detection, constraints, vehicles, ragdolls, soft bodies, GPU hair, buoyancy, deterministic simulation, and multiple language bindings.
Behind this: 1 more real example · 3 things the video never answers.
An open-source workshop and public template teaching agentic software-development workflows through markdown lessons and an optional browsable site.
Behind this: 5 build steps · 5 tools and how each is used · how to validate demand · 2 things the video never answers.
An extension that lets a coding agent delegate reviews, research, planning, implementation, and other tasks to foreground or background child-agent sessions.
Behind this: 1 tool and how each is used · 3 things the video never answers.
A command-line tool that runs an AI coding agent inside a fresh Linux virtual machine with host code mounted, restricted outbound traffic, forwarded SSH access, and disposable virtual-machine state.
Behind this: 2 tools and how each is used · 3 things the video never answers.
An MCP server that replaces repeated artifact pasting with short tokens resolved into model-specific views, while recording attribution and reads and propagating corrections.
Behind this: 5 tools and how each is used · 3 things the video never answers.
A local visualization tool that replays coding-agent logs on a 3D map of a repository, highlighting files searched, read, and edited over time.
Behind this: 2 tools and how each is used · 3 things the video never answers.
An Electron workspace combining AI tasks, code, terminal, browser, Git, models, and memory while treating model output as untrusted input and requiring policy-controlled approval for sensitive actions.
Behind this: 3 tools and how each is used · 3 things the video never answers.
A free Windows desktop application that identifies clickable screen elements with computer vision and assigns short keyboard labels for mouse actions.
Behind this: 3 things the video never answers.
An open-source TypeScript web engine that converts 3D scenes into pencil-style SVG drawings using geometry-derived strokes, exact silhouettes, hidden-line handling, curvature-based hatching, and stable animation wobble.
Behind this: 3 things the video never answers.
An agent skill that reads a repository, identifies its clearest value and proof, and redesigns its README using SVG visual identity with searchable, copyable markdown explanations.
Behind this: 6 build steps · how to validate demand · 3 things the video never answers.
A developer hub for generative Arabic speech-recognition models with model cards, benchmarks, and copy-paste inference for Arabic, code-switched Arabic-English, and additional languages.
Behind this: 2 tools and how each is used · 3 things the video never answers.
An open-source TypeScript library that lets SaaS platforms add customer hostnames, return required DNS and certificate records, track readiness, and remove domains safely through provider adapters.
Behind this: 4 tools and how each is used · 3 things the video never answers.
A source-available infrastructure platform that connects alerts, evidence, topology, human approval, controlled remediation, recovery verification, and learning into an auditable loop.
Behind this: 8 build steps · 8 tools and how each is used · how to validate demand · 3 things the video never answers.
An open-source personal model that learns from permissioned macOS activity, builds an evidence-backed layered representation of a person, and exposes grounded context to AI agents through MCP.
Behind this: 1 tool and how each is used · 3 things the video never answers.
Searchable transcript of Top Open-Source GitHub Projects : LobeHub, clawk, Jolt Physics, Clodex, mindwalk & Persome #276 — ManuAGI - AutoGPT Tutorials (14:54). 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 new open-source projects on GitHub, and this is our weekly GitHub project update video covering the best of them. Top trending open source GitHub projects. This week, you'll discover useful and trending developer tools from local AI models to agent workflow tools. Without wasting time, let's get started. >> Before we jump into today's project updates, here's a quick announcement for everyone.
00:22 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 realtime project breakdowns.
00:48 The link is right there in the description. Don't miss it. All right, let's get into today's video. >> Project number one, Bonsai demo. Run one bit and turnary Bonsai LLMs locally. Bonsai demo is an open-source repository that lets you run Prism ML's Bonsai language models locally. Bonsai is a one-bit model and Turnary Bonsai, a 1.58 bit model. Each offered in 8B, 4B, and 1.7B sizes, so compressed weights fit modest hardware.
01:17 One setup command installs dependencies, pulls the models from hugging face and fetches or builds binaries. It runs on Mac with Metal Linux and Windows with CUDA, Vulcan or Orochem, and on plain CPU in GGUF through llama.cpp or an MLX on Apple silicon. You chat through a built-in server or open web UI. Clone it and run a tiny model locally. Project number two, exercises data set.
01:40 1,324 fitness exercises with data, images, and GIFs. Exercises data set is a readytouse collection of 1,324 fitness exercises for building workout and health apps. Every entry carries a name, body part, target, and secondary muscles, equipment, bilingual English and Turkish instructions, a thumbnail image, and an animation GIF stored as a single JSON file with matching media folders.
02:08 It ships with two browser tools that need no server. One to search and filter every exercise and one that generates database tables, insert statements, and copy-paste API code in languages like Python, JavaScript, and Go. It's licensed for educational non-commercial use only. Clone it and prototype your fitness app. Project number three, LoHub. Self-hostable platform to run your AI agent team.
02:31 LobHub is a self-hostable platform that turns your AI agents into a managed team. It calls itself a chief agent operator, letting you hire, schedule, and get reports from agents that run around the clock. An agent builder sets one up from a plain description. And each agent can reach any model, plus a library of 10,000 skills and MCP compatible plugins.
02:51 Agent groups let several work together on shared pages, projects, and schedules, while a white box memory learns how you work and stays editable. You deploy it on Versell, Alibaba Cloud, or Docker. Spin it up and manage your agents. Project number four, Vibe Trading. Open-source natural language agent for trading research and back testing. Vibe Trading is an open- source research workspace that turns plain language finance questions into runnable analysis.
03:16 Installed with one command, it connects your prompts to market data loaders, back test engines, reports, and persistent memory. You can ask a question, back test a strategy idea, or run multi-agent teams for investment, quant, and risk work. pulling data across equities, crypto, futures, and forex through free sources with automatic fallback. A shadow account feature reads your broker journal and compares your trades to your extracted rules.
03:41 It targets research and simulation, trading only through a broker you authorize and can halt. Install it and start researching. Project number five, Kangi Skill. Distill books and videos into callable agent skills. Kangji skill is an open- source pipeline that distills high-v value content, books, long videos, podcasts, interviews, and courses into callable agent skills instead of summaries.
04:03 It runs a seven-stage method called Riadal TV++. It first reads the whole source, then sends five parallel extractors after frameworks, principles, cases, counter examples, and terms. Each candidate must pass a triple check for cross support, predictive power, and uniqueness, so only a fraction survive. Survivors become structured skills with quotes, rewrites, examples, executable steps, and boundaries.
04:25 Get linked and stress tested, then install into clawed code or cursor pointed at a transcript and turn knowledge into tools. Project number six, and new checkout v7, GitHub action that checks out your repo and workflows. Checkout is a GitHub action that pulls your repository into a workflow so later steps can use the code. You add it as a uses action/checkout step and it clones the repo into the workspace fetching only the triggering commit by default or full history on request.
04:52 It handles sparse and partial checkouts, subm modules, git lfs and several repositories at once and sets up token or SSH off for git before cleaning those credentials afterward. The newest version also refuses to run untrusted fork pull request code unless you opt in. Add it to your workflow and start from clean code. Project number seven, Jolt Physics.
05:13 Multi-core C++ rigid body physics engine for games. Jolt Physics is an open- source multi-corefriendly rigid body physics and collision detection library written in C++. It's built for games and VR and ships in titles like Horizon Forbidden West and Death Stranding 2. It simulates rigid bodies with continuous collision detection plus constraints, motors, ragdolls, wheeled and tracked vehicles, soft bodies, GPU hair, and water buoyancy.
05:40 Its deterministic simulation replicates across machines from inputs alone. It runs on Windows, Linux, Mac OS, Android, iOS, and web assembly, depends only on the standard library, and offers bindings for C, Rust, Python, and more. Drop it into your engine and start simulating. Project number eight, Copilot Workshops. Hands-on lessons for GitHub Copilot Agentic Futures.
06:02 Copilot workshops is an open- source set of hands-on workshop lessons for learning GitHub Copilot's agentic features across the software development life cycle. It walks you through C-pilot CLI, VS Code agent mode, the C-pilot app, and the Copilot cloud agent. The lessons are plain markdown you can read on GitHub, and an optional Astro and Starlight site publishes them to a browsable page.
06:25 As you work through them, you build a sample crowdfunding app, Tail Spin Toys. Several tracks also ship in translated loces. It's a public template so teams can run their own workshop, fork it, and start learning agentic workflows. Project number nine, PI sub aents. Delegate work to child agents inside PI. PI Sub Aents is an extension for the PI coding agent that lets it delegate work to focus child agents.
06:49 You install it with one command, then ask PI in plain language to review a diff, scout a codebase, or run several reviewers at once. Each sub aent is a separate PI session running in the foreground or background so you can check it later. It ships with built-in roles like scout, researcher, planner, worker, reviewer, and Oracle. And you can chain or parallelize them, isolate parallel edits and get work trees and define your own agents in Markdown.
07:14 Install it and start delegating. Project number 10, Clockon, disposable networkrestricted Linux VMs for coding agents. Clockom is an open-source command line tool that gives an AI coding agent its disposable networkrestricted Linux VM. You cd into a repo, type clock and clock on or codeex runs inside a fresh Linux VM with your code mounted and root inside while your files and keychains stay out of reach.
07:40 Outbound traffic is denied by default behind a per sandbox allow list. Though your SSH agent is forwarded, so git push works. Any OCI image becomes the root FS and destroying a VM keeps your code and conversations on the host. It runs on Apple Silicon Max. Give your agent its own machine. Project number 11. Waggle. A 30 by token layer for agent handoffs.
08:02 I will say Waggle is an open- source reference layer for agent handoffs. Instead of pasting an artifact to each sub agent, one agent mints a 30 by token that each consumer resolves into a view for its own model. Behind it, an attribution manifest records who made it, and every read is logged, so you see which sub agent opened it. Corrections propagate to every holder.
08:23 It runs as an MCP server, so one config line wires it into Cloud Code, Codeex, or cursor. The same token resolves across machines and at Cloudflare's edge, while the file never leaves, handoff by reference, not by paste. Project number 12, Mindwalk. Replay coding agent sessions on a 3D codebase map. Mindwalk is a visualization tool that replays coding agent sessions on a 3D map of your codebase.
08:46 A log shows what an agent did, not how it understood it. So, Mindwalk draws your repository as a night map and plays it back as light. The files the agent searched, read, and edited glow while everything else stays dark. One go binary reads clawed code and codeex logs on your machine, then serves a local web view. You can scrub the timeline, switch tree and terrain views, and click any file for its history.
09:09 Run it and watch your session light up. Project number 13. Cloex local first zero trust agentic IDE with governed execution. Cloex is an open-source local first agentic IDE built on one idea. A model's output is untrusted input. It brings persistent AI tasks, code, terminal, browser, git, models, and memory into a single electron workspace. Treating each job as a durable task with its own state, permissions, and evidence.
09:37 Each task keeps context across restarts. Routes work between models and runs locally or on Docker, SSH, or the cloud. Sensitive actions pass through a guardian that authorizes outside the model and fails closed while pending edits and approvals keep the final say with you. Put your agent under policy. Project number 14. NeverClick. Control the mouse with your keyboard using Vision.
09:59 NeverClick is a free Windows desktop app that lets you perform mouse actions with your keyboard. When you press a hotkey, it uses computer vision to find the clickable elements on your screen and overlays each with a short letter label. You type the letters to click there. Because it reads raw pixels, it works in any application. You can left or right click, aim at any point in an element, select several at once, and switch windows without lifting your hands.
10:21 It runs entirely on your machine. Offline with no online AI service. Put the mouse down and try it. Project number 15, KRBN, pencil style rendering engine for 3D scenes to SVG. KRBN is an open- source web engine that renders 3D scenes as pencil style drawings in SVG. Instead of asking what color each pixel is, it asks which lines an artist would draw.
10:46 So, it deres and styles strokes from geometry rather than shading surfaces. Silhouettes of spheres, cylinders, and cones are computed exactly. hidden lines split into visible and ghosted runs. And shading comes from hatch density flowing along each surface's curvature. It handles analytic primitives, curves, and triangle meshes through one pipeline with handdrawn wobble that stays stable across animation frames.
11:08 Written in Typescript. Browse the gallery and sketch your scene. Project number 16. Beautify-gll that redesigns your GitHub readme. on beautify-g is an open- source agent skill that helps AI coding agents redesign a reposiito's readme homepage around its real content. Most repos have enough information but show jargon and directory trees before a visitor understands the point.
11:33 So the skill reads the repository first finds the clearest value and proof then decides how the page looks. It keeps visual identity in SVG heroes and diagrams and explanations in plain markdown so the page feels designed while staying searchable and copyable. You run it in whole redeem or SVG only mode and it never publishes without approval. Give your readme a clear entry.
11:53 Project number 17. Odd ASR V1. Arabic first speech recognition models and inference hub. Odd ASR V1 is a family of Arabic first generative speech recognition models from Odd AI. And this repository is the developer hub with model cards, benchmarks, and copypaste inference. It treats transcription as next token prediction over audio using a language model decoder rather than a CTC objective.
12:21 Two tiers share one architecture and prompt interface. Flash for realtime and ondevice work. Turbo for the lowest error on hard dialectal audio. It handles modern standard Arabic major dialects code switched Arabic English and 30 languages overall. Weights run through transformers or gguf with llama.cp. Clone it and transcribe your Arabic audio. Project number 18, Domain-STK.
12:46 One TypeScript API for customer custom domains. Domain SDK is an open- source TypeScript library that handles custom domains for your customers through one API. You add a host name to the platform you already run, and it returns the exact routing, ownership, and certificate records your customer needs for their DNS, tracks the domain until it is ready, then removes it safely.
13:05 It ships adapters for Versal, Cloudflare for SAS, Railway, Render, and Netlefi, plus an in-memory testing adapter that never calls a real provider. Add and remove or identification comes from the provider and it runs server side on node or bun. Install it and ship custom domains. Project number 19, flawless AI S control plane for Kubernetes remediation.
13:29 Flawless is a source available AI sur plane for Kubernetes and cloud infrastructure. It joins alerts, evidence, topology, human approval, controlled remediation, and recovery verification into one audible loop. Discover, diagnose, preview, approve, execute, verify, learn. The model plans and explains while the platform keeps the execution boundary through arbback and action catalog dry runs and audit records and it retests the original symptom instead of trusting a successful command.
13:58 It connects to Kubernetes, Rancher, Prometheus, Loki, Tempo, and CMDB with any open AI compatible model and deploys by Docker or Helm. Deploy it and let your infrastructure prove recovery. Project number 20, Persome, a local personal model every AI agent can use. Person is an open- source personal model that gives AI agents context about you. It learns how you think and work from focused Mac OS activity you grant permission to capture, then builds an evidence-backed model in layers from single observations up to
14:28 patterns and one current picture of you. Every claim keeps receipts and you can inspect, correct, export or delete anything. It exposes that model over MCP soced clients can resume your work with grounded context. Data stays under a local folder with no cloud account or sync. Install it and build your human.mb. Thanks for watching.