A brand story is converted into a connected camera flight with generated isometric scenes, transition clips, and a portable scrolling playback engine.
Text is checked for canned AI-writing habits, with the smallest useful revisions applied and a separate detection mode that reports patterns without claiming to identify the author.
A single browser workspace combines paper search, scientific database queries, code execution, connected compute, artifact storage, and provenance tracking.
AgentENV (AENV) is an open-source distributed platform developed by kvcache-ai for running agent environments at scale, including environments used for agentic reinforcement-learning training such as Kimi K3. It runs Firecracker microVM sandboxes across machines, loads OCI-compatible images on demand through overlaybd, and uses local disks as bounded caches for image and snapshot data. The aggregate image and snapshot footprint can exceed local disk capacity because cold data is evicted from the cache. Snapshot-backed environments can boot or resume in under 50 ms and pause in under 100 ms. AENV incrementally snapshots memory and filesystem changes, stores snapshots in S3-compatible object storage or a shared distributed filesystem, and can fork a running environment into independent sandboxes for parallel workflows. It uses ublk for I/O and memory ballooning to return reclaimable guest memory to hosts; the documentation reports a 9.6× memory overcommit ratio in production. The distribution includes an AENV server and the `aenv` command-line client for pulling OCI images as templates, starting, attaching to, executing commands in, pausing, resuming, and deleting sandboxes. It exposes an E2B-compatible HTTP API that works with the standard E2B Python and TypeScript SDKs. It requires Linux kernel 6.8 or later and `/dev/kvm` access for Firecracker execution. API requests are authenticated, but traffic is not encrypted; the documentation recommends a trusted network or HTTPS termination at a reverse proxy or load balancer.
Amicro (@subhanhq/amicro) is an open-source React library of micro-interactions, transition components, Motion-powered primitives, custom hooks, spring presets, and animated card layouts. Its components are copied as TSX source into a project rather than delivered as a runtime service, with support for React, Next.js, and Vite applications using Tailwind CSS and Motion or framer-motion. The library includes entrance transitions such as fade-in, fade-up, and zoom-in; interaction components such as tilt-card and magnetic-button; text-reveal effects; scroll-progress utilities; and card layouts including fanned arcs, corner fans, cascades, radial wheels, scatter spreads, interactive 3D carousels, CoverFlow, and a timeline-controlled perspective stack. The npm CLI initializes project configuration and adds individual components, hooks, or utilities. Components can also be installed through a shadcn/ui registry using direct raw URLs or the @amicro namespace. The project supports light and dark UI palettes, requires Node.js 18 or later, and is distributed under the MIT License. It was created by Syed Subhan, associated with Subhan-code.
AVAL is an open-source web format and runtime for short prerendered motion with continuous loops, named application states, authored triggers, bounded transitions, reversals, frame-accurate timing, and packed-alpha transparency. It packages a logical animation as codec-specific .avl files containing motion units and a state graph; the browser selects a decodable candidate from an AV1, VP9, H.265/HEVC, and H.264 ladder, while graph routes handle state changes rather than timestamp seeking. Projects are compiled from a motion.json file with the @pixel-point/aval-compiler package, producing a directory containing the .avl output and build metadata. AVAL requires the integrating application to handle unsupported browsers and fatal playback errors and does not itself create or reveal fallback video, image, text, or other content.
An open-source agent skill for redesigning GitHub README homepages around a repository's actual content. It reads the repository first, identifies the clearest value and supporting proof, and then derives a project-specific visual system rather than applying a shared template. In whole-README mode, it works across content, visual-system, and engineering layers: it removes repetition and internal jargon, moves proof toward the opening, derives typography, color, composition, and project-native motifs, and keeps assets GitHub-safe, accessible, searchable, and copyable. The skill separates visual and content layers by using SVG for editable heroes, section transitions, comparisons, diagrams, and identity while retaining maintainable Markdown for searchable body text. It supports hybrid SVG compositions with optional AI-generated elements, local previews, and an approval step before publishing. The repository documents examples used by eight public repositories, including project-specific heroes and real outputs for slide creation, UI reconstruction, icon generation, agent delegation, vehicle telemetry, interactive mapping, and a solo werewolf game.
Bento is an open-source PowerPoint alternative and office suite distributed as a single HTML file. Each deck contains its own viewer, presenter, editor, fonts, images, charts, animations, and readable JSON document data, so it can be opened, edited, presented, and shared in a modern browser without an account, installer, or network connection. The editor rewrites the deck's data block when saving, using the File System Access API with a download fallback. It supports morph presentations in which elements sharing an ID animate between slides, built-in bar, line, pie, and scatter charts, speaker view, comments, layouts, interactive states, motion paths, PDF export, and multiple interface languages. Optional collaboration uses end-to-end encryption with AES-GCM, a CRDT for merging offline edits including character-level text changes, and a relay that stores ciphertext rather than document contents; Offline mode blocks synchronization and collaboration.
Canvas UI is an open-source, framework-agnostic library of creative UI components that render canvas and WebGL effects over live, interactive HTML. Its components use the experimental HTML-in-canvas API to read and redraw the DOM as a texture while preserving selectable text and clickable links; where that API is unavailable, they fall back to WebGL overlays. The library includes fluid simulations, shader effects, 3D scenes, distortion, particles, glass, fire, retro, and glitch effects, including Liquid, Ripple, Blaze, Glass, VHS, Shatter, and Decrypt Reveal. Each component is implemented as a plain TypeScript and WebGL engine with thin wrappers for React, Solid, Preact, Vue, Svelte, and vanilla JavaScript. Components are copied into an application through a shadcn-compatible registry rather than installed as a conventional dependency, and provide typed props with default configuration. The registry is MCP-ready for AI-assisted component discovery and installation. HTML-in-canvas effects require Chrome with the relevant flag or an origin-trial token; other environments use the WebGL-overlay fallback, while the library’s 3D effect components work without that flag. The project includes the component source, documentation site, demos, and registry builder, and is distributed under the MIT license with the Commons Clause, which restricts selling the library itself.
A browser-based first-person shooter built with Three.js r180 and WebGL2. It uses procedurally generated textures, meshes, animations, and Web Audio synthesis rather than external art or sound files; its systems include a market-street world with enterable interiors, custom physics and character movement, procedural weapons and ballistics, GPU visual effects, navmesh-based enemy AI with perception and cover behavior, and a DOM/CSS HUD. The repository also provides a development and testing harness: headless-Chromium screenshot capture, isolated reproducible shotsets with pixel-diff gating, scripted playtesting, and a gameplay profiler that reports frame-time distributions and attributes hitches to per-frame WebGL program compilation. The project is run with npm and has Three.js as its only runtime dependency.
Codex Security is an OpenAI CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities in authorized code repositories. It scans directories and codebases, including repository changes, uses AI-assisted analysis to trace attack paths and apply custom threat models, validates findings, and can generate reviewable patches and exports in SARIF, JSON, and CSV formats. The CLI supports CI use with an API key and can stop builds at a selected severity threshold. The package also supports configurable deep scans, worker and subagent counts, time limits, containerized bulk scans, and a preview findings service. The service stores findings and embeddings in SQLite, provides paginated listings and a read-only dashboard, identifies potential duplicates through embedding similarity, and allows completed findings to be published and deduplicated through the CLI or SDK. It requires Node.js 22.13.0 or later and Python 3.10 or later, and can use OpenAI or other documented inference providers.
Colibrì is a pure-C inference engine and open research platform for running large mixture-of-experts models on consumer and heterogeneous hardware without engine dependencies. It treats VRAM, system RAM, and storage as a unified multitier inference hierarchy, streaming expert weights from disk and coordinating placement, storage I/O, scheduling, kernels, speculation, and CPU/GPU overlap. The project provides chat, server, and web front ends, with support for CPU, CUDA, Metal, and dual-SSD streaming; its design rules prohibit silently changing model precision or router semantics when fast memory is insufficient.
An open-source native ARM64 port of Command & Conquer: Generals — Zero Hour for Apple Silicon Macs, iPhone, and iPad. It runs the real 2003 engine rather than an emulator, supporting campaign, skirmish, and Generals Challenge modes with RTS touch controls including tap selection, drag-box selection, long-press deselection, two-finger scrolling, and pinch zoom. The rendering path translates DirectX 8 through DXVK to Vulkan and then through MoltenVK to Metal. The port also reroutes the engine's writable configuration, cache, and save paths for iOS's code-signed app environment, adds an iOS DXVK cross-build and Vulkan-loader patch, and handles app-resume behavior when iOS removes the Metal drawable. It is based on EA's GPL v3 source release and community work including GeneralsX, with additional engine fixes and iOS/iPadOS support. The repository does not include or distribute game assets; users need their own copy of the game. The project is hosted on GitHub.
An open-source project that runs a 28.9-million-parameter language model locally on an ESP32-S3 microcontroller, without network connectivity. It uses a tiered memory layout based on Per-Layer Embeddings: frequently accessed activations and normalization weights stay in SRAM, the model core and output head reside in PSRAM, and a roughly 25-million-parameter embedding table is stored in flash; only the few rows needed for each token are read into faster memory. The included TinyStories model generates short stories, while a Barista model provides espresso question answering. The repository notes that the TinyStories model is not designed for general question answering, instruction following, code generation, or factual knowledge, and provides separate scripts for downloading and verifying model assets and deploying a selected model to the board.
GC Minimal Zine Poster is a Codex skill for turning a theme, sentence, article idea, object, mood, photograph, or reference set into a sparse editorial poster, production-ready image prompt, or reusable visual system. It compiles requests into a vertical aged-paper composition with 70–90% negative space, one small visual subject or event, restrained typography, a high-chroma color accent, and xerox, risograph, halftone, letterpress, or scanned-paper textures. The skill supports Generate, Reference Analysis, Prompt-only, Analyze + Generate, and Photo Input routes; its reference-analysis and quality-review workflows require image inspection, while image generation depends on the host runtime's available model. It is distributed as a Git repository for use with Codex or another compatible Skill runtime and contains no scripts, external fonts, API keys, private paths, or downloaded runtime assets.
Harness Engineering is Ryan Lopopolo’s anthology, field guide, and agent-context repository for improving coding-agent output without changing the underlying model or coding agent. It shapes the surrounding environment through context and tools so an agent can recover intent, operate the real system, respect authority, prove its outcome, and incorporate lessons from accepted work, corrections, failures, and user responses. The repository uses AGENTS.md to route agents to relevant arguments, cases, and proof, and provides a thesis index, playbooks, source material, and executable constraints for encoding organizational requirements and decisions. Repository-authored material is licensed under CC BY 4.0.
img2obj is a Codex plugin that reconstructs an object from an attached image, screenshot, or local image path as a code-only, procedural Three.js model. It first validates the reference, then plans an ObjectSculptSpec covering the component hierarchy, geometry, materials, pivots, sockets, motion requirements, and visual priorities. The model is built through blockout, form, look-development, and interaction phases; browser renders are compared with the source image, with quality evidence and approval required before advancing. The plugin generates procedural Three.js geometry rather than extracting or downloading a mesh. It supports hard-surface and organic assets, vegetation, fabric, glass, hair or fur, emissive elements, and decals; can define animation-ready parent-child relationships, detachable parts, pivots, and sockets; and can produce reference-derived PBR evidence such as palette, roughness, height, normal, and ambient-occlusion maps. It is intended for stylized props, mechanical objects, plants, scene assets, and interactive models, not photogrammetry, exact mesh extraction, or guaranteed production-ready geometry from a single image. The repository requires Codex with local plugin support, Python 3.10 or newer, and a browser project using Three.js.
img2threejs is an agent skill that reconstructs an object or character from a reference image as a code-only, procedural Three.js model. It generates a TypeScript THREE.Group factory using primitives, procedural shaders, and generated geometry rather than photogrammetry, mesh extraction, or downloaded art assets. The generated scene includes runtime structures such as pivots, sockets, and colliders for animation, and the project documents separate hard-surface and anatomy-aware reconstruction paths. It runs under Claude Code, Codex, or OpenCode and provides live browser demos whose models can be inspected as generated source.
Jacobian Lens (jlens) is a research and interpretability tool from Anthropic for examining what an internal language-model activation is disposed to make the model say. It linearly transports a residual-stream vector from any layer and position into the final-layer basis using an average input–output Jacobian computed over a text corpus, then applies the model’s own unembedding to produce ranked vocabulary tokens. The reference implementation fits lenses on open-weight decoder transformers, applies pretrained or newly fitted lenses, and renders an interactive layer-by-position view with top-token ranks, pinned-token tracking charts, and rank heatmaps. It is distributed as an Apache-2.0-licensed Python package and repository; the code is described as not maintained, and model weights and text corpora are not bundled.
Knockoff is a browser extension that filters pseudo-brand listings from Amazon search results by hiding, dimming, or labeling them. It runs locally in a content script and resolves each listing's brand against user allowlists and blocklists, bundled lists of flagged and established brands, and name heuristics for patterns such as all-caps strings, unusual consonant runs, non-English letter pairs, non-Latin characters, and irregular capitalization. The extension supports Chrome, Firefox, and Safari; the public repository is a frozen, fully local snapshot with no accounts, tracking, or network requests, while maintained auto-updating versions are distributed through browser extension stores.
LingBot-World 2.0, also called LingBot-World-Infinity, is an AI world-modeling system from the Robbyant Team for generating interactive video environments. Its causal pretraining paradigm is designed for an unbounded interaction horizon, while a distilled real-time variant is claimed to drive 720p video streams at 60 frames per second. The system supports diverse actions, including attacking, archery, spell-casting, and shooting, along with text-driven events. Its agentic harness uses a pilot agent to plan and execute character behavior and a director agent to synthesize new environmental elements as a scene progresses. The causal inference implementation processes video frames chunk by chunk with KV caching rather than processing the entire sequence at once. The release provides inference code and model weights, including a 14B causal-fast model, with a codebase built on Wan2.2; models are distributed through Hugging Face and ModelScope, and the real-time system can be tried through Reactor on the web or LingGuang on mobile. The project is licensed under CC BY-NC-SA 4.0 for non-commercial use. It states that deployment code will not be released and points to SGLang or flashdreams for self-deployment references.
An open GitHub guide and configuration project by jamesob for running state-of-the-art large language models locally. It documents a multi-GPU system using PCIe switches so GPUs can communicate peer-to-peer, including hardware selection, BIOS bifurcation, link-speed and ASPM settings, kernel and GRUB parameters, ACS configuration, and GPU power limiting. The repository also provides Docker-based serving configurations, a local speech-to-text configuration, and a GPU peer-to-peer bandwidth and latency benchmark.
An open, machine-readable taxonomy of learning across the primary and elementary years, produced by Marble. It represents fine-grained micro-topics as nodes in a prerequisite graph, with each topic containing a plain-language description, subject and domain, conceptual or procedural type, approximate age range, mastery evidence, assessment prompt, and links to source curriculum standards such as NGSS, Common Core, and the UK National Curriculum. Dependency edges identify hard or soft prerequisites and include reasons; the dataset also provides parent-friendly domain summaries. The JSON files include topics, dependencies, curriculum standards, domain clusters, and a manifest with counts and checksums. The repository describes it as pure data with no runtime or dependencies, and provides an interactive curriculum explorer at withmarble.com/curriculum.
No AI Slop is an AI-writing editing skill that removes more than 20 canned machine-writing patterns while preserving the author's vocabulary, cadence, humor, and imperfections. Its rules target binary contrasts, throat-clearing openers, faux-insight setups, colon reveals, dramatic fragments, superficial analysis, importance puffery, weasel attribution, synonym cycling, and fake-profound endings; they also cover active voice, leading with the point, untangling difficult sentences, and preferring concrete details over abstractions. The skill supports editing text, detecting and quoting suspected patterns without judging whether AI produced the writing, and generating deliberately exaggerated AI writing for satire. It can be invoked as `/no-ai-slop` in ChatGPT, Claude Code, Codex, or another coding agent, and is distributed through the Skills package manager or an npx command. The repository contains the editing rules in SKILL.md, evaluation checks in eval.md, ChatGPT and Codex plugin metadata, and a plugin build script; it is licensed under MIT.
OpenScience is an open-source AI workbench for scientific research developed by Synthetic Sciences. It runs as a browser-based workspace where a single research agent takes a goal through literature review, hypothesis formation, code writing and execution, experiments, analysis, and a written report. The agent can load domain skills, delegate bounded exploratory or execution work, query scientific databases, and preserve sessions as observable research traces. The local server hosts the workspace UI, agent runtime, skill library, and tool layer. The agent plans with a research harness and calls shell, editor, LSP, MCP, scientific database, and skill tools; sessions, skills, artifacts, and provenance are stored on disk. The workbench includes a file tree, code editor, terminal, session history, and inline rendering for molecules, structures, genomes, and plots. Its bundled skills cover training, evaluation, datasets, molecular and clinical biology, cheminformatics, papers and LaTeX, figures, and scientific runtimes, while database connectors provide direct access to services including UniProt, PDB, Ensembl, ChEMBL, PubChem, arXiv, OpenAlex, and Semantic Scholar. It supports frontier and open-weight models through user-provided keys, eligible ChatGPT or Codex access, local models, and optional Ace-managed models. Models are routed per request, allowing providers to be switched without changing the workspace. Extensibility includes LSP integration, MCP servers, plugins, custom agents and commands, experimental Python environments and BioNeMo adapters, and a TypeScript SDK. It is installed with the @synsci/openscience npm package or launched through npx; platform binaries and desktop installers are distributed through GitHub Releases. A free Synthetic Sciences account links an installation and can provide credential synchronization, private research graphs, enhanced search, and optional credit-backed models, while BYOK and local-model usage remain separate from Ace.
OpenWorker is an open-source desktop AI coworker that produces finished deliverables from everyday tasks, including code-security reviews with proposed fixes, cloud-configuration audits, incident reports, documents, spreadsheets, drafted messages, and scheduled briefs. It runs on the user's machine through a native desktop app and a local Python agent server built on aisuite, working across local files, the terminal, connected applications, and more than 25 connectors. Specialist coworkers cover security review, cloud posture, incident triage, everyday work, and recurring automations; security review combines deterministic scanners such as Semgrep with model reasoning, then re-scans and diff-reviews proposed fixes before approval. The agent breaks requested outcomes into steps and uses models from providers such as OpenAI, Anthropic, and Google, open-weight providers, or a local Ollama deployment, with the user supplying model access. Consequential actions—including sending messages, changing calendars, and running commands—are approval-gated. Its governance design includes human-only floors for dangerous or irreversible operations, explicitly granted and revocable autonomy rules, circuit-breaker escalation for uncertain or repeatedly denied actions, and an audit trail recording tool calls, approval provenance, and reviewer reasoning; unattended runs cannot self-approve. The project is in open beta and provides downloads for macOS on Apple Silicon and Windows on x64.
PenEcho is an open-source shared canvas for handwriting, equations, diagrams, and spatial context. It reads marks and their spatial relationships on a 20,000 × 20,000 canvas, then places hints, explanations, formulas, plots, diagrams, and other AI-generated drafts beside the original content. Drafts can be moved, resized, copied, accepted, or discarded before becoming part of the canvas ink; users can draw with a stylus or mouse, lasso confirmed ink to move, resize, recolor, or delete it, and send selected content to Typeset without triggering an AI request when editing ordinary ink. The application supports editable sandboxed HTML widgets, professional diagrams, animations, and live-data plugins, which can be refined in place through incremental unified-diff edits rather than full regeneration. It supports up to ten API or CLI connections, including Kimi, MiniMax, Codex, and Claude Code configurations, with a selectable active connection per client. Canvases can be organized into projects, shared across authorized devices, and exported as cropped PNG files. PenEcho provides a desktop app and an npm installation requiring Node.js 20.3 or newer. It can run offline with a user's own API or authenticated CLI setup; API credentials remain on the host device and are not sent to browser code. PenEcho Cloud is an optional companion service for private versioned projects, synchronized favorites, linked-device remote access through a relay, and public sharing of Canvases and Widgets as Echoes or Crafts. The project is supported through Moonshot AI's Kimi Open Source Friends program.
Personal Model is an open-source, local-first AI memory runtime from Intuition Lab for giving coding agents evidence-linked context about a user. After macOS permissions are granted, it captures focused activity from applications on the user's Mac and builds a portable HUMAN.md-style personal context model, storing the data locally and allowing it to be inspected, corrected, exported, or deleted. The runtime progressively organizes sourced observations and events into relationships and changes, supported patterns, higher-order structures, and an integrated current model. It retains source receipts for important claims so later evidence can strengthen, revise, or overturn inferences, then exposes this context through MCP to trusted clients including Claude Code, Codex, Cursor Agent, Claude Desktop, and other compatible tools.
quill is a fully local, ultra-minimalist macOS meeting recorder and transcriber developed by digimata. It runs as a single Swift menu-bar binary and uses macOS Core Audio process taps to capture microphone and system audio as separate CAF tracks without a virtual audio device or kernel extension. After recording stops, quill transcribes each track on the device, shifts the results by their start offsets, and merges the timestamped segments into a speaker-tagged transcript. The separate microphone and system tracks provide two-party diarization without a speaker-identification model. Each session produces the audio tracks, metadata, canonical JSON and Markdown transcripts, and a transcription log in a recordings directory. The default on-device transcription engine is Parakeet TDT 0.6B v2 through FluidAudio's Core ML port. It requires macOS 15 or later, supports optional launch-at-login operation, and can resume unfinished transcription jobs after relaunch. A WhisperKit fallback or re-transcription engine is planned.
riddle is a Rust application that turns a reMarkable Paper Pro into a diary-style AI interface. It reads pen strokes from the tablet, waits for an idle pause, commits the page as a PNG to a resident oracle LLM process, and streams the response back as animated handwriting. The handwriting pipeline rasterizes the response, applies Zhang-Suen thinning, traces the result into single-pixel pen paths, and replays those paths stroke by stroke. The app supports a qtfb display backend for running inside xochitl and a quill-based takeover backend that drives the vendor e-ink engine directly; the prebuilt bundle uses takeover mode. It can be configured with an OpenAI-compatible API key or used with pi, according to the repository instructions. Installation requires a reMarkable Paper Pro in developer mode with xovi and AppLoad, or the remagic installer. Takeover mode stops the normal reMarkable interface, runs as root, and takes control of the display; the repository says it has been tested on the Paper Pro ferrari model with OS 3.26–3.27 and warns that installation modifies the device and may not work on other models or versions.
scriptc is an experimental TypeScript-to-native compiler from Vercel Labs. It uses the TypeScript compiler for parsing and type checking, then emits typed IR, readable C, textual LLVM IR, native assembly and object files, native executables, or WebAssembly modules. Static builds include a small native runtime and do not require Node.js or a JavaScript engine when run; code that cannot compile statically is reported through diagnostics, and coverage reporting identifies the statically compilable portions and blockers. For npm packages and other dynamically typed code, an optional dynamic mode embeds quickjs-ng. The compiler itself requires Node.js 24 or newer and targets macOS, Linux, Windows, and WebAssembly through WASI Preview 1.
scroll-world is an agent skill for Claude Code, Codex, and other SKILL.md-compatible agents that builds scroll-scrubbed 3D world landing pages for brands and industries. It generates cohesive isometric diorama scenes and image-to-video camera flights, linking scenes through first/last-frame conditioning so the camera appears to move continuously as the visitor scrolls. The skill provides prompt templates, an AI image and video rendering pipeline using Higgsfield, Monid, Seedance, Kling, or Codex image generation, and a framework-agnostic vanilla JavaScript scrubbing engine that can be used with plain HTML, Next.js, Vue, or a Python-served page. It can also render a separate portrait chain for mobile devices and requires external rendering services or CLIs, ffmpeg/ffprobe, and Python with Pillow.
An agent-facing collection of design and engineering skills by Emil Kowalski for building and reviewing user interfaces. Its guidance covers animation curves, durations, properties, gestures, haptics, screen transitions, motion placement, borders, shadows, typography, Apple interface principles distilled from WWDC talks, UI-library selection, modern Swift, prototyping, and product-interface details. The skills can create animations from scratch, review animations against defined rules, audit a codebase and produce prioritized implementation plans, identify worthwhile opportunities for motion while flagging what should not be animated, and build multiple UI variants navigated through a switcher. A React Native and Expo skill covers gestures, sheets, haptics, screen transitions, and keeping motion off the JavaScript thread; other skills guide use of the Sonner toast library and library selection. The collection is installed with `npx skills@latest add emilkowalski/skills` and is based on Kowalski’s stated experience at Vercel and Linear.
The Fable Method is an agent-workflow repository that distills the Fable Workflow into skills that models can run. Its process classifies the request, defines completion with named verification, gathers evidence from primary sources in parallel, commits to one recommendation, changes the smallest correct thing, verifies the result by observation, and reports the outcome with caveats. The repository provides four related skills—fable-method for thinking, fable-loop for acting, fable-judge for evaluating results, and fable-domain for generating domain adapters—along with evaluation cases, transcripts, judge outputs, and logs. The project reports testing the method across fifteen evaluation rounds and more than 260 agent runs, with judges checking diffs and execution rather than relying on agent reports.
thinking-orbs is a React component package providing dotted animated loading indicators for AI and agent interfaces. It offers nine hand-tuned states— including working, searching, solving, listening, connecting, weaving, composing, breathing, and shaping—rendered with plain 2D canvas rather than WebGL or filters. The indicators have separate 64-pixel chat-avatar and 20-pixel inline-text designs, automatic or pinned light/dark themes, adjustable speed, pause control, and pass-through canvas properties. It includes per-state accessible labels, renders a static frame for prefers-reduced-motion, pauses offscreen instances through IntersectionObserver or when the tab is hidden, and resumes them in phase using a shared clock. The package is distributed through npm under the MIT license.
TurboFieldfare is a model-specific Swift and Metal runtime for running the instruction-tuned Gemma 4 26B-A4B model on Apple Silicon Macs, including systems with 8 GB of RAM. It keeps the model’s shared 1.35 GB core and FP16 key-value cache in memory while streaming the expert weights selected for each token from an SSD, allowing inference with about 2 GB of resident weights instead of loading the approximately 14.3 GB model. The model uses MLX affine 4-bit weights with group size 64, an 8-bit router, 4-bit shared and routed experts, and about 3.88 billion active parameters per token. The project includes a native macOS app, command-line interface, Swift runtime library with Metal kernels, decode service, streaming model installer and verifier, and experimental loopback OpenAI-compatible Chat Completions server. These products share a repacked `.gturbo` model directory; the installer streams byte ranges from a pinned Hugging Face revision, repacks them without materializing the source checkpoint, and validates the resulting manifest and file hashes. The app and CLI support instruction and completion generation with optional system guidance but do not execute tools; the server accepts function-tool declarations and returns model-produced tool calls for client authorization and execution. An optional companion vision tower adds image input on M2 or newer Macs, while text-only inference remains available on M1. The package is arm64-only and targets macOS 26, Metal 4, and Swift 6.2 or newer, with approximately 14.3 GB of storage for the text model plus about 1.1 GB for the image pack.
An open-source AI agent skill for Claude Code, Codex, and similar coding agents that produces cinematic product, marketing, launch, and demo videos with Remotion. It provides shot recipe cards, motion previews, a reusable video template, and workflows for storyboarding, real page capture, animation, 2.5D camera moves, sound design, beat-synced cuts, and visual checks. The repository includes native Remotion components driven by normalized progress values, a gallery for browsing motion previews, and an optional JianYing project export that separates shot plates, captions, sound effects, and background music into editable tracks. It can be installed through the skills CLI, by cloning the repository, or by linking it into a Claude Code or Codex skills directory.
Wardrobe is a local clothing-library and outfit-generation application that uses OpenAI image and vision APIs. It detects garments in photos with the OpenAI Responses API, creates clean clothing cutouts with the OpenAI Images API, and can generate modeled editorial previews using a local reference photo. The application stores original images, generated images, processing jobs, and a JSON database in its local data directory, and provides drag-and-drop, paste, editing, review, regeneration, and approval workflows. The repository includes Codex skills for importing clothes from a folder and for generating modeled outfit lookbooks. The importer reviews cutouts and modeled images before writing them to the local library; the outfit skill curates, generates, verifies, and saves complete looks. It runs as a local npm application, requires an OpenAI API key and a PNG model-reference image, and is licensed under MIT.
Searchable transcript of GitHub Trending Monthly #9（2026.07） — Github Awesome (15:26). 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 Github Awesome. 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 Welcome back to GitHub Awesome. Today, we're looking back at 35 popular projects from July and checking which ones developers actually got excited about last month. Let's get into it. Colibri runs enormous mixture of experts models by treating VRAM, system memory, and NVMe as one weight hierarchy. Its pure C engine keeps dense weights resident, caches experts based on measured routing heat, and pre-fetches the next layer while computation continues.
00:28 It supports CPU, CUDA, metal, and even dual SSD streaming. The clever part is placement, not smarter inference. Open Worker is a desktop co-worker that produces actual files and app updates instead of stopping at a chat response. It can work across local documents, the terminal, and more than 25 connectors, then schedule recurring jobs with transcripts.
00:54 Consequential actions, including messages, calendar edits, and shell commands, require approval. IMG3JS rebuilds or builds a reference image as procedural 3JS code instead of downloading or extracting a mesh. It inventories defining details, generates the model in staged passes, and compares each render against the source before moving forward. The output is a TypeScript 3.group with pivots, sockets, and colliders ready for animation.
01:23 Codex Security brings application security scanning to a CLI and TypeScript SDK. Pointed at a code base, and it finds, validates, and helps fix vulnerabilities, while scan history compares runs by root cause and tracks reopened or resolved findings. Bulk scans can run in containers against pinned Git revisions with resumable jobs and optional AppArmor hardening.
01:47 Scroll World turns a brand story into one camera flight controlled by scrolling. It generates connected isometric scenes, creates dive-in and transition clips from matching boundary frames, then wires them into a portable vanilla JavaScript scrub engine. It can render a separate portrait chain for phones instead of cropping desktop footage. The result works across frameworks, but generation uses paid image and video services.
02:15 Marble skill taxonomy turns elementary curriculum standards into a machine-readable prerequisite graph. It's JSON data set maps 1 590 micro topics across eight subjects, connects them with 3,221 hard or soft dependencies, and includes plain language mastery evidence plus age ranges. There's no runtime or database to deploy, so apps can load the files directly.
02:40 No AI slop is an editing skill for removing canned machine writing habits without sanding away the author's voice. It flags patterns like fake contrasts, vague attribution, dramatic fragments, and inflated claims, then makes the smallest useful revision. There's also a detection mode that quotes every pattern it finds without pretending to prove who wrote the text.
03:04 Open science puts literature review, coding, experiments, and write-ups inside one browser workspace. For machine learning or physics work, its research agents can search papers, query scientific databases, run code on connected compute, and keep artifacts with their provenance on disk. You can switch among hosted or local models using your own keys.
03:27 Video shot craft gives CodeX or Claude Code a production playbook for building product films in promotion. It includes 104 shot recipes with real timing and easing parameters, 161 motion previews, and a complete 36-second template that swaps in your screenshots and branding. The workflow covers storyboarding, page capture, beat synced editing, sound design, and visual checks.
03:53 Canvas UI puts WebGL effects over real interactive HTML instead of replacing the interface with a dead canvas. It's 33 components cover fluid simulations, optical distortion, particles, and 3D object treatments with versions for React, View, Svelte, Solid, Preact, and vanilla JavaScript. Components copy directly into your project through a Shad CN compatible registry.
04:17 Agent ENV runs large fleets of agent sandboxes as Firecracker micro VMs instead of heavyweight always-on containers. It loads OCI images on demand, snapshots memory and file system changes incrementally, and can fork a running environment into independent branches for parallel work. The maintainers report resume times under 50 milliseconds and pauses under 100.
04:42 Script C compiles ordinary TypeScript into native executables that don't need Node, V8, or a JavaScript engine. A coverage command shows exactly what can compile statically, rejects unsupported constructs with rewrite hints, and can embed QuickJS only when you explicitly enable dynamic mode. The maintainers report roughly 2 milliseconds startup and binaries around 170 to 200 kilobytes for static programs.
05:10 ESP32 AI fits a 28.9 million parameter language model onto an $8 ESP32 S3 with every token generated locally and no network connection. The trick is memory placement. A 25 million parameter embedding table stays in flash, while only the rows needed for each token move into faster memory. The maintainer measures roughly 9.5 tokens per second. It's an architecture experiment, though.
05:38 Quill is a one-click macOS meeting recorder that keeps both audio and transcription on your machine. It captures your microphone and system audio as separate tracks, then merges their on-device transcripts into timestamped, speaker-tagged markdown and JSON. Interrupted transcription jobs resume after relaunch. And a post-processing hook can feed finished sessions into your own workflow.
06:03 Harness engineering is a field guide for improving coding agents without changing the model. It's argument is that better context, tools, permissions, and executable checks let an agent recover intent, operate the real system, and improve the result. The repository packages that thinking into thesis notes, source material, playbooks, and agent-facing routing files.
06:26 Cloud of Duty is a browser FPS generated by a fleet of coding agents, but the interesting part is the engineering reality check. It's 3JS world has procedural textures, weapons, audio, enemy AI, and no external art assets. Reproducible screenshot tests catch single-pixel changes, while gameplay profiling exposed shader compilation stalls that static benchmarks missed.
06:51 Jacob Crail's skills collection gives coding agents a structured interface review checklist instead of vague requests to make it polished. Six focused skills cover typography, color, accessibility, layout, UI details, and product writing, while better interface coordinates them into one review. You can run a quick pass or target a complete flow such as checkout.
07:15 Turbo Field Fair runs Gemma 4's 26 billion parameter mixture of experts model on an 8 GB Apple silicon Mac without loading all 14.3 GB of weights into memory. It's Swift and metal runtime keeps the shared core resident, streams selected experts from SSD, and exposes a native app, CLI, plus loopback API. The maintainer reports roughly 2 GB of working memory and 5.1 to 6.3 tokens per second on an M2 MacBook Air.
07:45 The Fable method turns careful agent work into a literal sequence. Classify the request, define proof, gather evidence, make one decision, change the smallest correct thing, and verify the result. Four installable skills handle execution, adversarial review, and domain-specific adapters. The repository also keeps raw evaluation logs, including failures and null results.
08:09 Knockoff is a Chrome extension that cleans up Amazon search results by spotting pseudo brands, the random all-caps storefront names that exist mainly to game brand registry. The pain is buying a charger, tool, or cable and realizing every result looks fake. Knockoff runs locally, uses allow lists, block lists, known brand data, and name heuristics, then hides, dims, or labels suspicious listings without sending your shopping path to a server.
08:38 Bento is a PowerPoint alternative where the presentation and its editor live inside one HTML file. Open it in a browser, change the deck, then save the same file back with no account or installer. Slides can carry fonts, images, charts, and animations, while the document data stays readable JSON. It even supports encrypted collaboration through an optional blind relay.
09:01 That's a refreshingly portable way to own a presentation. Peneko gives AI a shared canvas instead of trapping every idea in a chat box. Write an equation, sketch a diagram, or circle part of your work, and it reads both the marks and their spatial relationships before answering beside them. AI results stay as movable drafts until you accept them, while sparse tiles keep the huge canvas lightweight.
09:27 It's especially useful when translating visual thinking into text would destroy half the context. Riddle turns a remarkable tablet into Tom Riddle's enchanted diary. You write with the pen, the page drinks your ink, and a reply writes itself back in flowing script. Chatting with an AI on an e-ink device usually means a laggy keyboard and a chat bubble UI that kills the whole feel.
09:51 This reads your raw handwriting straight off the pen, sends the committed page to a resident on-device model, and animates the answer stroke by stroke. $40,000 gets you noticeably closer to Claude Opus running entirely on your own hardware. No subscription, no API key. This is one engineer's actual build log for getting there. Four RTX Pro 6000s linked through PCIe switches, so the GPUs talk directly to each other instead of routing through the CPU.
10:21 He walks through the exact BIOS settings, kernel flags, and ACS fix that took his card-to-card bandwidth from broken to full Gen4 speed. Snap a photo of a pile of clothes on your bed, and wardrobe pulls out every garment as its own clean product cutout, no background. Then optionally drapes each piece onto a modeled photo of you. The app keeps its image library and JSON database local, while bundled Codex skills can import entire folders or assemble new outfit ideas.
10:51 It feels like a personal inventory system built around clothes you actually own. Jacobian Lens is a research tool for asking what an internal language model activation is preparing the model to say. It transports a vector from any layer into the final output basis, converts it into ranked vocabulary tokens, and renders an interactive grid across layers and positions.
11:15 You can fit lenses on open-weight hugging face decoders or apply saved ones. It's companion code for an interpretability paper. General's Mac iOS iPad isn't an emulator wrapping the old Zero Hour. It's the real 2003 engine compiled straight for ARM64. Getting a 20-year-old Windows RTS running natively on an iPad, touch controls and all, means solving problems nobody wrote down.
11:44 It renders through a DirectX 8 to Vulcan to metal chain, ships real RTS touch gestures like drag box select and pinch zoom, and its docs folder is a full engineering log of every bug it took to get there. Ask an agent to rebuild an object from a photo in 3.js, and it one-shots a blobby mesh that's kind of the right shape but loses the details that made it recognizable.
12:06 This Codex plugin makes it sculpt instead. It's explicitly not photogrammetry. It guides the agent from blockout to fine surface, writing pure procedural code with real pivots and sockets, so the result is animation-ready. Thinking Orbs gives AI interfaces a loading indicator that says more than still working. It ships six-dotted canvas animations for states like searching, solving, listening, and composing, each tuned separately for avatar and inline sizes.
12:38 There's no WebGL, and the theme follows your app automatically. It also respects reduced motion settings and pauses offscreen. Lingbot World V2 is a generative world model built for long interactive video worlds. Not another clip generator that falls apart after a few seconds. You ask for a character action, and the scene forgets where it was headed.
12:59 It's causal model generates frames chunk by chunk with KV caching. The real-time variant targets 720p at 60 FPS, and pilot plus director agents plan actions while adding new scene elements. GC Minimal Zine Poster is a Codex skill that turns a loose theme or reference image into a restrained editorial poster, then generates the raster result. Its visual rules are unusually specific.
13:26 A vertical aged paper canvas, lots of negative space, one small subject, and a single bright color anchor. It also layers in Xerox, risograph, or halftone texture. So, outputs feel printed and imperfect instead of like glossy ad mock-ups. Beautify GitHub readme is an agent skill that redesigns a repository homepage around the project itself, rather than dropping in another generic template.
13:49 It reads the repo first, moves the clearest proof forward, then separates decorated SVG assets from searchable markdown and copyable commands. You can request a full readme overhaul, or just the visual pieces with local previews before anything is published. Making a web video react to hover or state means seeking to timestamps, and it stutters at every seam.
14:14 Avel is a new format. One .avl file packs decodable motion units and a state graph, so the browser runs a decoder timeline forward instead of seeking. Hover and state become graph routes, not hand-timed seeks. It has packed alpha transparency and an image fallback in one web component. A micro is a gallery of copyable React micro transitions, instead of another design demo that leaves you rebuilding every animation from scratch.
14:42 Browse its live button and control previews in grid, list, or matrix layouts, then copy the generated React, Tailwind, and motion code. It also handles light dark transitions and gives you a defined place to add your own interactions. Persome builds a local personal model from the apps you use, instead of leaving every agent blind to your work context outside its chat window.
15:08 It reads focused macOS accessibility data, uses on-device OCR only as a fallback, and exposes receipts-backed memory over MCP to trusted clients. You can inspect, correct, or delete the model. Screenshots are encrypted, no telemetry. >> [music]