3dviz-pro-max is a standalone agent skill for turning plain-language ideas into interactive Three.js or Blender scenes. It is designed for Claude Code and Codex, and provides a ten-step workflow covering intent, object reasoning, representation, construction routing, first-view creation, meaningful behavior, capture, inspection, refinement, and reporting. The skill includes searchable recipes and knowledge records, runnable Three.js templates, reusable scene kits, style and camera guidance, and a capture helper that drives a built scene in a browser and writes PNGs and console logs for inspection. Its catalog covers subjects including environments, anatomy, mathematics, physics, architecture, logistics, robotics, and abstract systems. The supplied Python helpers search the catalog and drive an existing scene; they do not themselves render or judge scenes, and screen capture requires an optional Playwright installation. The repository includes 37 runnable studies and a site at 3dviz.dev. Its own code, data, and authored media are released under the MIT license, while bundled third-party assets retain separate terms. Blender 4.2 LTS or newer is recommended but not required; without Blender, the repository states that scripts still run but kit output is limited to T2.
Agent Launcher is a local desktop workspace for configuring and running existing coding-agent command-line interfaces. It detects installed CLI binaries, links or installs supported agents, applies account or provider profiles, and runs them in an embedded terminal or chat view without automatically reinstalling or updating already installed CLIs. It supports project-aware sessions by launching an agent in a selected project directory, synchronizes provider settings with CLI-native configuration files and environment variables, and performs a minimal model request to check credentials, endpoints, models, network access, and account status. The application reads local conversation histories in JSONL or SQLite formats, can resume or delete local sessions, displays installed MCP servers and Skills, and summarizes locally stored usage data. Agent Launcher distributes macOS, Windows, and Linux builds and is released under the MIT License. It is local-first rather than offline-only: launched agents send requests to the provider or relay selected by the active profile, while API keys are stored as plaintext in local configuration files and masked only in the interface.
argo9s is a K9s-inspired terminal user interface for monitoring Argo CD resources in real time. It watches Applications, ApplicationSets, and AppProjects through either the Kubernetes API, using kubeconfig or in-cluster access, or the Argo CD HTTP API with token authentication. The interface provides list and type views, project/name/sync filters, keyboard navigation, application resource graphs, pod logs, live YAML, and describe and edit actions. It is read-only by default, with optional synchronization, refresh, suspend/resume, autosync, and delete actions when a connection is configured as writable. Connections can be stored as named profiles in ~/.argo9s/config.yaml, and an offline --demo mode provides sample applications. The project is written in Go, distributed as a Homebrew package, Go installation, or release binary, and licensed under the MIT License.
ARTEMIS is an open-source Google tool for natural-language Android automation and testing. It drives real phones or emulators through end-to-end workflows, generates test plans, captures screenshots and logs, profiles performance, and produces structured reports. Its element-locating system combines accessibility hierarchies and OCR with visual-model and coordinate fallbacks for custom interfaces. ARTEMIS provides a CLI, web visual console, Python SDK, and native Model Context Protocol (MCP) server for connecting Android devices to AI coding assistants such as Antigravity, Claude Code, Codex, and Windsurf. Its Flash profile uses a reactive observe-and-act loop with compressed history, while the Pro profile uses planning, pre-execution checks, action bursts, incident recovery, checkpoint verification, and optional final reports. The project also includes an accessibility helper for reading screen layouts and can fall back to UIAutomator2. It is distributed under the Apache License 2.0; the repository reports 99%+ completion on Google Research's AndroidWorld benchmark.
Birdview is an open-source developer tool that makes AI coding agents map a project's architecture before editing. Its Birdview workflow defines modules, responsibilities, file ownership, evidence, relationships, and layout in an architecture JSON file; an agent then declares planned task scope, targets, files, lifecycle phases, and verification records against that map in an activity JSONL stream. Validators check the schemas and cross-record rules before a renderer produces a standalone interactive HTML view with architecture, changes, comparison, module evidence, and activity-history views. Birdview supports automatic or on-demand activation through project-level agent guidance and includes Node.js scripts, JSON Schemas, example maps, and tests. The generated output has no server or network dependency, but the v0.1 implementation is file-based: activity is agent-declared rather than automatically observed, and updates require regenerating the HTML. It is MIT-licensed and is not published to npm.
Cloudflare is an internet infrastructure and security platform that provides hosting and delivery infrastructure for applications, agents, websites, and data. Its network combines compute, connectivity, and security at edge locations, with Workers running code close to users and backend data; the company states that its network operates in more than 335 cities and reaches 95% of the world's Internet-connected population within 50 milliseconds. Cloudflare also provides AI crawler controls that let site owners inspect, allow, or block crawler activity, along with pay-per-crawl and a monetization gateway for pages, data sets, APIs, MCP tool calls, files, and search indexes. The described x402 flow uses an HTTP 402 response to request payment, after which an agent pays, retries with proof, and Cloudflare verifies the payment at the edge.
An open-source Codex configuration that uses Astra as the root orchestrator and independent reviewer, with Luna-based subagents for exploration, implementation, testing, and research. It provides separate Pro and Plus profiles, TOML role configurations, an Astra orchestrator skill, project-level AGENTS.md instructions, and shell and PowerShell installers that copy the selected configuration into a target repository. The profiles set model, reasoning-effort, approval, sandbox, and concurrent-thread settings, while named role files can override the defaults. The repository also includes guides for orchestration patterns, plan-specific setup, iteration, and token-usage reporting from Codex session logs. It is licensed under Apache License 2.0.
DeepJIT is a lightweight, header-only C++20 library for just-in-time compilation, caching, loading, and launching of device kernels on NVIDIA CUDA GPUs and Huawei Ascend NPUs. It provides a shared runtime interface for both backends while keeping kernel source and compiler options backend-specific. The runtime hashes kernel source, tracked includes, compiler information, effective options, and application-provided dependency signatures to reuse compiled artifacts through in-memory and on-disk caches. Cache directories can be shared across users, processes, and nodes on local or distributed filesystems, with completed entries published through atomic directory renames. Lazy initialization defers device and compiler discovery until first use. DeepJIT integrates with PyTorch CUDA and torch_npu streams and can expose a runtime through pybind11. The CUDA backend uses NVCC and the CUDA Driver API to produce, load, and launch CUBIN files; the Ascend backend uses Bisheng, ld.lld, and ACL to compile, link, load, and launch kernels. It supports per-kernel compiler and launch options, compilation diagnostics, and optional PTX, SASS, or Ascend assembly dumps. The host environment requires Linux, a C++20 compiler with std::format support, Python, pybind11, and the dependencies for the selected backend; it is intended to be embedded in another C++ or Python extension as a header-only dependency.
deepseek-recipe is a collection of Rust libraries and Python bindings for converting API requests in Messages, Chat Completions, and Responses formats into a shared conversation representation, encoding conversations into prompts or token IDs for DeepSeek V4 and V4.1 models, and converting model output back into protocol-specific responses. It supports streaming, text and image inputs, thinking and reasoning settings, function tools, tool calls, JSON object output, and stop sequences; its image component provides DeepSeek V4.1 preprocessing with OpenCV. The package parses model output and handles prompt encoding, while inference, tool execution, and HTTP transport are supplied externally. It includes separate components for core conversation types, prompt encoding, image processing, Python bindings, and example Axum and FastAPI servers with mock inference. Project code and public documentation are licensed under the MIT License.
edge0 is an open-source streaming Mixture-of-Experts inference framework for running sparse MoE models on Apple Silicon Macs. Its MLX backend memory-maps expert weights from SSD storage and loads them on demand, so peak memory depends on the active experts rather than the model's total parameter count. A trained prerouter predicts the next expert routes so storage reads can overlap the forward pass, while Recover-LoRA adapters distill from the full-precision model to compensate for quantization loss; the base weights remain read-only and the adapters stay separate. The framework ships two end-to-end model tiers with bundled 4-bit checkpoints, LoRA adapters, and prerouter heads. It provides a command-line interface and Python API, supports demo, chat, and server modes, and exposes an OpenAI-compatible `/v1/chat/completions` endpoint. The current backend supports macOS on Apple Silicon M1 through M4; CUDA support is identified as planned. The repository is licensed under Apache-2.0.
Eliza OS is a software compatibility target for converted robot demonstration records.
geiger is a read-only command-line scanner from Atomburst that inventories AI agents, harnesses, MCP servers, plugins, skills, hooks, browser extensions, desktop AI applications, and related IDE configurations on a machine. It reads known configuration files and directories, identifies each finding's origin and evidence path, and labels capabilities such as code execution, secret storage, broad filesystem access, broad web access, and network access. Credential-shaped values are reported by key name, file, and secret type without exposing value contents; the scanner performs no telemetry and writes only an explicitly requested report file. It can emit terminal, HTML, or versioned JSON reports, scan specified project or home directories, enforce a strict exit status for findings that can execute code or hold secrets, and compare JSON snapshots to identify newly appeared, removed, or escalated findings. It runs through npx with Node.js 18 or newer, has no runtime dependencies or account requirement, and is licensed under MIT. The project notes that it reads configuration rather than runtime behavior, covers known locations, and does not determine whether a package is malicious.
gpu-time is an experimental local neural parser for English date and time expressions. It converts text into ISO dates and time ranges, RFC 5545 recurrence rules, diagnostics, and source spans. The parser extracts token features without a word list, classifies tokens into time-related roles, detects expression boundaries, and uses TypeScript to build schedules from the predictions. It runs on CPU or WebGPU, with WebGPU processing parallel blocks and overlapping windows for long inputs; a calendar resolver applies the supplied reference instant, time zone, daylight-saving rules, and occurrence limit. The package provides parse, parseMany, and defineParser APIs, and does not send input to a server.
HKTex is an intrinsic texture representation and research software for triangular meshes. Instead of using a UV atlas, it represents appearance directly on the mesh with anisotropic heat kernels, the geodesic counterparts of Gaussians; kernel positions, optimisation, pruning, and densification operate on the surface. It can fit textures from textured meshes or multi-view observations and supports differentiable and physically based rendering through a Mitsuba-based pipeline. The repository includes KNN-accelerated HKTex experiments, adaptive density controllers, rendering and optimisation workflows, neural-texture and vertex-colour baselines, benchmarking utilities, and interactive notebooks. The documented setup targets Python 3.11 on Linux with an NVIDIA GPU and a CUDA 12.9-compatible driver. It is released under the MIT License.
Infinite World is an open-source system for building persistent worlds with multimodal AI. It combines multimodal understanding, world simulation, and interaction in a continuous loop: prompts, images, or other inputs are interpreted into world state, rules, and possible actions; generated scenes respond to that state; and people, agents, and events can update the world. The runtime records scenes, choices, state changes, history, and branches so worlds can preserve context across scenes and be replayed. The project supports text, voice, and image interaction, local previews, saved-branch replay, and cached generated video. Its current local setup uses a Node.js and pnpm workspace, with FFmpeg and whisper.cpp for voice transcription and live output. Interactive live streaming is in development, while additional inputs such as chat, audio, mouse, and keyboard are planned. The repository describes the project as early development, with APIs and stored data subject to change, and distributes it under the Apache License 2.0.
A browser-based Three.js study of foldable-screen transitions that recreates an iPhone Duo fold animation with projected screen UI, progressive blur, and custom image uploads. Only the cover half rotates while the rear-camera half remains fixed; screen content uses a front-view projection, and blur and darkening follow image coordinates during the fold. It provides Wallpaper, Launcher, and Custom modes, with orbit controls, zooming, a fold slider, and a play button. The application is static HTML, CSS, and JavaScript with Three.js bundled locally, and can be deployed to Vercel or another static host after preparing the Apple reference assets. The original application code is MIT-licensed; Apple models and imagery are excluded from that license and remain subject to Apple's terms.
Lid Plane is a macOS menu bar app that creates a MacBook lid-folding illusion using the laptop's hinge-angle sensor. It captures the built-in display with ScreenCaptureKit, reshapes the captured image with a Metal shader as the lid moves, and blends increasingly blurred copies using Metal Performance Shaders; the blur increases with lid movement and is stronger near the top of the display. The app can use a selected activation angle or a movement-based mode that anchors the effect after motion stops, while the underlying applications remain clickable and retain keyboard focus. It requires macOS 13 or newer, an Apple silicon MacBook, a supported lid-angle sensor, and Screen Recording permission. The effect applies only to the built-in display, pauses during closed-lid or unsupported-display states, and keeps captured frames in memory without saving or transmitting them. It is built with Swift, ScreenCaptureKit, IOKit, and Metal, and is distributed as a downloadable app for Apple silicon. The repository is licensed under GPLv3-or-later from version 0.3.2 onward; version 0.3.1 and earlier remain MIT-licensed.
Life Recorder is an open-source native iPhone recorder with a private Mac receiver. The iPhone captures roughly one-minute AAC chunks while the screen is locked or other apps are in use, then uploads them over an authenticated HTTPS connection. The Mac receiver runs whisper.cpp locally, removes common stage-direction markers and repetitive hallucinated noise, and appends the results to one continuous Markdown transcript with hourly capture markers; audio is deleted after durable receipt and successful transcription. It requires an iPhone development build installed with Xcode, a macOS system running Python, ffmpeg, whisper-cli, and a downloaded GGML Whisper model. Setup creates a bearer token and self-signed TLS certificate, uses certificate pinning, and stores the transcript in a private runtime directory. The default connection is over the local network; cellular use requires a private VPN. The project is licensed under MIT.
Mac Duo is a free macOS menu-bar app that recreates the iPhone Duo visual effect on compatible MacBooks. As the lid closes, it uses the built-in lid-angle sensor and ScreenCaptureKit to capture and render the live screen, while Metal applies perspective, blur, and dimming. The perspective is adjustable, and the app supports English and Simplified Chinese. It requires macOS 14 or later, a compatible built-in lid-angle sensor, and Screen Recording permission. Prebuilt downloads are provided for Apple Silicon and Intel Macs; the project can also be built with Xcode using Swift 6.0 or later. The effect applies only to the built-in display, stops when macOS sleeps, and passes clicks through to the applications underneath. The code is licensed under Apache License 2.0.
Mac Duo is a native Swift and Metal macOS app that makes desktop effects follow the position of a MacBook lid. It provides six transitions—Duo, Ghost, Roll, Shutter, Flex, and Iris—that animate the desktop around the hinge, including panel sliding, display curling, bowing, or iris-like closure. The app includes live preview, floating controls, menu-bar access, pause controls, and an option to clear the screen after the lid remains still. It runs locally without third-party runtime dependencies, accounts, analytics, or screen-content uploads; screen frames remain in memory when Screen Recording permission is enabled. It supports macOS 13 Ventura or newer on compatible Apple-silicon MacBooks, with an Intel preview for the 2019 16-inch MacBook Pro, and is distributed under the MIT license.
MiniWatts is an iPhone battery, charging, and thermal monitor built with Apple's private iOS APIs. It reads the phone's power-management sensors to show charger input, power reaching the battery cell, heat-related power, temperature sensors, and USB Power Delivery information, with Power, Thermal, Adapter, and History views. It also provides Home Screen and Lock Screen widgets, a charging Live Activity, and a floating Picture-in-Picture meter that can keep recording while another app is open or the screen is locked. MiniWatts is distributed as a sideloaded IPA because its private API usage prevents App Store distribution. It requires an iPhone running iOS 17 or later, includes no network code, and can be installed using tools such as Sideloadly, AltStore, SideStore, or Xcode. Sensor availability and meanings vary by iPhone model; wireless charging input current, battery health, and cycle count are unavailable, while discharge power and charging holds may be estimated or inferred. The project is licensed under Apache 2.0.
mjbatch is a Python library for running thousands of MuJoCo physics simulations in parallel on CPUs. It uses a C++ thread pool with the Python GIL released, provides live array access to simulation states and controls across a batch through bindings to MjData fields, and supports per-simulation model parameters with expansion of MjModel fields and recomputation of derived constants. The library includes examples for reinforcement learning, model-predictive control, system identification, and hardware co-design, including robot controllers, predictive-sampling control, parameter fitting, and joint optimization of robot properties and controls. It is distributed under the Apache-2.0 license.
motion-panels is an unstyled, framework-agnostic panel-layout library with a React adapter. Its core uses the DOM for resizing and Motion for animated folding, collapsing, pinning, and snapping, without relying on a layout library. Panels support pixel or percentage sizing, nesting, separator intersections, keyboard controls, right-to-left layouts, and live size updates; the package is distributed through pnpm and can also be added through shadcn. It is licensed under the MIT License.
Mousiki is a keyboard-driven terminal music player (TUI) for browsing and playing local music files or searching and streaming online tracks through yt-dlp. It provides word-by-word synchronized lyric highlighting, an FFT spectrum visualizer, waveform rendering, spinning-disk artwork, and queue controls for shuffling, repeating, and track management. Colors, animations, visualizer behavior, hotkeys, and local music paths are configured through a file in the user's home configuration directory. It has native support for Linux, macOS, and Android through Termux; Windows support is unverified. Building requires CMake, a C++17 compiler, FFmpeg, yt-dlp, and the Python syncedlyrics package, and the project is licensed under Apache License 2.0.
Mural is an open-source native iPhone and Android language-learning app built around live conversations with an AI voice companion. It lets learners choose a learning language and subtitle language, practise through voice or typed replies, receive corrections and optional meaning subtitles, look up words, and review vocabulary from later conversations. Challenge levels and provisional ability observations are adjusted from evidence in the learner’s replies, while each language has separate progress records and conversation themes. The iPhone client uses SwiftUI, SwiftData, Keychain storage and WebRTC; the Android client uses Jetpack Compose. Conversations, vocabulary and preferences remain on the device, and learning data can be exported or imported as JSON. The app connects directly to OpenAI using the learner’s own API key, requires an internet connection, and sends practice audio, selected text, learning context and requested searches to OpenAI. Hosted free conversations, minute purchases, and public App Store or TestFlight distribution are not active according to the repository. Mural supports Norwegian Bokmål, Spanish from Spain, international English, French from France, German, Italian, Brazilian Portuguese and Standard Mandarin with Simplified Chinese. It is released under the MIT License; third-party components retain their own licenses.
OpenAI is an AI model provider whose language models are available through an API and can serve as one provider in multi-model routing architectures. The videos describe its models being used for AI replies and extraction, Amazon-review analysis and customer-service email generation, legal workflows, complex software development and debugging, and experimental or auxiliary agent tasks.
OpenSend is a self-hosted transactional and marketing email platform built on Amazon SES for sending transactional mail, newsletters, and campaigns. It provides a dashboard, a TypeScript SDK distributed as the `opensend-js` npm package, and an MCP server over one public API; the video also describes send simulation and delivery-status logs. The project supports Docker or Cloudflare deployment, uses Google-only sign-in, keeps API-secret setup out of the browser, and recommends starting with simulated sending before provisioning infrastructure or sending real email.
Phyzical is a browser-based robot teleoperation data platform for embodied-AI training. Users control simulated robot arms through a web interface; inverse kinematics converts end-effector movements into joint actions, while sessions record demonstration trajectories containing joint positions, end-effector poses, object poses, and actions at 30–60 Hz. Episodes use a documented JSON Schema format and can be converted with the included Python tool into elizaOS-compatible SQLite trajectory databases for imitation learning and related robot-training pipelines. The project also documents quality checks and on-chain episode provenance using a data ID, content hash, and contributor, and includes human, scripted, and Fly-controller action sources. The repository is MIT-licensed and provides the converter and example episodes; it requires Python 3.10 or later and has no external dependencies for the quickstart converter.
ResolveHQ is a Cloudflare-native, self-hostable helpdesk for small support teams. It provides a shared inbox with tenant-isolated customers, tickets, assignment, status, priority, tags, full-text search, team roles, email threading, attachments, saved replies, internal notes, reports, rule-based automations, and a public help center generated from knowledge-base articles. It runs as a single Cloudflare Worker that serves the REST API and React application through Hono. Cloudflare D1 stores tickets and customers, R2 stores attachments, Queues handle inbound and outbound mail jobs, and Email Routing delivers incoming mail. Outbound mail uses Resend by default, with an option on Workers Paid to use Cloudflare Email Sending. The system includes retries and recovery for stalled mail jobs, RFC 5322 thread handling, CSV report exports, and customer data export or erasure workflows. AI assistance for summaries, reply drafts, classification, and translation is optional and disabled until enabled per workspace. It can use Workers AI or OpenAI. The project can run on Cloudflare's Free plan for small deployments, subject to current platform limits, and requires a Cloudflare account and a domain on Cloudflare for Email Routing.
Roam Control is a source-available SwiftUI iPhone app for simulating reported device locations during location-based development, quality assurance, and personal testing on an iPhone the user owns and controls. It provides an Apple Maps interface for searching places with MapKit, entering coordinates, selecting map points, starting fixed locations, and previewing walking routes; active walks can be paused, resumed, reversed, or redirected, and locations can be saved as favourites or restored to the device's real position. The app pairs natively with the iPhone through LocalDevVPN. It stores an RPPairing record in the device-only Keychain, discovers the remote-pairing service, verifies the device identity, and opens an encrypted developer session to set or clear the simulated location. It is distributed as an IPA through GitHub Releases rather than the App Store or TestFlight and must be signed with SideStore or Xcode. The public beta requires iOS 27 or newer and Developer Mode. The current source is licensed under PolyForm Noncommercial License 1.0.0; the earlier 0.9.0 Beta 1 release remains under the MIT License.
Rune is a GPU-rendered, keyboard-driven IDE for power users, developed by Unstable Build. It combines code editing, terminals, command-line tools, language intelligence, debugging, file and workspace management, and AI agents in a composable multi-workspace environment with a character-grid interface. Rune Agent is provided separately as an extension in the same repository rather than being part of the core editor. Rune is an application written in Go and built with standard Go tooling; its GPU renderer uses cgo and requires platform-specific graphics and development libraries. The application is licensed under the GNU General Public License version 3 or later, while the separate rune-go-sdk extension module is licensed under Apache-2.0.
SoL-Pi is an open-source standalone extension for the Pi coding agent, developed and maintained by NVIDIA. It packages four opt-in efficiency mechanisms: Action Fusion runs a follow-up validation command in the same edit or write call; ObservationPack replaces repeated large tool results with stable handles that support exact paged recall; the Evidence-Preserving Reducer condenses diagnostic logs into receipts only when retained quotations match archived source material; and Online Context Compact marks completed plan steps for Pi's native compaction when economic and context-window checks allow it. The extension operates through Pi's public APIs without patching or vendoring Pi, leaves the mechanisms disabled by default, and preserves original observations locally. It requires Node.js 22.19 or newer and is released under the MIT License.
SQLite is a self-contained, serverless, zero-configuration, transactional SQL database engine implemented as a C library. It was created and is primarily maintained by D. Richard Hipp and the SQLite development team and is widely used as an embedded database in applications, devices, and browsers.
trynix is a browser-based environment for running packages from the historical Nixpkgs archive. It resolves a package and version through the nixpkgs-multiverse index, fetches its signed runtime closure from cache.nixos.org, and runs it in an x86_64 Linux virtual machine powered by QEMU compiled to WebAssembly. The browser decompresses the package archives into an in-memory filesystem shared with the guest over virtio-9p; the guest resumes from a migration snapshot rather than booting from scratch, then exposes the selected packages through shell commands. Environment state is encoded in the URL, including package selections, store paths, and additional binary caches. A GitHub Action can comment on pull requests with links that launch the store paths produced by a build, allowing reviewers to run branch builds in a browser. The project is MIT-licensed and can also be run locally with Nix.
Viserys is a self-contained pack of Markdown engineering workflows, reviewer personas, validation scripts, commands, and lifecycle hooks for AI coding agents. Its workflow routes development through DEFINE, PLAN, BUILD, VERIFY, REVIEW, and SHIP stages, with 28 skills covering requirements clarification, specification and task breakdown, implementation, test-driven development, debugging, browser testing, review, security, performance, migrations, CI/CD, documentation, observability, and release preparation. The pack also includes four reviewer personas, shared checklists, evaluation cases and fixtures, and validators for skills, commands, artifact paths, reference links, and versions. The repository includes an OpenCode agent definition that registers the skills and routes requests through the lifecycle workflow. The Markdown skills can also be supplied to skill-aware agents such as Claude Code, Cursor, and other agents, or read directly as standalone process documents.
whisper.cpp is a C/C++ implementation of OpenAI's Whisper automatic speech-recognition model for speech-to-text inference. It is designed for offline, on-device use and supports CPU-only inference, GPU and NPU acceleration, integer quantization, mixed F16/F32 precision, voice activity detection, and a C-style API. The implementation targets macOS, iOS, Android, Linux, Windows, WebAssembly, Raspberry Pi, Docker, and other platforms, with optimizations including ARM NEON, Apple's Accelerate and Metal frameworks, x86 AVX, POWER VSX, Vulkan, CUDA-related NVIDIA support, ROCm, OpenVINO, and other backends. The repository includes the whisper-cli example, which transcribes 16-bit WAV audio using Whisper models converted to ggml format; its zero-dependency implementation is intended for integration into applications such as on-device captioning tools.
YOINK is a Python terminal note-taking and forecasting tool that stores notes as plain Markdown files with Obsidian-compatible frontmatter and wikilinks. It adds claims, deadlines, confidence levels, machine-testable conditions, and data sources to notes; when a claim is due, `yoink settle` reads the source, applies the test, and writes the outcome and reading back into the note. Supported sources include local files, CSV values, public EVM-chain JSON-RPC readings, and manually recorded URLs. The terminal interface provides capture, search, link navigation, tags, graph views, claim settlement, and vault health commands. It recomputes the vault from disk without a database or index, and `yoink score` reports note activity, settled-claim accuracy, confidence gaps, and Brier scores. An append-only, hash-chained JSONL ledger records captures, reads, claims, and settlements; the project describes it as tamper-evident rather than tamper-proof. The repository states that it targets Python 3.10 or later, uses only the standard library, has no external dependencies, and is licensed under MIT.
Zetta is an open-source closed-loop embodied harness for self-evolving physical intelligence. It keeps a base policy frozen while evolving code-based runtime critics and recovery skills when rollouts fail. Its evolution protocol clusters failed trajectories, uses a diagnostic stage to identify an observable causal failure mechanism, generates Critic-Recovery candidates under fixed tool schemas, and checks them through shadow replay, paired same-seed evaluation, and held-out seeds. Candidates are rejected or promoted based on these gates; the critic may propose recovery actions, a higher-level role approves them, and a bounded recovery actor executes only approved programs. The repository provides campaign management, rollout workers and environment backends for systems including LIBERO, RoboCasa, RoboTwin, and Genie Sim, along with manifests, tool catalogs, audit records, deployment helpers, and contract tests. It supports Python 3.10–3.12, while simulator assets, model weights, credentials, and host runtime files remain outside the repository.
Run an email platform on top of Amazon SES so users can manage transactional mail, newsletters, and campaigns without building the infrastructure from scratch.
Provide a help desk on Cloudflare infrastructure with shared inboxes, ticket workflows, search, role-based workspaces, attachments, and a public help center.
A native iPhone app lets learners practice languages through live conversations with an animated AI character, receive corrections, look up words, and revisit vocabulary in later conversations.
Searchable transcript of GitHub Trending Today #49: SoL-Pi, viserys-agent, Edge0, Mac-Duo, artemis, rune, YOINK, opensend — Github Awesome (15:37). 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 today number 49. 35 trending open-source projects on GitHub right now. Let's get into it. So, PPI is NVIDIA's extension for cutting repeated work in the PI coding agent. It can combine a file edit with its validation command, replace bulky repeated outputs with references the agent can retrieve, and shorten diagnostic logs while checking quotations against the originals.
00:28 Everything is optin and pi itself stays unmodified. The optional log reducer may send diagnostic content to a configured model. This is a codeex config that splits the work across a chain of command instead of one model doing everything. GPT6 Astra sits on top as orchestrator deciding what needs doing while GPT 5.6 Luna sub aents explore the code, implement, test, research or review it.
00:56 Readon roles like explorer and reviewer can't touch your files the way worker and tester can. Vizeris gives AI coding agents a written engineering process from clarifying requirements to reviewing and shipping changes. Its markdown workflows define steps and exit criteria for tasks like test-driven development, debugging, and migrations. Separate reviewer personas examine code quality, security, and performance, while shared checklists load only when needed.
01:26 Edge Zero runs Mixture of Experts models on a Mac by streaming expert weights off the SSD instead of loading the whole model into RAM, which is how the 35 billion parameter version gets by on just 2.9 GB of peak memory. A pre-outer predicts which experts get used before it needs them. The maintainers say that's worth up to 59% more decode throughput.
01:48 and a distilled Laura adapter claws back most of the quality lost to 4-bit quantization. Mac Duo takes that iPhone Duo effect where the screen tilts and blurs as you close the lid and ports it to MacBooks that actually have a lid angle sensor. It captures your live screen with Screen Capture Kit and runs the perspective, blur, and dimming through Metal on the GPU with adjustable angle settings sitting in the menu bar.
02:16 It only works on the built-in display and needs screen recording permission granted first. Zeta lets a robot fix itself while it works. Instead of retraining the whole policy, it freezes the base model and evolves small code-based critics and recovery skills on top of it whenever something goes wrong. There's an actual evolution protocol that diagnoses the failure, generates candidate fixes, and only keeps ones that pass staged validation gates with held out test seeds.
02:47 So, it's not just gaming its own benchmark. Artemis drives a real Android phone or emulator off plain language instructions instead of brittle UI selectors with a planner, operator, and checker doing the actual work. It runs in two speeds. flash at three to five seconds a step, pro at 15 to 40 when it needs to actually plan. It self-heals by clearing popups that would otherwise derail a test.
03:14 And Google says it's held up through 10 plus hours of continuous exploratory testing. Rune is a GPU rendered IDE built entirely around the keyboard. No mouse first design, just a character grid interface running editor, file tree, and workspace management through one composable environment. AI coding lives in a separate extension called Rune Agent instead of being baked into the core, so you can still do it all by hand if you want.
03:44 Yoink is a terminal note app that checks how your predictions turned out. It stores everything as markdown compatible with Obsidian, but lets you attach a deadline, confidence level, and testable condition to a note. When you run its settle command, it checks the source and records the outcome. You can then compare confidence with accuracy while also spotting notes you saved and never revisited.
04:10 OpenSend is a self-hosted email platform that sits on top of Amazon SCEs, handling transactional mail, newsletters, and campaigns without making you build any of that from scratch. O is Google only, so there's no password system or team admin panel to maintain. And the dashboard SDK and MCP server all hit the same public API. It'll let you simulate a send before anything real goes out and logs every delivery with its status so you can filter through what failed.
04:40 Guyer runs one command, npxgger scan, and inventories every AI agent, MCP server, plugin, and extension actually installed on your machine. Then tags each one with what it can touch, execute, hold secrets, broad file system, network. It never shows you actual credential values, just their shape, and it's entirely readon, apart from an optional JSON export.
05:00 You can snapshot a baseline and diff later scans against it to catch drift. Deepjit handles the plumbing around compiling and running device kernels at runtime. Deepseek's header only C++ library gives Nvidia CUDA GPUs and Huawei Ascend MPUs a shared compile, cache, and launch workflow. It can reuse compiled binaries across processes and machines and integrates with PyTorch's current device stream.
05:28 It unifies the infrastructure, but your kernel code and compiler options still depend on the hardware back end. 3D Viz Pro Max is an agent skill that turns a plain description into an actual 3JS scene. And the key part is it doesn't just generate code and guess. It captures the rendered frame with playright and checks its own output before calling it done.
05:49 It ships with 223 recipes across 24 subject areas from anatomy to physics and 22 blueprint kits backed by real GPU captured frames verified with Shaw 256 hashes. Try lets you try historical Nyx packages in a Linux virtual machine running inside your browser tab. Pick a package and version and it fetches the pre-built package and its dependencies from Nyx's cache, then gives you a shell.
06:17 The environment lives in the URL so you can share it as a link. There's also a GitHub action that turns pull request builds into browser previews for reviewers. MJBatch lets Python run batches of Muoko physics simulations. across CPU cores. A C++ thread pool handles execution outside Python's global interpreter lock, while live arrays let you read states and update controls across the whole batch.
06:45 You can also vary parameters like friction between simulations. The repository includes examples for robot learning, predictive control, and fitting physical parameters. This recreates the fold animation on the iPhone Duo entirely in the browser with 3.js. Only the cover half rotates while the rear camera half stays put and it applies a dynamic blur up to 72 pixels that darkens and tracks the fold as it happens.
07:12 You get three display modes, wallpaper, launcher or your own uploaded image, plus drag to orbit and scroll to zoom controls. This project tested whether GPT6 Astra can actually control a robot arm, not just reason about code. Running it directly on 10 Robo-Dogjo manipulation tasks, five trials each, it scored 37.81 on average with a 26% success rate.
07:35 But when they used it as a correction layer on top of an existing PI0.5 policy, stepping in on just 14.4% of control steps, the combined score jumped to 62.6 with a bird view makes an AI agent show its work before it touches your codebase. It builds an evidence- linked map of your architecture, stable module IDs, who owns which files, and then has the agent declare what it plans to change against that map before any edit happens.
08:05 The output is a single self-contained HTML file, no server, with sidebyside architecture, changes, and comparison views, plus light and dark themes. Rome control simulates GPS location on your own iPhone for testing location based features. The same thing Xcode Simulator does, but on real hardware with a proper Apple Maps interface. You can search a place, tap a spot on the map, or type coordinates directly, and it'll preview and step through an actual Apple Maps walking route with pause and reverse.
08:40 Argo 9 brings a K9 style terminal interface to Argo CD so you can watch application health and sync status without opening the browser. It connects through Kubernetes or Argo CD's API with filters and a resource tree for drilling into deployments and pods. You can inspect logs and live YAML from the keyboard. It starts readonly, enabling writes, adds actions like syncing applications and toggling automatic synchronization.
09:09 Life recorder turns an iPhone into a continuous voice recorder that keeps running even with the screen locked or another app open, capturing audio in roughly 1 minute AAC chunks. It ships those over a local HTTPS connection, LAN by default, PrivateVPN if you need it over cellular to a Mac that transcribes everything locally with whisper.cpp. CPP no cloud service involved.
09:33 Deepseek recipe translates different AI API formats into prompts. Deepseek models understand then converts their outputs back into matching responses. Its Rust libraries and Python bindings cover messages, chat completions and responses, including streaming and tool call parsing that gives developers shared conversion code instead of separate plumbing for each format.
09:57 GPU time takes something like satsson 1 p.m. to 8 p.m. and turns it into an actual ISO date range with an RFC545 recurrence rule behind it entirely on your machine. Nothing gets sent to a server. It runs on plain CPU by default, but switches over to web GPU once a batch gets big enough past 32 inputs or 512 tokens and it's time zone and daylight saving aware.
10:23 Cargo Pretty replaces Cargo's usual scrolling build log with a live terminal dashboard tracking up to six crates or tests at once with their own timers. And the progress bar is actually built from Cargo's real unit graph instead of a guess. Warnings collapse into a navigable panel and failed tests get diffled output plus a keyboard shortcut to retry them without rerunning the whole suite.
10:48 Infinite World generates a persistent environment that actually remembers what happened, and you steer it with text, voice, or images instead of just type prompts. Every scene it generates reflects the current state of the world, and branches you take stay branches. You can preview locally and replay back through past paths with cached video. Physical connects browser-based robot demonstrations to training pipelines.
11:15 The platform lets people move simulated robot arms, recording joint positions, actions, and object poses as demonstration data. This repository publishes the episode format, a sample task, and a dependency-free Python converter that writes Eliza OS compatible SQLite records. Lid plane creates that lid fold illusion on a MacBook driven by the actual hinge sensor instead of guesswork, pulling the lid angle about 30 times a second and blurring the desktop progressively as you move it.
11:46 Stop moving and an auto anchor settles the effect after a pause. You can tune anywhere from 0.15 to 2 seconds. Your apps stay fully clickable and keep keyboard motion panels gives resizable interfaces. animated folding, collapsing, and snapping powered by motion. Its resizing engine works directly with the DOM with a React adapter on top, so the core isn't tied to React.
12:12 Panels can use pixel or percentage sizes, nest inside each other, and support keyboard controls and right to left layouts. The components come unstyled, or you can add a styled separator through its Shad CN registry. This Mac Duo app syncs desktop effects to your MacBook's actual lid position. And you get a choice of named transitions, Duo, roll, shutter, flex, iris that all play out as you close it.
12:38 It auto clears after a stationary delay you can set between 1 and 5 seconds and esque or control option command F pauses the tracking entirely. It's pure swift and metal. No third party runtime dependencies, no analytics. Everything stays local. Musiki is a terminal music player built for people who'd rather not touch a mouse. It plays your local library or pulls tracks in through YTDLP for online search and streaming.
13:03 And lyrics sync in real time with wordbyword highlighting as the song plays. There's a real FFT spectrum analyzer, waveform rendering, and album art right there in the terminal, plus shuffle, repeat, and Q management. Mural is a native iPhone app for practicing languages through live AI conversations. You speak to an animated orb, get corrections, and use optional subtitles or word lookup when you're stuck.
13:31 It revisits vocabulary in later conversations and keeps separate progress for each language. Learning records stay on your phone, but practice sends data to OpenAI. Agent Launcher brings existing coding agent CLIs into one desktop workspace. It detects what's already there instead of forcing a reinstall and one clicks whatever's missing. It manages multiple provider profiles with actual connectivity checks, remembers which project you were working in per session, and reads your local conversation history straight out
14:04 of JSONL or SQLite, so you can resume or delete past sessions. Mini Watts taps into the same private power management sensors Apple uses internally to show you exactly what's happening during a charge. Real wattage coming out of the charger, how much of that's actually reaching the battery cell and readings off all 30 temperature sensors in the phone.
14:27 It even shows the USB PD handshake as it negotiates. Nothing leaves the device. There's no network connection at all. HK text puts textures directly on a 3D mesh instead of flattening its surface into a UV map. It represents appearance with heat kernels that follow the surface, avoiding seams introduced by UV layouts. You can fit textures from an existing textured mesh or multiple images, then use them in physically based rendering.
14:56 The repository includes optimization and comparison tools with a setup targeting Linux and an NVIDIA GPU. Resolve HQ is a help desk you run entirely on Cloudflare's own infrastructure instead of paying for Zenesk or Intercom. It's a shared inbox with ticket assignment, priority and status tracking, full text search, and proper multi-tenant roles, owner, admin, agent across workspaces.
15:24 Email threading follows RFC 5322. Attachments go through Cloudflare R2, and there's a public help center generated straight from your knowledgebased articles.