ABYSSAL is an open-source, browser-based real-time ocean and extreme-weather simulation built with Three.js, WebGL2, GLSL3, and Vite. It procedurally generates the ocean, clouds, rain, spray, foam, lightning, hurricanes, waterspouts, whirlpools, rogue waves, and tsunamis on the GPU without external textures, meshes, HDRIs, or sound files. Its ocean uses three FFT wave cascades driven by a JONSWAP spectrum and a per-frame butterfly IFFT; a projected screen-space grid renders the water, while analytic disaster height fields deform its surface. Volumetric clouds are raymarched from Perlin-Worley noise and a procedural weather map, with atmospheric-scattering lookup tables providing shared lighting for the sky, clouds, ocean, and spray. Temporal anti-aliasing, adaptive quality scaling, bloom, depth of field, motion blur, and tone mapping form the post-processing pipeline. The browser experience includes an automatic cinematic storm sequence and a sandbox with free-flight controls.
cc-prune is a Python command-line tool for surgical context recovery in Claude Code sessions. It removes already-processed tool output from a Claude transcript while preserving conversation text, reasoning, thinking blocks, transcript structure, and tool-call relationships; it does not summarize or truncate the retained content. The workflow measures actual API usage, inspects transcript byte buckets, optionally audits individual results, and clears selected buckets such as tool results, inputs, attachments, or orphaned command output. Clear operations use a character-length threshold, can protect a tail of recent records, create numbered snapshots, validate invariants, and support restoration. Its validation checks include unchanged record and UUID/parent-UUID chains, byte-identical thinking signatures, and matching tool-result and tool-use identifiers. The tool operates on Claude Code JSONL transcripts under the user's .claude/projects directory and requires the session to be stopped before editing. It supports usage verification after pruning, snapshot listing and restoration, and density calculations based on measured token counts. The project states that the Claude Code transcript format is undocumented and version-sensitive, and that cleared results may need to be regenerated; it is distributed under the MIT license and requires Python 3.9 or later with no dependencies.
CDAF is an open sidecar format and toolkit for video that stores a timestamped plain-text description beside the corresponding video file. Its Python library and CLI can generate, parse, validate, read, and report the status of `.cdaf` files, while an agent skill teaches video agents to check for a matching sidecar before processing footage. Each sidecar contains a minimal versioned header with the video filename, SHA-256 hash, byte size, duration, generator, and creation time, followed by sections such as summary, timestamped segments, transcript, on-screen text, and tags. Conforming tools verify the video's freshness and refuse to use a stale sidecar after the video changes; the format is model-agnostic even though the included generator uses the Gemini Files API, with optional local-model support. The repository includes a normative specification, reproducible sidecar-versus-direct-video benchmarks, an agent skill installable with `npx cdaf-skill`, and CLI commands for generation, validation, reading, and status checks. The core validation functions require only the Python standard library; generation requires Python 3.10 or later and a user-supplied Gemini API key. The project is licensed under MIT.
Codewhale is an open-source coding agent for the terminal, built in Rust and independently maintained. It reads repositories, edits files, runs commands, inspects results, and works toward user-defined goals. It connects to hosted providers or local models through Ollama, vLLM, or SGLang, supports switching providers and models during a session, and can run interactively in a terminal UI or non-interactively with the `exec` command. Its control mechanisms include read-only planning, Ask, Auto-Review, and Full Access approval modes; `/undo` reverts the last turn and `/restore` returns the workspace to an earlier snapshot. It also supports saved sessions, durable goals, reviewable workflows, agent coordination, MCP servers, skills, hooks, and configurable agent roles. The project runs on the user's machine with the access granted to it, with optional operating-system sandboxing where supported, and is distributed under the MIT license.
Defragger is a graphical Linux filesystem defragmenter written in Rust with Qt Quick/Kirigami. It visualizes storage blocks and performs online XFS defragmentation through kernel extent-management calls.
Editable Visual Design is an open-source, coding-agent-driven toolkit for creating editable visual artifacts from prompts. A persistent coding agent interprets a brief, plans a composition, generates assets, implements the design in semantic HTML, renders and observes the result, and applies repairs. An image model supplies visual direction for composition, hierarchy, color, and spatial relationships, while the delivered artifact rebuilds typography and layout in HTML rather than shipping reference pixels. The resulting artifact contains real text, independent assets, selectable and movable layers, a visual editor, rendered PNG output, an animated layer breakdown, and a replayable creation process. Deterministic checks cover the canvas, fonts, layer contracts, rendering, and editor round trips. The repository distributes the workflow as two Codex skills, including editable-design for fixed-canvas designs and a separate html-to-pptx skill for converting compatible HTML designs into editable presentations.
FixAnything is an open-source video-refinement tool that repairs rendering artifacts from 3D representations, including 3D Gaussian splats, NeRFs, meshes, and sparse point clouds. It repurposes the pretrained Wan2.1 video diffusion model with minimal modification and fine-tuning, using a FixAnything LoRA to turn a rendered camera-path video into a refined video. The inference pipeline accepts a video file or folder of frames; its documented workflow uses 61-frame renderings resized to 832×480 and writes a generated video, the resized input, and a side-by-side comparison. An optional pipeline reconstructs scenes from a small set of photographs with MapAnything, renders the reconstruction along an interpolated camera path, and then refines that rendering. The repository contains inference code and downloads the underlying Wan2.1 model and FixAnything weights; the code and weights are released under the Apache 2.0 License.
genart-skill is a Claude Code plugin that provides an AI coding skill for generative art. It teaches deterministic, hash-seeded compositions, including seeding a pseudorandom number generator from a token hash, using named sub-streams, rendering the same composition at different resolutions, and designing traits and rarity tables for editions. It covers Canvas 2D, p5.js, Three.js/WebGL, and SVG, with guidance on ethics, platform-specific workflows, debugging, keyboard shortcuts, PNG and video export, and print and pen-plotter output. The plugin includes scripts for checking same-machine repeatability, distinctness, global-state isolation, and feature stability; rendering a hash-specific PNG or contact sheet; measuring rarity across a census; and exporting batches of images. The scripts can be run in a project and use Playwright when browser rendering is required. The repository explicitly limits its determinism claims, noting that GPU shader compilers, floating-point behavior, rasterizers, MSAA, and JavaScript engines can prevent cross-machine equivalence. It is installed through the Claude Code plugin marketplace and is maintained by Camille Roux. The repository says that CI tests the scripts against a known-good fixture and a deliberately broken variant, while a monthly check verifies that platform-documentation URLs still resolve. It is licensed under the MIT License.
A Claude Code skill for converting scripts into MiniMax H3 shot lists and directing character performance. It breaks scenes into shots based on emotional-beat density, using controlled comparisons and documented inferences about MiniMax H3 output; its central rule is to split shots that contain too many facial-expression beats because H3 may otherwise produce a frozen face. The skill also covers dialogue-tag effects on frame allocation, camera placement, observable body actions for silent characters, reference-image use for shapes, edit-based handling of pauses and emotional transitions, PSNR-based verification, dubbing constraints, tail degradation, and a six-step breakdown workflow. It is installed with `npx skills add https://github.com/phileiny/h3-storyboard-skill --skill h3-storyboard` or by copying the skill directory into `.claude/skills/`. The repository distinguishes controlled findings from partly verified and inferred rules, and describes its scope as serialized short-form drama produced with local ComfyUI and MiniMax H3.
HexStellar is an agent-first computational platform whose Cortex service executes structured optimization, decision, scientific-computing, and verification requests through a Python CLI and API. An agent submits a JSON formulation for problems such as QUBO/Ising optimization, maximum cut, traveling-salesperson routing, facility assignment, mixed-integer optimization, selection, ranking, scheduling, coloring, and business-rule feasibility; the managed service returns a structured result with execution metadata, a receipt, and an assurance label distinguishing certified optima, heuristics, operations, and abstentions. The CLI also provides free validation, estimation, service re-checks, local witness recomputation for supported families, reproducible seeds and model versions, batch and compressed-binary transport, and an MCP server over standard input/output, with read-only mode for free analysis and verification tools. The public package is a zero-dependency Python thin client: the proprietary solver runs on HexStellar-managed infrastructure rather than inside the package. A separate enterprise runtime is licensed for compatible customer-controlled compute paths under NDA; it has no public runtime download in version 1.0.
hqtui is a TypeScript terminal UI library developed by profullstack for building btop-style dashboards with a one-import API and zero runtime dependencies. It renders directly through ANSI/VT sequences using a typed-array cell framebuffer, differential rendering, truecolor with automatic 256- and 16-color fallback, and Braille or block graphics, without ncurses, a browser DOM, React, native add-ons, or network access. Its API provides layouts, panels, tables, trees, logs, meters, graphs, forms, dialogs, themes, mouse input, and responsive terminal handling. A headless renderer can produce text, screen, ANSI, or HTML output for testing without a TTY. Bun is the default runtime; Node 22.6 or newer is supported, with best-effort Deno support. The project is MIT-licensed and includes a demo dashboard with real or simulated system metrics.
Keyword Pro is a standalone, local-first keyword research console developed as an independent open-source project. It uses the user's own DataForSEO credentials rather than a hosted account, signup, or subscription system, and stores research sessions in a local database. Its guided report can make up to 19 keyword-research calls covering search volume, CPC, difficulty, intent, trends, SERP competitors, and related terms. The application supports 94 markets and 46 languages, validates country-language combinations, skips unsupported paid calls, and provides an advanced mode with 332 keyword-oriented DataForSEO endpoint definitions. It displays estimated costs before paid runs, saves completed reports for later reopening without another API call, and exports results as PDF, CSV, and JSON. The application runs locally with Node.js, pnpm, OpenSSL, and Podman or Docker, binding to 127.0.0.1 by default for a single local user. DataForSEO credentials stored by the application are encrypted with AES-256-GCM. The project is distributed under the Apache License 2.0.
LightNav-0 is an open-source generalist embodied-navigation model from the Light Origins Team. It uses a pretrained Qwen3-VL backbone with a single egocentric RGB stream and natural-language instructions to control humanoid, quadruped, wheeled, and aerial robots, transferring across tasks, embodiments, and scenes without per-task or per-benchmark fine-tuning. The model extends the vocabulary with dual-channel pointing tokens and residual vector-quantized action tokens rather than adding navigation-specific prediction heads. At each step, it emits an affordance point representing a feasible local direction or waypoint and an object point representing the goal, followed by three action tokens that decode to ten future SE(2) waypoints for an embodiment-specific low-level controller. Its temporally aware history compressor samples older frames less frequently, pools them more coarsely, and preserves ordering with timestamp tokens while bounding the visual context. The repository includes a released checkpoint and action decoder, command-line prediction, a vLLM-based server with WebSocket streaming, a MuJoCo simulation demo, evaluation harnesses, and a ROS 2 deployment stack with adapters for the Unitree Go2 and LimX TRON 1. The project is released under the Apache License 2.0; the included EVT-Bench is separately licensed CC BY-NC-SA 4.0.
loadersz is a framework-agnostic, dependency-free browser library of Canvas loading animations, distributed as a native Web Component with an imperative TypeScript API. It provides 150 motion states, including working, searching, connecting, solving, and racing, and supports configurable size, speed, density, color, theme, pausing, and reduced-motion behavior. The root import registers a switchable <loadersz-loader> element, while individual state entry points such as loadersz/racing include the Canvas core and one geometry builder for smaller fixed-state bundles. The API can also attach a fixed or switchable loader controller to an application-owned canvas and supports lifecycle methods such as setOptions and destroy. Typed integration adapters are provided for React, Vue, Svelte, and Angular; plain HTML, Lit, Solid, and Qwik can use the native element directly. The library pauses in hidden tabs, respects prefers-reduced-motion by default, and caps Canvas device-pixel ratio at two. The package has no regular runtime dependencies, and its repository is licensed under the MIT License.
markdown-graphs is a collection of React components for ASCII-style tables, charts, and diagrams in MDX. It provides more than 30 graph components, including tables, bar charts, sparklines, heat maps, Gantt charts, timelines, trees, and flow diagrams. Components render inside dashed frames using text glyphs and blocks rather than SVG, with optional single-, dual-, or three-accent palettes. The project distributes source files through a shadcn registry: users can add individual components or the full collection to an existing shadcn project, then import them locally. It is not an npm package and requires an existing shadcn project and Motion. The repository is licensed under MIT.
Monocolor Editorial Print is an agent skill for generating one-ink or controlled two-ink editorial images for posters, zines, portraits, packaging, and visual field notes. It accepts a theme, phrase, object, article idea, or supplied photograph and produces a raster image when image-generation tools are available, the exact production prompt, and a short recipe describing the print mode, palette, layout, typography, process, and originality changes. The skill identifies the subject and intent, selects a layout family, assigns one or two ink plates, composes the page with visible paper and an asymmetric grid, and generates and inspects the result. Its visual system uses adaptive white, gray, or beige substrates; halftone, risograph grain, cyanotype exposure, or photocopy breakup; responsive typography; and controlled negative space. Without image generation, it returns the prompt and states the limitation.
Noty is a native macOS sticky-notes app built with Swift, SwiftUI, and AppKit. Notes rest as a pill on the screen edge and fan out into tabs when the pointer reaches it, avoiding floating windows and a Dock icon. It supports inline checkbox tasks, quick capture, search, archiving, Markdown editing and export, import, multiple displays, customizable global shortcuts, and automation through a noty:// URL scheme. Notes are stored in a local SQLite database, with note bodies encrypted using AES-GCM through CryptoKit; titles, colours, and timestamps remain in plaintext. The app requires macOS 15 or later, does not require an account or system permissions, and reports no server, analytics, telemetry, or tracking SDK use. It is distributed as a macOS application and DMG, is MIT-licensed, and is not sandboxed or notarized according to the repository.
Omakade is a local-first Linux game library for Omarchy, developed as an independent community project. It discovers games from installed Steam, Lutris, Heroic, Faugus, RetroArch, Epic, GOG, and Amazon sources, then presents them in a cover-focused library without modifying launcher data. Games are launched through their owning platform, while discovery, browsing, artwork, local Steam achievements, search, favorites, collections, tags, completion states, source filters, and controller navigation work offline. Optional Steam Web API synchronization adds public owned games and ready-to-install views, and optional IGDB credentials provide critic aggregates and game-length estimates. It is distributed for Omarchy and Arch as a package and stores its library in a local SQLite database.
Procedura is an open-source agentic 3D-modeling tool from SpatiaOS that converts text prompts into editable procedural assemblies rather than point clouds or triangle meshes. It generates OpenSCAD source with named parts and typed mates, plans and builds parts incrementally, and can use reference images and optional Blender render feedback during refinement. Optional passes assign per-part PBR materials and plan articulation, exporting motion to OpenUSD and URDF with Isaac-based validation. It runs locally as a Bun/TypeScript pipeline using a configurable OpenAI-compatible, Gemini, or local model endpoint; it does not provide hosted inference or API keys. The pipeline uses a Manifold-capable OpenSCAD build to compile the generated programs and Blender for renders, and includes a web Studio for composing runs and inspecting intermediate artifacts. The repository is MIT-licensed.
An Apache-2.0-licensed serving setup for Qwen3.8-27B on a single 24 GB NVIDIA GPU, built around vLLM and exposing an OpenAI-compatible API with optional key authentication. Its preparation pipeline starts from a W4A16 AutoRound model, requantizes the language-model head and embedding matrices to int8, requantizes the MTP components, and applies vLLM patches for the serving stack. The runtime combines int8 tensor-core GEMMs with an fp16 recurrent state, continuous batching, split-KV verification attention, and speculative decoding through either Qwen's MTP path or the optional DFlash2 block drafter; DFlash2 can also draft from the request's cached context for document-reproduction workloads. The repository provides Docker Compose profiles for batch throughput and single-user latency, prebuilt container images, model-download and requantization scripts, benchmark and quality-test tools, and launchers for fast, long-context, and experimental KVarN cache modes. The standard configurations target roughly 64k to 150k tokens of context, while KVarN and alternative int4 KV-cache paths extend the claimed capacity to the 256k-token range with lossy KV quantization. The setup requires a recent NVIDIA driver and a compatible Ampere-or-newer GPU; the repository notes that its benchmark figures were measured on an RTX 3090 subject to a 250 W power limit.
Shrimply is a free, open-source video editor for creating videos from start to finish. Its main application is written in Rust with a GTK 4 and libadwaita interface; rendering uses Skia, wgpu, Slang, and CUDA, while FFmpeg and PipeWire handle media and a Python component provides the compute server. The project is pre-alpha software and warns of undocumented footguns, performance regressions, irreversible project-file breakage, crashes, resource leaks, and security limitations. It is licensed under GNU GPL version 3 or later, although some features depend on separately licensed NVIDIA components and model weights.
Skill Cabinet is a local catalog for agent skills installed on a machine. It scans user-level skill locations such as .agents, .claude, .codex, .cursor plugins, Hermes profiles, and other ~/.* /skills folders, then lets users filter skills by drawer, metadata, risk, invocation mode, and status; inspect rendered or source bodies, YAML frontmatter, extra files, symlinks, duplicates, origins, and broken links; and review disk usage. It runs with Node 20 or later through npx skill-cabinet, starting a server bound to 127.0.0.1 and opening the catalog in a browser. Users can quarantine skills to ~/.skill-cabinet/quarantine and restore them, or delete individual skills or groups; deletion removes folders, files, or symlinks from the scanned locations, while a symlink's target is retained. The project is distributed under the MIT license.
SkillRadar is open-source discovery, security, ranking, and routing infrastructure for Agent Skills and Codex. It discovers public SKILL.md files, parses their contents, performs conservative static checks for capabilities such as shell commands, dynamic execution, secret access, networking, package installation, filesystem writes, and deployment tooling, then classifies and ranks candidates in a safety-gated registry. D and Blocked candidates are kept audit-only and excluded from automatic routing. Its Codex plugin provides task-to-Top-3 routing, skill search, provenance and safety inspection, and a read-only Skill Budget Doctor. Routing can use a bundled offline registry and returns relevance, SkillRadar score, security grade, provenance, reasons, and match details without executing candidate repositories or depending on live GitHub discovery. The repository includes radar data, a matching system, a router-quality benchmark, a local registry UI, and daily bot-refreshed generated data; it is released under the MIT license.
An Apache-2.0 C++23/GGML port of SkinTokens and TokenRig for automatic 3D mesh rigging on CPU or Vulkan. It takes a static GLB mesh, uses a Michelangelo point encoder, Qwen3-based TokenRig policy, FSQ expansion, condition encoder, and SkinVAE decoder to predict a skeleton and per-vertex skin weights, then exports a rigged GLB. It can also generate weights for an existing static or animated skeleton, inspect GLB assets, and retarget recognized SOMA30 motion to the 52-joint Mixamo order or to a generated humanoid rig. The project provides a command-line interface, shared library, C11-compatible API, and a Go/WebGL demo; converted GGUF model weights are available separately, with optional Vulkan support.
slotstream is an open-source macOS local-LLM inference engine that runs Qwen3.8-Flash-Next, a 125-billion-parameter mixture-of-experts model, on Apple Silicon Macs whose memory cannot hold the model. Implemented as a single Swift binary using MLX, it keeps the dense trunk resident and reads routed expert weights from an SSD into a shared cache of slots, sizing that cache to available memory instead of loading the full model. It supports prefix caching for follow-up turns and exposes Ollama-compatible and OpenAI-compatible chat APIs, as well as a command-line prompt mode; the documented API subset supports streaming, CORS, and sampling options but not tools, images, JSON-schema output, or logprobs. The project requires macOS 14 or later, Apple Silicon, and roughly 110 GB of free disk space for the model weights. Its installer provides signed release binaries, and the repository can also be built with Swift and the macOS Command Line Tools. The source is MIT-licensed; the model weights are distributed under the Qwen community license.
Strata is a fast, keyboard-first file manager for Linux, developed for Omarchy and compatible GTK4 environments. It combines Miller-column navigation with thumbnail Grid and sortable Explorer views, supports fuzzy filename and path search, and provides previews for text, source code, images, camera RAW files, PDFs, audio, and video. The application loads directories incrementally and virtualizes large directory listings. Its recursive search indexes a selected tree on a background thread and retains the best fuzzy-ranked matches while results are published in bounded updates. It supports native Linux paths, including filenames containing invalid UTF-8 bytes, and GIO/GVfs locations such as SMB shares. Native parser-backed previews run in isolated Bubblewrap helpers with no network access, restricted mounts, resource limits, and fail-closed behavior on errors or timeouts. Strata is distributed as dynamically linked release archives for x86_64 and aarch64 Linux systems with glibc 2.39 or newer, and is licensed under GPL-3.0-or-later.
TokensBurned is a privacy-first tool that publishes AI coding activity on a GitHub profile through a live SVG card. It collects token counts and model metadata from AI coding harnesses, reduces raw sessions locally, aggregates usage into 15-minute buckets, and serves totals, activity heatmaps, harness/provider/model comparisons, and an optional site-wide rank. It supports native or official integrations for Claude Code, Codex, Gemini CLI, Cline CLI or SDK, and GitHub Copilot CLI, with OTLP or standalone CLI fallbacks for other harnesses. The tool uploads aggregate token counts, harness/provider/model metadata, hashed session identifiers, time buckets, and request counts, while its stated privacy boundary excludes prompts, responses, source code, tool payloads, repository names and paths, transcript files and paths, and API keys. Public display is a separate explicit opt-in tied to a verified GitHub account; users can build cards with selectable layouts, themes, heatmaps, comparisons, and rankings. It is distributed through harness plugins and an npm CLI, installs no cron job, daemon, proxy, or Git synchronization task, and is licensed under the MIT License.
TrustMeBro is a Go command-interception tool for controlled red-team testing of coding agents such as Codex, Claude Code, and pi. It uses PATH shims to intercept configured command-line tools without requiring a plugin, hook, or MCP integration; rules can spoof generated or fixed output, rewrite stdout from a real command while preserving stderr and its exit status, pass through to the real binary, or reject the call. Each decision is recorded in a timestamped JSONL audit log. Rules match command names, domains, DNS record types, argument globs, and regular expressions. The bundled DNS generators support dig, nslookup, and host, while custom shims can use fixed output, exit codes, and standard streams. On Linux, lab mode uses Bubblewrap to shadow PATH lookups and discovered absolute paths so an agent cannot bypass interception merely by invoking a resolved system binary. TrustMeBro is distributed as a single MIT-licensed Go binary with installation, configuration validation, status, rule-listing, and uninstall commands. Lab mode requires Linux and Bubblewrap and is explicitly an interception namespace rather than a security sandbox: it reuses the host filesystem, workspace, network, environment, and agent credentials. Outside lab mode, absolute paths, changed PATH environments, and in-process DNS clients can bypass command shims.
Utopia is a self-hosted enterprise knowledge platform from DeepLethe, implemented as a Rust binary with a PostgreSQL service using pgvector. It ingests documents and synchronized web, RSS, GitHub, and Jira sources, then combines Tantivy full-text search with pgvector retrieval through reciprocal-rank fusion; its chat answers include inline citations linking to source passages and can use OpenAI-compatible endpoints. Its knowledge model is a bitemporal graph: extracted entities and facts retain validity intervals, evidence, and revision history, so corrections link new facts to closed versions instead of overwriting them. An editable ontology drives three-stage entity resolution, temporal Datalog forward-chaining, provenance-preserving derivation, conflict detection, and ontology-driven queries over mounted PostgreSQL databases. Low-confidence extractions, entity-merge candidates, and cardinality conflicts enter a review queue, while decisions such as confirmations, rejections, merges, and graph rebuilds are recorded in a queryable ledger. The application includes a browser console, graph browser, ontology workbench, and multi-user knowledge-base permissions, and can be deployed offline with Docker. The repository identifies the project as version 0.1 and licenses it under Apache-2.0.
vol-rs is a Rust port of the Volatility 3 memory-forensics framework. It reads the same memory-image formats, runs the same plugins and options, and is designed to produce byte-for-byte equivalent output while running faster than the Python implementation. The project ports Volatility 3's 197 plugins and command-line interface, including plugin listing, CSV rendering, filtering, symbol handling, and configuration options. Its built-in components include symbol decompression, RC4, DES, AES, YARA matching, x86 disassembly, PNG writing, and tar-archive creation, avoiding the Python version's external dependencies for those functions. Symbol packs can be loaded from configured directories, and Windows symbols can be fetched from Microsoft's symbol server and cached locally. The repository reports identical plugin help for all 197 plugins, matching output across tested Windows and Linux captures, and substantially shorter benchmark runtimes on those captures. Testing remains incomplete for real macOS images, 32-bit images, several image formats, and additional kernel versions. vol-rs is distributed under the Volatility Software License 1.0 as an addition to Volatility 3 and is not affiliated with or endorsed by the Volatility Foundation.
Utopia stores facts with the time period in which they were true, supports hybrid search, cites source passages, and sends low-confidence extractions to review.
Neo is a local word processor for novelists with text-based chapter structure, a recovery area for deleted material, and timestamp-fingerprinted PDF exports.
Slotstream runs a 125-billion-parameter model by streaming its 104 GB weights from an SSD and exposing an Ollama-compatible API.
Searchable transcript of GitHub Trending Weekly #47: utopia, neo, slotstream, Codewhale, TokensBurned, mono-color-skill — Github Awesome (15:06). 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. This is GitHub trending weekly number 47. 34 trending open- source projects on GitHub right now. Let's go. Utopia is a self-hosted rag platform that tracks not just what a fact is, but when it was true. Ask who owned a project last Q3, and it shows the graph as it looked then, instead of overwriting old facts. Search is hybrid.
00:25 Full text plus vector embeddings and every answer comes back cited to the source passage. Low confidence extractions sit in a review queue instead of autoinserting. Neo is a word processor built for novelists by Hugh Howe. No accounts, no subscription. Everything saves as plain text on your machine. Chapters fall out of how you type. One enter for a paragraph, two for a section break, three for a new chapter.
00:51 A feature called Darlings keeps everything you cut in its tab instead of deleting it. It exports a PDF snapshot with a SHA 256 fingerprint, proof of when you wrote something. Slotstream runs a 125 billion parameter model Quen 3.8 flash. Next on a Mac that doesn't have enough RAM to hold it, streaming the 104 GB of weights off SSD. It speaks the Olama API, so it drops into your setup and it caches the prefix so follow-up replies don't slow down.
01:21 On a 48 gig M5 Pro Mac, the team clocks warm decode at about 12 tokens a second with a RAM floor at 16 gigs. Code Whale is a terminal coding agent written in Rust and it doesn't lock you into one model. Point it at a hosted API or run it fully local through Alama or VLM and switch midsession with /model. It gives you control. Three approval modes from ask first to full access and undo for reverting a turn and restore snapshots if it wrecks your workspace.
01:56 Tokens burned puts a live badge on your GitHub profile showing how much you use AI coding tools. No cron job or bot committing to your readme. It hooks into clawed codecs and only reads token counts and model names. No prompts, source code, repo names, or API keys leave your machine, just counts. You get heat maps and rankings against everyone else, and it stays private until you publish it.
02:21 This clawed skill turns any subject into an editorial print image, and it refuses the usual AI cliches. No monochrome filter over a color photo, no gradient poster, no centered template, no automatic vintage look. It works like an actual print run. One dominant ink covering 70 to 85% of the page, one accent ink for the rest. It leaves empty paper space, and every output ships with its prompt and recipe.
02:49 Lightnav is a navigation model that runs on one RGB camera, no depth sensor, no adometry, no panoramic, and the same checkpoint drives a humanoid, a quadruped, a wheeled robot, or a drone with zero fine-tuning. It reasons by pointing. one output for where it's feasible to go, another for where the target is. They report 68.5% success on VLNC2R, 79.5 on object navigation, and over 91% tracking a single target.
03:16 Abyssal is an ocean and weather simulator that runs in a browser tab with no textures, meshes, or HDRI. Everything comes out of math at runtime. It renders a clear day or a hurricane, eye and eyewall, plus water spouts, rogue waves, and volutric lightning with branching bolt geometry. The engine is about 400 kilob of JavaScript and GLSL. On a GTX 1650, they clock 108 frames a second in clear weather and 72 in a storm.
03:48 Trust me, Bro answers one question. Does your AI coding agent verify what a command tells it or does it believe it? It sits on your path and intercepts things like dig, nsookup, and host. Then spoofs a fake result, rewrites the real one, passes it through, or rejects the call. You can target it by command, domain, or rejects. And every decision gets logged to a timestamped file, so you can see what your agent was fed.
04:15 Hexellar gives AI agents a solver for problems text generation can't guess through. traveling salesman, graph coloring, scheduling, portfolio allocation. Every answer comes back labeled either certified optimum or huristic, so you always know which one you got. It plugs in as an MCP server, ships zero Python dependencies, and rechecks a candidate answer for free instead of billing twice.
04:40 Ni is a Mac OS sticky notes app that skips the floating window and the dock icon. It sits as a pill on the edge of your screen and moving your cursor there fans notes out as tabs even over full screen apps. You get checkbox task lists, a 10-second undo on anything you delete. And every note body is encrypted with AESGCM 256-bit stored locally with no account and no network access except an optional update check.
05:09 Shrimply is an open- source video editor built in Rust with a GTK4 and libida interface that looks like a modern desktop app, not a decade old fork. Rendering runs through Skia and WGPU with CUDA support, and it leans on FFmpeg and Pipewire for the media pipeline instead of reinventing codec support. It's GPLv3 with a Python backend handling compute services alongside the main app.
05:35 Skill Cabinet scans every folder where Claude and Codeex hide their installed skills and gives you one interface to browse them. You can search by name, description, or drawer. Read the source and front matter for anything you forgot you installed and delete skills one at a time or in bulk with keyboard shortcuts. One warning, delete removes the folder from disk.
05:56 There's no undo. Markdown Graphs gives you over 30 chart and diagram components for MDX and none of them render as SVG. Every bar, spark line, heat map, and Gant chart is built from character glyphs, blocks, and shading characters in Gist Mono, so it reads like it belongs next to your code. You don't install it as an npm package. You copy the source into a shad CN project.
06:23 Open SEO MCP skills gives Claude eight SEO skills, audits, keyword research, rank tracking, competitor gaps, back links, content briefs, and skips the part where Semrush or Hrefs charge you $120 a month for estimated data. It connects to your search console, G4, and Google Ads accounts, so what comes back is your ranking data, not a guess. There's even a skill tracking how much traffic you pull from chat GPT and perplexity.
06:51 HQA builds Btop grade terminal dashboards with one import. And it gets there by refusing the usual shortcuts. No endcurses, no browser DOM, no React, no native add-on, no network access ever. It renders through raw ANIE sequences at 30 frames a second, 15 over SSH. On a 160x 50 grid, they clock an unchanged frame at 0.068 068 milliseconds and a full redraw at 1.4.
07:18 Skill radar scans public skill.mmd files across GitHub and grades them from a down to blocked using static security analysis, not popularity. It matches on technical evidence, app router, tool calling instead of substring guesses. And every recommendation comes back with its relevant score, security grade, and reasoning, not a black box. It runs offline, never executing a candidate repo.
07:43 Ginart is a clawed code skill built by a generative artist for platforms like art blocks and highlight where a piece has to render from the same hash seed every time. It covers seeding a PRNG from a token hash, keeping the composition consistent from 400 pixels to 4,000, and designing trait and rarity tables for an addition. It works across canvas 2D, 3.js, JS and SVG and ship scripts that verify determinism before you mint.
08:12 Most AI design tools hand you a flattened image, so changing one word means regenerating the whole poster and hoping the layout survives. Editable design uses an image model for art direction, composition, color, hierarchy, but none of those pixels make it into the final file. An agent rebuilds the thing in HTML with text and selectable layers you can move, edit, and export independently.
08:35 And it logs a replay of every step it took. Vers rewrites Volatility 3 in Rust, and it's not a partial port. All 197 plugins are there with identical output down to the bite. The speed difference is the story. Detecting suspicious threads on Windows drops from 281 seconds to 47, about 600 times faster. And a Linux sweep goes from 3,130 seconds down to 68.
09:00 It ships symbol decompression, yara matching, and x86 disassembly. Nothing external to install. Strata is a Linux file manager built around Miller columns, so you see several folder levels at once instead of drilling in and losing your place. You can peek into a folder by hovering without committing to navigate. And it stays responsive in directories with up to 100,000 entries because everything's virtualized.
09:30 It handles file names with invalid UTF8, keeping the original path bytes and converting for display. Blink lives in your Mac's menu bar and shows the dev servers and iOS simulators running with port, framework, and project name colorcoded so you can tell a Nex.js JS server from a Rails one at a glance. Stop or restart any of them with one click. No terminal tab hunting required.
09:55 If a restart fails to come back, the row tells you why. Output one click away. Loaders gives you 150 different loading animations named motion states like working, solving, racing, imagining, all rendered in pure canvas with no SVG filters. Import the state you're using and it's 2.5 to 3.1 kilobytes gzipped treehaken down from the full library in one import.
10:19 It's framework agnostic at the core but ships typed adapters for react view and spelt. This clawed code skill fixes a bug in Miniax H3. Cram too many emotional beats into one shot and it drops the facial animation for a frozen frame. The creator measured it. A 7-second shot with nine beats came back at 42 bpsnr. frozen, but splitting those beats into three 2 to 3 second shots, drop that to 22 to 23dB, so the face moves.
10:47 It's built from tested H3 output, not the docs. AAD pulls your Linux game library, Steam, Lutris, Heroic, Epic, RetroArchch into one interface, and it never touches the launcher data. It reads it and hands the launch back to whichever platform owns the game. It works offline for browsing and launching, tracks local steam achievements and matches its look, pallet, fonts, transparency to the Omari desktop theme.
11:17 Live keyword pro is a keyword research tool that runs on your machine using your data for SEO credentials instead of a monthly SAS subscription. A guided report pulls up to 19 API calls, volume, CPC, difficulty, intent, trends, SER competitors, and shows the cost before it spends your money. Once a report finishes, it's saved locally, so reopening it costs nothing.
11:43 Defragger is a graphical defragment for Linux, something that's never existed. The creator built it because fragmentation variance was messing with reproducible storage layouts while testing high throughput logging, even on NVMe. It shows a visual block map and does X4 defragmentation online using the kernel's extent management calls, not a wrapper around other tools.
12:07 Fix anything takes a rendered video from any 3D reconstruction, Gaussian splatting, nerf, a mesh, even a sparse point cloud, and cleans up the artifacts without needing a separate model per representation. It's fine-tuned from a pre-trained W 2.1 video diffusion model with a Laura adapter. So, it's a generalist, not rebuilt from scratch per format. Procedura turns a text prompt into an editable 3D program, not a mesh or point cloud.
12:35 It writes open SCAD source with named parts joined by typed mates you can keep editing. Feed it a reference image for better fidelity and it shows the build before generating each part. It exports motion plans to open USD or URDF for anything meant to move. Quack lets you tell a small duck robot what to do in plain language and an LLM picks from 13 preset moves.
12:57 Walk, kick, sit, quack, not motor commands. Balance runs on the robot's onboard loop at 50 hertz with a safety stack of allow lists, confirmation, and kill switches. Their find and kick test needed three to eight decisions and 1 to two seconds, 10 out of 10 runs. Cedaf caches a plain text description next to a video file so agents stop reanalyzing the same footage from scratch every time.
13:24 It's hashverified, shot 256 plus byte size, so a stale sidecar gets refused instead of silently serving wrong info. On their benchmark, sidecar answers hit 20 out of 20 versus 19 out of 20 for direct analysis using 303 tokens instead of 3,66, about a tenth. Fitting a 27 billion parameter model onto one 24 GB gaming card usually means stopping at it loads.
13:51 This recipe pushes past that to 417 tokens a second batched and the wins are unglamorous in the best way. Quen's untied embeddings ship as two unquantized 2.5 gig matrices nobody bothered with. So re-quantizing both to int8 hands back 2.6 GB and on this architecture spare memory converts straight into batch size. CC prune clears out tool output. Claude's already processed from a session transcript without touching the conversation or thinking blocks.
14:22 Nothing gets summarized. No reasoning gets deleted. It audits what it's about to remove, takes a numbered snapshot so you can roll back, and checks token counts against the API instead of a local guess. Skin tokens.cpp takes a 3D mesh and rigs it, predicting a skeleton and painting the vertex weights, then writes out a GB you can drop into an animation pipeline.
14:48 It can also just weight an existing skeleton and retarget soma 30 rigs into 52 joint mixo format. It chains a point cloud encoder, a small quen 3 model making the decisions and a skinva decoder on CPU or Vulcan.