Skip to content

Top Open-Source GitHub Projects: colibri, ArtCraft, openGym, WeKnora, SCSKiller & Whirl #299

Read the Portuguese translation →

Author: ManuAGI - AutoGPT Tutorials · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Top Open-Source GitHub Projects: colibri, ArtCraft, openGym, WeKnora, SCSKiller & Whirl #299 · Channel: ManuAGI - AutoGPT Tutorials · 13:03 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI content generation, Model inference and optimization, Open source and software ecosystems

This weekly roundup introduces 20 open-source projects, including developer tools, AI agent platforms, creative software, and utilities for gaming and fitness. Several projects focus on running or managing AI locally, such as Colibri, Paperclip, WeKnora, Mortiflick, and Whirl. Others offer practical tools for code review, workout tracking, shader compilation, structured data, and learning while coding. The video also announces a separate channel dedicated to AI agent projects.

  • Diagram Design creates brand-matched editorial diagrams as self-contained HTML and SVG, with 27 diagram types.
  • Artcraft is a desktop studio for creating AI images, videos, music, sound, 3D meshes, and worlds.
  • Open Gym is a self-hosted workout tracker that stores data in a user-controlled folder and can sync across devices.
  • Open Code Review uses deterministic checks and an LLM agent to review Git diffs and provide line-level comments.
  • Paperclip helps users manage AI agent teams with goals, budgets, governance, approvals, and activity audits.
  • WeKnora turns documents into a retrieval-based Q&A system, an agent, or a self-maintaining wiki.
  • Colibri is a C inference engine that streams large mixture-of-experts models from disk and offers an OpenAI-compatible API.
  • SCS Killer precompiles game shaders into the driver cache to reduce shader compilation stutter.
  • Whirl is a self-hostable AI chat app with memory, living documents, MCP-connected tools, and installable skills.
  • The roundup also features decision models, an Android radio observer, a Commodore 64-inspired font, and an SDK for building AI gadgets.
Start Chapter Summary
00:00 Introduction and first projects The video introduces the weekly GitHub roundup, announces the AI Agent Studio channel, and covers Diagram Design, Artcraft, Cursor Plugins, Open Gym, and Open Code Review.
04:08 Agent management, document knowledge, and local inference The roundup presents Paperclip, WeKnora, Colibri, an ADHD-focused coding assistant skill, SCS Killer, and OpenAI Math.
07:17 Motion design, decision models, and AI chat Projects include Mortiflick, Decider, Whirl, and Fieldwatch, spanning local video creation, probabilistic decisions, AI chat, and passive radio observation.
09:53 Developer tools, learning, and creative projects The video covers Redux for structured data, Vibe Wise for learning with AI coding tools, a Commodore 64 keyboard font, and Muse Gadgets.
12:31 Strands Decider and closing The final project is Strands Decider, a compact decision model for agent workflows, followed by the video’s closing thanks.

Every week, developers release new open-source projects on GitHub. — 00:00

Diagram design is an open-source claude code skill that makes editorial quality diagrams as self-contained HTML and SVG. — 01:04

Its hybrid design pairs deterministic pipelines with an LLM agent. — 03:37

The results are machine generated and not peer-reviewed. — 07:17

Thanks for watching. See you in the next update. — 13:03

ArtCraft, Open Gym, Open Code Review, Paperclip, WeKnora, colibri, SCSKiller, Whirl, Strands Decider

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 inference engines to AI agent skills. Without wasting time, let’s get started. >> Before we jump into today’s project updates, here’s a quick announcement for everyone. We’ve launched a brand new YouTube channel called AI Agent Studio dedicated entirely to AI Agent projects,

00:32 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. The link is right there in the description. Don’t miss it. All right, let’s get into today’s video. >> Project number one, diagram design editorial brand match diagrams for claude code. Diagram design is an

01:04 open-source claude code skill that makes editorial quality diagrams as self-contained HTML and SVG. It covers 27 types: architecture, flowchart, sequence, timeline, quadrant, pyramid, and more. Each in three variants with no build step. Point it at your website and it reads your pallet and fonts, maps them to semantic tokens and checks contrast so every diagram matches your brand. Export to PNG or SVG and install it into Claude code or codeex. Install

01:34 it and draw your diagram. Project number two, Non Artcraft, an IDE for AI image and video creation. Artcraft is an open-source desktop studio for interactive AI image and video creation. an IDE for artists, designers, and filmmakers who want real control. On one canvas, you select models, add reference images, and edit prompts to generate images, video, music, and sound, 3D meshes, and worlds, drawing on a catalog of 62 models across providers. It’s a native Tari app for Mac OS, Windows, and

02:06 Linux, written in Rust, and part of the open-source crafting app suite. Download it and craft. Project number three, Cursor Plugins plug-in spec and official plugins for Cursor. Cursor plugins is the open- source home for Cursor’s plug-in specification and its official plugins. Each plugin is a standalone directory with its own manifest and can bundle agent skills, editor rules, and MCP server definitions listed in one marketplace manifest. Official plugins cover developer workflows. Continual

02:36 learning writes memory into agents. MD PR review canvas renders diffs as canvases, orchestrate fans tasks across parallel cloud agents, and the SDK plug-in helps you build on cursor, browse it, and extend your editor. Project number four, Open Gym, self-hosted gym and workout tracker you own. Open Gym is an open-source self-hosted gym and bodyweight tracker. Your data lives in a folder you control, plan routines, and log workouts, super sets, warm-ups, and cardio. And a muscle

03:07 map shows what’s trained, fatigued, or d-trained. Progression sets each session’s weights and explains why. Mist reps hold the load, and stalls trigger a de lo. It imports from fitnotes, strong and heavy, signs in with pass keys, and installs as an offline app that syncs across devices. Self-host it with Docker and own your training. Project number five, nonopen code review. AI code review CLI from Alibaba. Open Code Review is an open-source AI powered code

03:37 review CLI from Alibaba. Its hybrid design pairs deterministic pipelines with an LLM agent. It reads your git diff and the agent reads full files, searches the codebase, and inspects changes, then returns line level comments. A multi- language rule set catches issues like null pointer bugs, thread safety, XSS, and SQL injection, and it can auto apply fixes. It works with OpenAI and Anthropic with plugins for Claude Code and Codeex. Install it and review your code. Project number

04:08 six, Paperclip. Self-hosted app to run a company of agents. Paperclip is an open-source self-hosted app for managing a team of AI agents. If an agent is an employee, Paperclip is the company, a NodeJS server and React dashboard that orchestrates agents to run a business. You bring your own agents, any provider, any runtime, assign goals, set budgets, and track work and costs. It looks like a task manager, but adds org charts, governance, approval gates with roll back and activity audit. It self-hosts

04:39 on Postgress. Deploy it and manage your agents. Project number seven, be wig Nora. Turn documents into rag agents and auto wiki. Wiggora is an open-source LLM powered knowledge platform from Tencent that turns documents into knowledge three ways. Ragbased Q&A for lookups, a React agent that orchestrates retrieval, MCP tools, and web search for multi-step tasks, and a wiki mode where agents distill documents into a self-maintaining interlin markdown wiki with a knowledge graph. It ingests 10

05:12 plus formats, autosyncs from FU, notion, and UK. Works with 20 plus LLM providers and swaps vector databases. Self-host it with Docker and query your documents. Project number eight, Gibbri. Run huge MOI models from disk. In Pure C, Galibri is an open-source inference engine. In Pure C with zero dependencies that runs mixture of experts models on hardware you already own, ane model uses only a slice of itself per token. So, Calibri keeps that slice in RAM and streams the

05:44 rest from disk on demand. So, a model of hundreds of billions of parameters, runs on an ordinary PC. It’s a single C file, no blaz, no runtime Python, and serves an open AI compatible API. Install it and run a huge model. Project number nine, I have 8HD skill that stops coding agents burying the answer. I have ADHD is an open-source skill for AI coding assistants that stops them from burying the answer instead of a wall of preamble and hope this helps. It makes the assistant lead with the next action

06:15 number multi-step tasks. Cap lists give concrete time estimates and end with one clear next step. It’s 10 rules adapted from an 8HD toolkit. You install it into claw code or codeex and invoke it. Install it and get straight to the point. Project number 10, Necess Killer. Kill shader compilation stutter before you play. SCS Killer is a free open-source Windows tool that kills shader compilation stutter before you play. The hitch happens when your GPU driver compiles a new shader mid-frame.

06:46 SCS Killer reads the shaders a game ships with and compiles them into the driver cache with the game closed so they’re there when you launch. It recompiles after driver updates, skips anti-che games, and keeps everything on your PC. It fixes shader stutter on Nvidia and AMD. Install it and play without stutter. Project number 11, OpenAI Math 722 AI generated math manuscripts, many with lean proofs. Math is OpenAI’s public open-source repository of 722 mathematical

07:17 manuscripts produced by an internal frontier model grouped into 372 families. It spans pure math, theoretical computer science, and mathematical physics, number theory, complexity, geometry, and more. With many results backed by computer checkable lean proofs, each ships as a PDF with source and citations, and nearly everyone came from a single prompt to a single agent. The results are machine generated and not peer-reviewed. Clone it and read the proofs. Project number 12, Mortiflick EAS, local AI motion studio, where

07:49 Claude makes videos. Mortifix is an open- source motion design studio that runs on your machine and uses Claude to make animated videos step by step. You approving each stage. It’s the harness, pipelines, gates, a review room, and memory across sessions. You start a brief, let it run until something needs you. Review and approve or request changes until delivered. It adds narration, your own voice or a model, a music engine, and Blender 3D. Bring your own key, install it, and direct a video.

08:20 Project number 13. Decider. Open-typed decision models with calibrated probabilities. Decider is an open-source family of decision models that return typed answers with calibrated probabilities in a single forward pass. A language model that doesn’t generate text, feed it a state, and typed questions, and it returns a probability per question, a choice of up to 255 options, a 2 to 10 score, or a yes, no. It’s an open reproduction of the system one class from mode 8b up to a 35b plus

08:52 a vision model that decides from images. Clone it and decide. Project number 14. Whirl self-hostable AI chat with memory docs and tools. Wh is an open-source self-hostable AI chat app. The full code behind World. at under the MIT license. It puts every top model in one place with real memory, living documents, and your own tools connected over MCP, OOTH included. Installable skills teach it new tricks. And if you leave a provider’s keys out, Whirl hides the

09:23 features they’d power. It’s built on NextJS and Convex, and you can run it locally or in production. Deploy it and own your AI chat. Project number 15, Fieldwatch. Receive only Wi-Fi and Bluetooth observer for Android. Fieldwatch is an open-source receive only Wi-Fi and Bluetoothle observer for Android. It passively listens for the Wi-Fi and BLE radios around you and never transmits. Watching without connecting, it decodes drone remote ID broadcasts, logging a drone’s ID and

09:53 coordinates and flags anomalies like a radio that changes its address on every packet. A privacy mode masks remote ID details, and you can export sightings to Wiggle and CSV. It’s a sideloaded APK. Install it and watch the airwaves. Project number 16- Redux. Capcom’s tokenbased structured data engine for.net. Redux air ducks is an open-source tokenbased structured data engine for.NET from Capcom’s technical research division, a core of Rex, the technology behind its next generation

10:24 game engine. It’s more than a JSON serializer. It parses formats into a token representation. So you can read, write, and convert them. JSON, JSON 5, CBOR, message pack, TOML, XML, HTML, NI, CSV, and a fast binary docs. It adds dynamic DOM access and system.ext, JSON, and NOSH compatibility as Nougat packages. Add it and work with structured data. Project number 17, Vibe Wise. Learn how to build while AI writes

10:56 code. Vibe Wise is an open-source claude code plug-in that helps you learn how to build while the AI writes the code. It puts learning first, forcing design checkpoints and step-by-step reasoning before any code is written, so you stay in control. Run vibewise, learn in a project to start, pause, and resume any time or reset to learn a codebase from scratch. Your preferences, notes, and a project map live in a local.vibewise/folder. Install it and learn while you build. Project number 18, C64 keyboard font. A

11:28 font rebuilt from Commodore 64 keycaps. C64 keyboard is a free public domain font reconstructed from photographs of classic Commodore 64 keycaps. It recreates the printed key legends, not the C64 screen font. So you get uppercase letters, numbers, punctuation, smooth vector PKY graphics, the function keys, and the Commodore logo. It ships as desktop fonts, a web font, and editable outlines with an interactive type tester. Click a key to insert its symbol. It’s unofficial, not endorsed by

12:00 Commodore. Install it and type like a C64. Project number 19, Muse Gadgets. Build your own AI hardware that runs Muse. Muse Gadgets is Meta’s open-source SDK and firmware for building your own AI hardware that runs Muse. Two SDKs cover it. Flash the ESP32 firmware onto a board and wire up displays, buttons, and sensors, or install the Linux SDK to turn a Raspberry Pi into a gadget. Muse can run commands on. Gadgets pair with the Muse app and coding agents like Muse

12:31 code can build it for you. It’s Apache licensed. Build it and make a Muse gadget. Project number 20, Strands Decider, a small fast decision model for agents. Strands decider is a small open-source decision model, a system one model for agentic workflows. It reads a state and a question and picks an option or rates a scale faster than an LLM and with calibrated confidence. For hybrid agents, the LLM makes hard calls and a decider the wrote ones. In a strand agent, it gates a tool call on yes no

13:03 decisions. It ships a 2B model, a vision variant, and a CLI. Install it with pip and decide. Thanks for watching. See you in the next update.