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

AirLLM is an open-source Python library for running large language model inference on GPUs with limited memory. It reduces GPU memory use by loading and processing one model layer at a time instead of holding the entire model in GPU memory; for sparse mixture-of-experts models, it streams individual experts as needed. The project is designed to run large models without requiring quantization, distillation, or pruning, and provides a pip-installable package with support for multiple model families, CPU inference, macOS, and optional model compression and quantization features.

Archify is an open-source agent skill and Node.js rendering and validation system that turns a codebase or system description into searchable, shareable interactive diagrams. It supports architecture, workflow, sequence, data-flow, and lifecycle diagrams and can be used with Cursor, Claude Code, Codex CLI, and OpenCode. Agents produce typed JSON intermediate representations, which Archify deterministically validates and compiles into self-contained HTML and SVG artifacts. The generated viewer supports node search, optional revision-verified source inspection, upstream and downstream reach tracing, route inspection, semantic-role comparisons, guided stories, themes, presets, and finite motion. Archify can compare two validated snapshots as Before, Delta, and After views, identifying added, removed, changed, moved, and rerouted facts. Its exports include PNG, SVG, WebM, and 1200×630 share cards; it can be installed as a global skill with `npx skills add tt-a1i/archify -g`.

authentik is an open-source, self-hosted identity provider for modern single sign-on and identity management. It supports SAML, OAuth2/OIDC, LDAP, and RADIUS, and can be deployed with Docker Compose for small or test setups, Kubernetes via a Helm chart for larger clusters, or official AWS CloudFormation templates and a DigitalOcean Marketplace app. The project also offers an enterprise edition for large-scale identity management.

celld is an open-source daemon from Deno that runs Cloudflare Workers and Durable Objects on self-hosted machines. Each Durable Object is a named cell with its own SQLite database; cells store long-term state in an S3-compatible, Google Cloud Storage, or Azure Blob Storage bucket. A celld node embeds V8 and executes Wrangler bundles. Nodes sharing a bucket form a fleet, using conditional bucket writes for cell ownership and signed peer HTTP for routing and replicated-log transport, without a membership protocol, failure detector, or consensus service. Committed SQLite writes are captured as LTX data; nodes persist that data locally and upload it to the bucket, while takeovers recover open logs from the previous owner before restoring a cell. celld can run locally with `celld dev` and a local object store, or as a container for Linux x86-64 and ARM64. Its installer provides a verifiable binary release, and deployed Worker projects require esbuild on the PATH.

Cloudflare Computer is an open-source library that provides AI agents with a virtual filesystem backed by a Cloudflare Durable Object. The Durable Object stores the authoritative state in SQLite, while the workspace exposes a single execution interface through `workspace.runtime.exec(source, { backend })`. It supports three execution backends: Container projects the SQLite state into a sandbox container through a FUSE mount and synchronizes changes over capnweb RPC; Isolate shell runs `just-bash` in a Dynamic Worker; and Isolate JavaScript runs an ECMAScript module in a fresh Dynamic Worker with structured inputs and results, durable relative imports, configured libraries, workspace-backed `node:fs/promises`, and trusted `ws:git` and `ws:artifacts` modules. Workspaces can register multiple backends under stable identifiers, connect to them lazily, or be used solely as a filesystem without an execution backend. The repository includes examples for container, Worker shell, Worker JavaScript, egress policies, MCP, agent workspaces, tutorials, artifacts, and assets. The package is marked preview-only: its APIs are unstable and the repository says it is intended for experiments, exploration, and prototypes rather than production use.

Cloudflare OS is an open-source AI productivity environment developed by Cloudflare and built on Cloudflare Workers. It provides an agent chat interface preloaded with company-specific knowledge, sandboxed development of shareable applications called gadgets, and a security framework called Gatekeepers for controlling agents and apps. It is intended to be customized into an organization's own company operating environment rather than adopted unchanged as a traditional computer operating system. Each user receives a private, separately sandboxed instance of a productivity application, such as a slide-deck tool, whiteboard, game, or dashboard. Agents can create and modify these gadgets and perform tasks involving configured company systems and integrations. Gatekeepers are service-specific Workers that expose a Cap'n Web API, handle authorization such as OAuth, restrict access to the intended resource, log gadget and agent actions, and support human approval for side effects. Instead of stopping synchronously for approval, a Gatekeeper can simulate the outcome locally, let the agent continue and queue actions, then allow the user to approve or reject those actions later in bulk or individually. The full stack can be run locally with pnpm, Wrangler, and workerd using `pnpm run-local`, or deployed to a Cloudflare account. The repository describes the project as an early-access release under heavy development; the local setup is intended for evaluation rather than production use.

Embabel is an open-source framework for authoring agentic flows on the JVM. Written in Kotlin with a natural Java usage model and built on Spring, it combines LLM-prompted interactions with manually written code and strongly typed domain models. Flows are expressed through actions, goals, conditions, and domain objects; they can be authored with Spring-style annotations such as @Agent, @Goal, @Condition, and @Action, or with an idiomatic Kotlin DSL. The framework dynamically formulates plans toward goals using a pluggable, non-LLM planning step. Its default approach is Goal Oriented Action Planning (GOAP), while Utility AI is also supported for selecting actions by utility scores rather than strict preconditions and postconditions. Conditions are reassessed after each action, and the system replans based on new information and observed effects, allowing known actions to be combined in novel orders and enabling runtime decisions such as parallelization. The planning model is independent of the actions, goals, and plans themselves, and can extend application capabilities by adding domain objects, actions, goals, and conditions without changing finite-state-machine definitions. Execution is provided through an AgentPlatform with focused, closed, and open modes. Focused mode runs a requested agent from application code; closed mode classifies user intent or an incoming event to select an agent; and open mode searches across available goals and actions to build a custom agent, while allowing developers to restrict goal selection. Open mode can find paths not explicitly envisioned by developers and combine functionality from multiple providers, but individual steps must still be defined by the application. Embabel supports mixing language models, including local models for selected tasks, separates the programming model from platform internals for local or production execution, and integrates with Spring dependency injection, AOP, persistence, transactions, and testing. The framework is designed for unit and end-to-end agent testing.

Finger Frame AI is a browser app and Python CLI pipeline by Sophia Yang that places an AI-generated world inside a two-hand finger-frame gesture. It sends an uploaded video to Gemini Omni Flash for video-to-video restyling using a selected style or custom prompt, with alignment instructions intended to preserve framing, facial positions, expressions, blinks, gaze, and motion. MediaPipe Hand Landmarker then tracks both hands and the finger-frame quadrilateral frame by frame; the tracking pipeline uses anatomical corner ordering, spread and area gates with hysteresis, teleport rejection, velocity-adaptive smoothing, dropout holding, and presence fading. The compositor reveals the restyled video through the tracked quadrilateral and adds a dashed outline with pulsing corner markers. The browser app exports MP4 where browser recording supports it and otherwise WebM. The CLI separates restyling and compositing into `stylize.py` and `composite.py`, accepts formats readable by FFmpeg or OpenCV, and can produce H.264 MP4 while retaining the original audio track when present. Users provide their own Gemini API key; the app keeps it in the browser and sends it only to Google's API. Generation is billed per video, and the app recommends keeping uploads under about 15 MB. A no-key placeholder mode runs the tracking, compositing, and export pipeline with a hue-shifted stand-in.

GenOffice is a free, open-source AI office suite from Genspark AI for macOS, Windows, and Linux. It consists of six Electron applications sharing an engine layer: Docs for .docx files, Sheets for .xlsx spreadsheets, Slides for .pptx presentations, PDF for viewing and editing PDFs, Markdown for plain Markdown files, and a shell application that hosts the editors. It opens and saves Microsoft Office formats, and includes light, dark, and system themes. Its AI panel supports document-aware, block-level edits with version snapshots and diffs; in spreadsheets, presentations, and PDFs it provides a tool-calling agent over document state. Built-in tools include web search, image search, image generation, and media analysis. Users can sign in through Genspark or supply their own keys for providers including Claude, OpenAI, Gemini, DeepSeek, Kimi, GLM, Qwen, Doubao, MiniMax, Grok, Mistral, and OpenRouter, as well as custom OpenAI-compatible endpoints and local servers. GenOffice Docs uses byte-preserving .docx editing in which only modified paragraphs are regenerated, with Word-compatible pagination, tracked changes, comments, styles, equations, and ink. GenOffice Sheets uses the open-source Univer core with in-house extensions and a Rust .xlsx import/export sidecar, plus charts, pivot tables, slicers, conditional formatting, and formula tracing. GenOffice Slides has an in-house .pptx parsing, rendering, and editing engine with masters, layouts, charts, cropping, ink, and text shaping. GenOffice PDF supports annotations, forms, outlines, stamps, signatures, page operations, printing, direct text reflow and image editing, and local conversion to editable Word, PowerPoint, or Excel files; scanned PDFs can use system OCR on macOS and Windows. GenOffice Markdown is a Tiptap block editor that saves headings, lists, tables, images, and code blocks back to plain Markdown. The suite is distributed under the Apache-2.0 license with installers for the three supported desktop platforms.

Grok2API is a self-hosted, Go-based API gateway with a built-in React administration console. It manages separate account pools for Grok Build, Grok Web, and Grok Console, synchronizes account credentials, quotas, and models, and provides unified OpenAI- and Anthropic-compatible APIs with account and model routing. The gateway exposes Grok capabilities including chat, image generation and editing, and video. The project is intended for technical research and learning and asks users to comply with Grok's terms of use and applicable laws.

i-have-adhd is an open-source skill/plugin for coding assistants that formats responses as ADHD-friendly output. It leads with the next action, numbers multi-step tasks, suppresses tangents, restates state, gives specific time estimates, uses matter-of-fact errors, limits lists to five items, and ends with one concrete next step. The skill is installed through a coding assistant's CLI and can be customized by editing its SKILL.md file. It is licensed under the MIT License and does not require an ADHD diagnosis.

Omnigent is an open-source AI agent framework and meta-harness that provides a common orchestration layer for Claude Code, Codex, Cursor, OpenCode, Hermes, Pi, and custom agents. It allows agents to be swapped or combined without rewriting integrations, supports multi-agent supervision and collaboration, and keeps messages, sub-agents, terminals, and files synchronized across terminal, browser, phone, and desktop sessions. The framework can run agent sessions in disposable cloud sandboxes and applies governance policies at the server, agent, or chat level, including approval gates for risky actions, spending limits, and tool restrictions. It supports compatible model gateways and optional integrations for model providers, sandbox providers, SDK harnesses, storage, and memory. Omnigent is distributed as an open-source Python package requiring Python 3.12 or later, with installation through its installer, uv, pip, Homebrew, or the Git repository.

PDF Inspector is an open-source Rust library by Firecrawl for classifying PDF documents and extracting position-aware text without OCR. It samples content streams to classify documents as text-based, scanned, image-based, or mixed, returns confidence scores and per-page OCR-routing information, and detects broken font encodings. Its extractor uses font information and X/Y coordinates to determine reading order, including multi-column and right-to-left layouts, and converts content to Markdown with headings, lists, code blocks, tables, formatting, links, and page breaks. Native Rust and CLI consumers can opt into selective OCR for pages that need it; the Python and Node.js bindings can run the same workflow, while browser WebAssembly provides local parsing without a server round trip. The project also supports CID font decoding and shares a single parsed document between classification and extraction.

Plezy is an open-source cross-platform client for Plex, Jellyfin, and Emby, built with Flutter for desktop, mobile, and TV. It provides library, collection, playlist, watchlist, recommendation, and cross-server search views, along with media playback supporting direct play or transcoding, HDR and Dolby Vision, multiple codecs, customizable subtitles, playback progress synchronization, and external players. The client also supports music playback, offline downloads, live TV and DVR, synchronized watching, and integrations with services including Trakt, MyAnimeList, AniList, Simkl, and Seerr.

Reasonix is an open-source AI coding agent built around a local engine and distributed as a self-contained static Go binary. It is designed for DeepSeek, with DeepSeek provided as a preset, while configuration can target OpenAI-compatible endpoints and optionally use separate executor and planner models. The agent can be used through a terminal or desktop app, in a browser, or from an editor through ACP; it provides plan mode, permissions, workspace sandboxing, per-turn checkpoints, and MCP- or extension-based plugins. Its context maintenance is designed around prefix-cache stability, including a stable startup environment summary and pruning of stale tool output before compaction.

Soup is an open-source Python CLI for fine-tuning and post-training large language models from a YAML configuration and a single training command. It supports supervised fine-tuning, LoRA, QLoRA, preference optimization, quantization, automatic batch-size and GPU detection, and local training without a cloud service. Its optional layer-streaming mode keeps the frozen base model out of VRAM and feeds it to the GPU one decoder layer at a time, reducing memory use for low-VRAM systems. The project describes training an 8B model on a 4 GB laptop GPU with this mode; layer streaming is marked beta.

Torph is a dependency-free animated text morphing web component for React, Vue, Svelte, and vanilla JavaScript. Its TextMorph component transitions between strings and supports physics-based spring easing configured with parameters such as stiffness and damping. The package is installable through pnpm and is documented in the project's repository.

WARP (Weight-Aware Runtime and Paging), formerly WASTE, is an embeddable inference engine written in dependency-free C by SQLiteAI. It runs large mixture-of-experts models beyond available RAM by keeping the shared model trunk in memory, streaming selected experts from NVMe or other fast storage, and using remaining RAM as a bounded expert cache. Its container layout gives each expert an aligned read; reads overlap computation, while a lookahead router prefetches experts for the next layer without changing the decisions made by the actual router. Experts use 3-bit residual vector quantization, while more sensitive shared weights use 4- or 8-bit storage. The project is designed for local inference of frontier-scale models such as Kimi K3 and GLM-5.3-Flash on consumer hardware.

World Monitor is an open-source global intelligence dashboard developed in the koala73/worldmonitor project. It aggregates curated global and regional news, geopolitical information, infrastructure signals, and financial data into a unified situational-awareness interface, producing AI-synthesized briefs and tracking military, economic, disaster, and escalation signals. The application provides both a 3D globe and a WebGL flat map using a shared map-layer catalog, a server-authoritative Country Instability Index, and finance views for stock exchanges, commodities, crypto, and composite market signals. It supports local AI through Ollama without API keys, as well as other documented AI providers, and exposes an MCP server for programmatic access by agents and scripts. The repository supports multiple site variants from one codebase, including world, tech, finance, commodity, energy, and happy views. It also distributes a Tauri desktop application for macOS, Windows, and Linux, and documents self-hosting through Vercel, Docker, or static deployment.