← All transcripts

Top Open-Source GitHub Projects: Cloudflare Computer, AirLLM, Archify, Omnigent & WASTE #281 Transcript, AI Summary & Key Points

ManuAGI - AutoGPT Tutorials · 7 days ago · Science & Technology · 13:29 · EN-US

🧠 AI Summary

Twenty open-source GitHub projects are presented across AI agents, local model inference, developer productivity, system documentation, self-hosting, media, and office workflows. The projects include tools for giving agents computer environments, running large models on limited hardware, orchestrating multiple agents, generating interactive architecture diagrams, processing PDFs, editing office files with AI, and fine-tuning language models locally.

🔑 Key Points

  • Cloudflare Computer gives an AI agent a real computer inside a Cloudflare durable object, with a virtual file system in SQLite and pluggable execution backends.
  • AirLLM keeps one model layer on the GPU at a time, allowing a 70B model to run on a 4-gigabyte card and a 6.1B model on 12 gigabytes without quantization, distillation, or pruning.
  • Omnigent provides one orchestration layer across Claude Code, Codex, Cursor, Open Code, Hermes, Pi, and custom agents, with policies for pausing actions, capping spend, and limiting tools.
  • Archify turns codebases or system descriptions into searchable, shareable HTML diagrams covering architecture, workflow, sequence, data flow, and life cycle views.
  • WASTE streams active mixture-of-experts model layers from NVMe while keeping the shared trunk in memory, enabling the 2.78 trillion parameter Kimiko 3 to run on a 64 GB laptop.
  • World Monitor aggregates 500 feeds across 15 categories, generates AI-written briefs, maps events, and tracks military, economic, disaster, exchange, commodity, and crypto signals.
  • Cloudflare OS provides a self-deployable agent workspace with sandboxed gadgets, configurable internet access, service gatekeepers, action logs, and approval workflows.
  • GenOffice combines word processing, spreadsheets, presentations, and PDF editing with a shared AI panel and narrow patches that preserve untouched content byte for byte.

✅ Actionable items

  • Deploy Authentik with Docker Compose or Kubernetes to self-host login, multi-factor authentication, and user management.
  • Point an OpenAI or Anthropic client at Grok-2-API to access Grok capabilities through an OpenAI-compatible API.
  • Run Celld on personal machines and replicate its SQLite-backed durable objects to an S3-compatible bucket.
  • Install AirLLM and pass a Hugging Face model ID to run large language models on limited GPU memory.
  • Configure DeepSeek-Reasonix providers, tools, and plugins in a TOML file and run it through its NPM, Homebrew, desktop, or VS Code installation.
  • Install I-have-ADHD into Claude Code or Codex and invoke it to make coding assistants lead with the next action, number multi-step tasks, cap lists, provide time estimates, and end with one clear next step.
  • Run World Monitor locally with Ollama, or access it through MCP, REST, and CLI interfaces.
  • Use Archify in Cursor, Claude Code, Codex CLI, or Open Code to generate validated, self-contained HTML system diagrams.
  • Use Plezy to access Plex or Jellyfin libraries across desktop, mobile, and TV platforms.
  • Use Torph with React, Vue, Svelte, or plain JavaScript to animate text transitions with configurable spring stiffness and damping.
  • Use PDF Inspector to classify PDFs, identify pages needing OCR, extract position-aware text, and convert documents to Markdown.
  • Run Finger Frame AI in a browser or Python CLI, providing a Google Gemini key to restyle finger-frame videos.
  • Use GenOffice's shared AI panel to make granular edits to documents, spreadsheets, slides, and PDFs.
  • Configure Soup with YAML and run a single training command for LoRA, QLoRA, supervised fine-tuning, or preference optimization.

🤖 AI in practice

Used for

Give an AI agent a virtual computer where it can run code and access a file system. 00:57
Expose Grok's web capabilities through an OpenAI-compatible API for chat, image generation, image editing, and video. 02:11
Run large language models on GPUs that are too small to hold the full model. 02:25
Run a DeepSeek-native terminal coding agent during long coding sessions while reducing repeated prompt-token costs. 04:00
Let humans and AI agents collaborate in shared rooms where agents can access repositories, send patches, review code, and run workflows. 04:37
Build JVM agentic workflows that combine LLM calls, code, domain models, goals, and conditions. 05:15
Make AI coding assistants communicate in concise, action-oriented responses for multi-step coding work. 05:49
Aggregate global news and infrastructure signals into AI-written situational-awareness briefs. 06:20
Turn a codebase or system description into searchable, shareable architecture documentation. 07:02
Coordinate multiple AI coding agents and harnesses in one session. 07:43
Chat with a company-contextualized agent to write documents, run tasks, and build applications. 09:00
Stream mixture-of-experts models from NVMe so models larger than RAM can run locally. 10:11
Transform a finger-frame gesture video into a clip showing an AI-generated world inside the frame. 11:30
Edit documents, spreadsheets, presentations, and PDFs with an AI-native office suite. 12:05
Run local LLM fine-tuning and post-training workflows with less configuration and infrastructure work. 12:44

Agents

  • Cloudflare Computer — Give an AI agent a computer environment for running code and managing files. 2 held 00:57
  • DeepSeek-Reasonix — Act as a terminal coding agent for software-development sessions. 2 held 04:00
  • Buzz agents — Collaborate with human teammates on repositories, patches, code reviews, and workflows. 2 held 04:37
  • Embabel agents — Execute agentic JVM workflows that combine code and LLM actions. 2 held 05:15
  • Omnigent — Orchestrate multiple AI agents and let them review, share, and continue one another's work. 2 held 07:43
  • Cloudflare OS agent — Write company documents, run tasks, and build sandboxed applications called gadgets. 2 held 09:00
  • GenOffice tool-calling agent — Make targeted edits across spreadsheets, slides, and PDFs. 2 held 12:05

Advice

  • Install Cloudflare Computer when an AI agent needs a real execution environment and virtual file system. for Developers building AI agents
    It provides code execution through pluggable backends inside a Durable Object.
  • Use AirLLM to run large language models on small GPUs before resorting to quantization, distillation, or pruning. for Developers running models locally
    The transcript describes layer-by-layer GPU loading and gives examples of 70B on 4 GB and 6.1B on 12 GB.
  • Use Embabel when an agent workflow needs dynamic replanning after each step. for JVM developers building agents
    Its non-LLM planner determines the sequence from actions, goals, and conditions and replans as execution proceeds.
  • Install i-have-adhd into Claude Code or Codex when coding-assistant responses are too verbose or indirect. for Developers using AI coding assistants
    Its rules emphasize the next action, bounded lists, concrete estimates, and one clear next step.
  • Run World Monitor locally with Ollama when you need global intelligence briefs without API keys. for Analysts and developers building situational-awareness dashboards
    It aggregates 500 feeds across 15 categories and synthesizes them into AI-written briefs.
  • Use Archify to document a codebase or system as interactive diagrams. for Developers and technical architects
    It produces searchable HTML maps across five diagram types and validates them before image or video export.
  • Use Omnigent when several agent harnesses need to work together under spending and tool-use policies. for Teams orchestrating multiple AI agents
    It supports cross-agent review, harness swapping, session sharing, action pauses, spend caps, and tool limits.
  • Use WASTE to run mixture-of-experts models larger than available RAM on a local machine. for Developers running large models locally
    It streams active experts from NVMe while keeping the shared trunk in memory and using RAM as a cache.
  • Use Soup when local fine-tuning needs to be reduced to a YAML configuration and one training command. for Developers fine-tuning language models
    It supports LoRA, QLoRA, supervised fine-tuning, and preference optimization with optional layer streaming.

What it could not do

  • Cloudflare Computer is described as an early preview. — The project is presented as an early-preview library for giving agents a computer environment.
  • Cloudflare OS gadgets have no internet access unless it is explicitly granted. — Each gadget runs in its own sandbox.
  • Finger Frame AI requires the user to provide their own key. — The browser tool uses a Google Gemini video model and says, "You bring your own key."

🧰 Tools & AI usage

AI is used for

  • Provide agents with an executable computer environment — Cloudflare Computer lets agents access a virtual file system and run code through configurable backends.00:57
  • Generate coding-agent plans and execute terminal coding workflows — DeepSeek-Reasonix is a terminal coding agent, while Embabel uses dynamic planning to mix LLM calls with code.04:00
  • Coordinate multiple AI agents — Omnigent lets agents collaborate, review one another, switch harnesses, and share or fork sessions.07:43
  • Generate intelligence summaries and visualizations — World Monitor synthesizes news, geopolitics, and infrastructure signals into AI-written briefs and maps.06:23
  • Create interactive system documentation — Archify generates grounded, searchable diagrams from codebases or system descriptions.07:02
  • Restyle video — Finger Frame AI uses a Google Gemini video model to transform clips while preserving motion alignment.11:44
  • Edit office documents and files — GenOffice provides a shared AI panel and tool-calling agent for documents, spreadsheets, slides, and PDFs.12:05

🔗 Links mentioned

📄 Transcript

Searchable transcript of Top Open-Source GitHub Projects: Cloudflare Computer, AirLLM, Archify, Omnigent & WASTE #281 — ManuAGI - AutoGPT Tutorials (13:29). 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 that help teams build faster and work smarter. In this weekly update, we cover top trending open source GitHub projects this week, so you can quickly discover useful and trending developer tools, AI workflows, and productivity systems. 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 cutting-edge 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, Cloudflare computer. Give your AI agent a real computer. Cloudflare computer is an open source library that gives an AI agent a real computer inside a Cloudflare durable object. It holds a virtual file system in SQLite and exposes one execution entry point, letting the agent run code through pluggable backends, a full Linux container over a fuse mount, a bash shell in a worker, or a JavaScript

01:22 module in a fresh worker. Backends connect only when used. It's an early preview. Install it and give your agent a computer. Project number two, provider for single sign-on. Authentik is an open source identity provider for single sign-on that you host yourself. It sits between your users and your applications, handling login, multi-factor authentication, and user management through protocols apps already speak, SAML, OAuth 2, and OIDC, LDAP, and RADIUS.

01:53 Customizable flows let you shape sign-up, login, and recovery. It scales from a home lab to a production cluster and can replace providers like Octa, Auth0, and Entra ID. Deploy it with Docker Compose or Kubernetes and own your logins. Project number three, Grok-2-AP, OpenAI compatible gateway for Grok's web capabilities. Grok-2-AP is a self-hosted gateway built with FastAPI that turns Grok's web capabilities into an OpenAI compatible API.

02:24 You point any OpenAI or Anthropic client at it and reach Grok for chat, image generation, image editing, and video with streaming, thinking output, and tool calls. It manages an account pool with tiered selection, quota sync, and automatic maintenance and can cache media locally. It ships an admin panel and web UI. Deploy it with Docker and call Grok through familiar APIs.

02:47 Project number four, Celled, self-hosted distributed durable objects on your machines. Celled is an open-source daemon that runs Cloudflare Workers and Durable Objects on your own machines. It's a distributed system you program like a database. Each object is its own SQLite database, addressed by name and replicated to an S3-compatible bucket you own.

03:08 Nodes coordinate through that bucket alone with no control plane or consensus. And Litestream streams every change, so any node can restore a cell and resume. Idle cells hibernate to nothing. Install it and self-host durable objects. Project number five, Mesh-RLLM, run huge language models on small GPUs. RLLM is an open-source Python library that runs large language models on small GPUs.

03:34 It keeps one layer on the GPU at a time, so memory depends on a model's layer size, not its total size. That lets a 70B model run on a 4-gigabyte card and a 6.1B model on 12 gigabytes without quantization, distillation, or pruning. One auto model call takes any hugging face model ID working with Llama, Quinn, Deep Seek, and more. Install it and run a huge model.

04:00 Project number six, Deep Seek Resonex. Deep Seek native terminal coding agent tuned for prefix caching. Resonex is an open source Deep Seek native AI coding agent for your terminal shipped as a single static Go binary. It's tuned around Deep Seek's prefix cache keeping the prompt prefix stable across long sessions so token costs stay low. Providers, tools, and plugins are declared in one Toml file with no hardcoded models and any OpenAI compatible endpoint is just a config entry.

04:31 You install it through NPM, Homebrew, a desktop app, or a VS Code extension. Install it and leave it running. Project number seven, Buzz. Self-hostable workspace where humans and agents collaborate. Buzz is an open source self-hostable workspace from Block where humans and AI agents share rooms on a relay you own. It runs as a Nost relay so every message, reaction, workflow step, review, and Git event becomes one signed event in a single log with one identity and audit trail.

05:03 Agents join channels like teammates with their own keys and can open repos, send patches, review code, and run workflows. Run the relay and share a room with your agents. Project number eight, Embabel Agent Framework. Agent framework for the JVM with dynamic planning. Embabel is an open source agent framework for the JVM from the creator of Spring that lets you author agentic flows mixing LLM calls with code and type domain models.

05:28 You define actions, goals, and conditions and a non-LLM planner works out the sequence replanning after each step so it adapts. It's built on Spring so agents are managed beans and you write them in Kotlin or Java using annotations or a DSL. Add the starter and build your first agent. Project number nine, I have ADHD. Skill that stops coding agents bearing the answer.

05:53 I have ADHD is an open source skill for AI coding assistants that stops them from bearing the answer. Instead of a wall of preamble and hope this helps and makes the assistant lead with the next action. Number multi-step tasks, cap lists, give concrete time estimates, and end with one clear next step. It's 10 rules adapted from an ADHD tool kit. You install it into Claude Code or Codex and invoke it.

06:18 Install it and get straight to the point. Project number 10, World Monitor. Real-time global intelligence dashboard for situational awareness. World Monitor is an open source intelligence dashboard that pulls news, geopolitics, and infrastructure signals into one view. It aggregates 500 feeds across 15 categories and synthesizes them into AI-written briefs, plots events on a 3D globe or a WebGL map.

06:42 It correlates military, economic, and disaster signals, scores country instability, and tracks exchanges, commodities, and crypto. You run it locally with Ollama and no API keys or reach it via MCP, REST, and CLI. Clone it and watch the world. Project number 11, Wanarasea. Turn a codebase into verifiable interactive diagrams. Wanarasea is an open source agent skill that turns a codebase or system description into an interactive diagrams.

07:14 It works in Cursor, Claude Code, Codex CLI, and Open Code. You describe a system or a repository, and it produces a self-contained HTML map you can search and share. It covers five diagram types, architecture, workflow, sequence, data flow, and life cycle, and keeps nodes grounded in authored facts rather than inventing topology. It validates diagrams before exporting PNG, SVG, and WebM.

07:39 Map your system. Project number 12, Omnigent. One orchestration layer over all your AI agents. Omnigent is an open-source AI agent framework and meta harness that puts one orchestration layer over Claude Code, Codex, Cursor, Open Code, Hermes, Pi, and agents you write. You mix agents in one session, ask one to review another, and swap harnesses without rewriting.

08:04 Sessions follow you across terminal, browser, phone, and desktop app. And teammates can share, co-drive, or fork. Policies pause actions, cap spend, and limit tools. Agents run in cloud sandboxes. Install it and orchestrate your agents. Project number 13, Plezi. Cross-platform Plex and Jellyfin client built with Flutter. Plezi is an open-source client for Plex and Jellyfin that runs on desktop, mobile, and TV.

08:30 Built with Flutter and powered by MPV. It plays many codecs with HDR and Dolby Vision. Full ASS/SSA subtitles, progress sync, and skip intro and credits. Beyond playback, it offers live TV and DVR, offline downloads, synchronized watch together, and tracking through Trakt, AniList, and more. It's available on the App Store, Google Play, and every desktop platform.

08:55 Download it and stream your library. Project number 14, Cloudflare OS. Self-deployable agent workspace built on Cloudflare Workers. Cloudflare OS is an open-source agent workspace built on Cloudflare Workers. Used internally at Cloudflare, you chat with an agent preloaded with company context to write documents, run tasks, or build apps called gadgets.

09:17 Each in its own sandbox with no internet unless you grant it. A security layer called gatekeepers wrap services like GitHub, Google, and Slack. Logs actions and lets you approve later. It runs on Cloudflare or your servers via worker D. Deploy it and build your company's OS. Project number 15, Torf, dependency-free animated text morphing component for the web.

09:41 Torf is an open-source, dependency-free component that animates text as it morphs from one string to another. It works with React, Vue, Svelte, and plain JavaScript. So, you drop a single text morph tag around your text, and it smoothly transitions whenever the content changes. Written in TypeScript, it supports physics-based spring easing, letting you tune stiffness and damping for the exact feel you want.

10:05 You install it with one package and no extra dependencies. Add it and animate your text. Project number 16, Waste, stream trillion parameter models from NVMe beyond RAM. Waste, the weight-aware streaming tensor engine, is an open-source inference engine written in C with no runtime dependencies. It runs mixture-of-experts models beyond RAM by keeping the shared trunk in memory and streaming each token's active experts from NVMe, using RAM as a cache.

10:34 This lets the 2.78 trillion parameter Kimiko 3 run on a 64 GB laptop without distillation or pruning. It ships a CLI, an embeddable C library, and an OpenAI compatible server. Build it and run a frontier model locally. Project number 17, PDF Inspector, fast Rust PDF classifier and text extractor. PDF Inspector is a fast, open-source Rust library from Firecrawl that inspects PDFs, classifies them, and extracts text without OCR.

11:04 It detects whether a document is text-based, scanned, image-based, or mixed, returns a confidence score, and the pages that need OCR, then extracts position-aware text and converts it to markdown with headings, lists, tables, and links. It's pure Rust with no ML models, and ships bindings for Python, Node.js, and browser web assembly, plus a CLI. Install it and route your PDFs.

11:28 Project number 18, Finger Frame AI. Put an AI world inside your finger frame. Finger frame AI is a free browser-based tool that turns a video of the finger frame gesture into a clip with an AI generated world inside the frame. A Google Gemini video model restyles the whole clip as anime, claymation, watercolor, or a 3D look while keeping your motion aligned.

11:53 MediaPipe tracks both hands and the styled video is composited so the frame becomes a window. You bring your own key. It runs in the browser or a Python CLI. Frame your own world. Project number 19, Gen Office. AI native office suite for macOS and Windows. Gen Office is an open-source AI native office suite for macOS and Windows. It bundles a word-compatible word processor, an Excel-compatible spreadsheet, a PowerPoint-compatible presentation editor, and a PDF viewer and editor as Electron apps sharing one engine with

12:24 AI editing built in. Every app shares one AI panel. Block granular edits with version snapshots and docs and a tool calling agent over your spreadsheet, slides, or PDF. Edits apply as narrow patches, so untouched parts stay byte for byte. Download it and edit with AI. Project number 20, Soup. Simple CLI for LLM fine-tuning workflows. Soup is an open-source Python CLI that simplifies LLM fine-tuning and post-training through a YAML configuration and a single training command.

12:57 It supports workflows such as LoRA, QLoRA, supervised fine-tuning, and preference optimization with Hugging Face and PyTorch-based tooling underneath. Developers can train locally on their own GPU while optional layer streaming moves frozen model layers through RAM or NVMe to reduce VRAM use. This helps developers fine-tune models with less infrastructure work. Explore the repository to streamline local LLM training. Thanks for watching. See you in the next update.