An ESP32-S3 firmware and MCP server for the M5Stack StopWatch that lets an external phone or computer agent control an iPhone. The client observes the iPhone screen and decides the next action; the StopWatch executes MCP commands such as taps, double-clicks, long presses, drags, scrolling, keyboard presses, ASCII text entry, and stopping or revoking control. The firmware runs no model and does not provide screenshots. The client connects to the StopWatch MCP endpoint over USB networking or Wi-Fi, while the StopWatch sends input to the iPhone over USB or experimental Bluetooth Low Energy. It supports portrait-normalized coordinates, one-pointer input, US-layout ASCII typing, and iPhones with AssistiveTouch enabled. The repository includes a setup skill for identifying the board, flashing firmware, pairing MCP, and adapting the software to other ESP32 boards. The project is integrated with the client-side Anan agent but also supports other HTTP MCP clients. Wi-Fi requires a 2.4 GHz local network; BLE input is experimental, and Chinese or other Unicode text input is not supported. Existing code is GPL-3.0-only, with an AGPL-3.0-or-later control-avatar component and other third-party notices retained in the repository.
BINNY is an open-source agent harness for reverse engineering. It connects a selected hosted or local language model to decompiled functions, strings, imports, data, Ghidra analysis, and other analysis tools, then runs a triage, investigation, and report-drafting loop against a binary. It can classify samples using the STIX 2.1 malware vocabulary, apply malware or vulnerability-research playbooks, record findings, and save checkpoints after each step for later resumption. BINNY supports autonomous, supervised, and collaborative operating modes, as well as an interactive question-and-answer mode for checking existing report claims or inspecting individual functions. Its sandbox executes analysis scripts in a Docker container with no network, read-only mounts, dropped capabilities, and resource limits. Local Ghidra decompilation runs on the user's machine, and projects can include multiple samples. The tool supports hosted providers and local models through Ollama, LM Studio, llama.cpp, or vLLM, provided the model supports tool calling. It is distributed as cross-platform binaries and can also be built from source. BINNY is licensed under the GNU Affero General Public License version 3 or later.
c64-hw-sim is a Commodore 64 motherboard simulator for the ASSY 250407 board. It models the board's 216 parts and 249 electrical nets at pin level, including the 6510, VIC-II, CIAs, SID, RAM, ROMs, logic chips, PLA equations, and discrete components. Users can lift or short individual pins, remove chips, open resistors or other connections, and observe the resulting monitor output, keyboard behavior, logic states, and oscilloscope signals. Its simulated netlist and chip behavior compute fault symptoms from the circuit rather than selecting predefined results; the simulator can also load .prg programs, run headlessly, apply timed faults, inspect bus contention, and record graphical output. It builds with make for macOS and Linux with X11 and OpenGL. The project is intended for studying C64 hardware and investigating repair hypotheses, not for playing games, and runs below real-time because every chip pin is simulated. Its models do not include the video output, SID filter, power supply, or dot-clock PLL, and the analog floating-line and power-fault values are approximations.
Casa Serra is a scroll-driven interactive house visualization and work-in-progress web demo. It opens as a flat architectural floor plan, which is the actual Three.js 3D model viewed from above and cut at about 1.2 metres; scrolling raises the cut to form the walls and moves the camera through the rooms. Each room remains a line drawing until the camera reaches it, when materials appear. The scene uses geometry generated in code, GSAP ScrollTrigger and Lenis, plain files without a build step, and CC0 texture scans from Poly Haven. On devices that cannot run the real-time scene, it provides a still image and a video clip.
Co-op Anything is an open-source MCP server and playbook for directing AI coding agents to add cooperative multiplayer to single-player PC games. The MCP server searches a directory of more than 230 co-op mods, detects a game's engine from its installation folder, provides research and reverse-engineering guidance, and validates contributions. An agent either revives an existing mod by auditing and fixing its code or builds co-op from scratch by hooking into the game, synchronizing players, enemies, worlds, story, level changes, and saves. Builds are tested with bots driving real game clients, then packaged for installation and Steam-relay play without port forwarding. The repository includes agent skills, engine-specific guidance, networking and split-screen playbooks, automated QA procedures, game-specific projects, and contribution rules. It is licensed under the MIT License.
earshot is an open-source audio analysis and sound-generation tool for coding agents. It runs locally on CPU through a CLI, Python API, or MCP server and turns audio into evidence an agent can interpret, including calibrated sound descriptions, vocabulary matches, candidate rankings, moments found inside long recordings, real-versus-code-synthesized assessments, listening timelines, measurements, annotated images, and video sound-sync critiques. Its forge component generates effects from physical-synthesis recipes such as modal resonators, filtered noise, room and microphone models, and layered DSP, then searches and refines variants using audio descriptions, reference recordings, spectrograms, model-free measurements, and artifact checks. The fit component analyzes a recording by synthesis and writes a parameterized recipe that can render variations or morphs. A bundled sound-effect bank can render coded effects by identifier, while kit commands assemble deterministic film sound kits with measured onset, peak, and end points. The project includes an MCP server for Claude Code and other agents, static HTML galleries for generated sounds, and optional audio-LLM critique with explicit consent and a user-supplied key. Its code is MIT-licensed; the two open models it uses have separate licenses. Installation requires Python 3.10+ and ffmpeg, and the first model use downloads approximately 1 GB of model files for local caching.
Enclose is a desktop workbench for CAD projects built from code. It displays a model in 3D; users select a part and describe a change, after which Claude edits the project and rebuilds the model. Circuit boards are designed with tscircuit and rendered in the app using its libraries. The early release runs on macOS and Linux and opens STEP, STL, and OBJ files. It requires Claude Code to be installed and signed in, and Node.js for circuit-board work. Distribution includes macOS disk images for Apple Silicon and Intel and Linux AppImage, DEB, pacman, and RPM packages. The project is licensed under AGPL-3.0.
esp32sim is an open-source ESP32 emulator written in Rust. It uses instruction-level emulation to run unmodified firmware through the real mask ROM, second-stage bootloader, FreeRTOS, and application image for ESP32, ESP32-S3, ESP32-C3, and ESP32-C6 systems, with modeled peripherals such as GPIO, UART, timers, displays, cameras, audio, storage, and wireless interfaces. The desktop command-line version supports virtual Wi-Fi with DHCP, DNS, NTP, WPA2, and NAT to the host network, allowing emulated firmware to reach real services without hardware or firmware changes. The same emulator can be compiled to WebAssembly and run in a browser, although browser execution has no network NAT. It includes board models, debugging and tracing commands, regression tests, and differential comparisons against physical ESP32 hardware. The project is distributed under the MIT License.
Filee is a free, open-source desktop file-conversion tool developed and published by Myeonggu Kim (KnifeLemon). It opens a format menu at the cursor when a key is held and files are dragged, showing conversion targets appropriate to the selected files; conversions run locally and save results beside the originals. It supports image, document, spreadsheet, presentation, PDF, e-book, video, audio, subtitle, archive, CAD, font, data, and 3D-model formats, with batch and parallel conversion, automatically discovered multi-step routes, watch folders, a right-click menu, and a command-line interface. Optional engines such as Ghostscript, LibreOffice, Pandoc, calibre, and FFmpeg are downloaded only when needed. Filee runs on Windows, macOS, and Linux, and is licensed under the MIT License.
flipflop is an open-source file- and text-sharing tool that establishes a trusted tunnel between two or more devices, allowing subsequent transfers without repeated pairing codes. It uses Iroh to connect devices over Bluetooth, a local network, or the internet, including automatic NAT traversal, and can share files, links, and plain text. Builds are provided for Linux, macOS, Windows, and Android; an iOS build is not available.
Frametell is a local-first desktop app and CLI from Sitor AI that turns an idea or source material— including notes, documents, PDFs, slides, images, and Markdown—into an explainer video. Its AI director converts the input into pages and story beats for approval or editing, writes the narration, generates a voiceover, draws each page using configurable styles and camera moves, checks pages with lint and screenshots, and performs a final video check before rendering a subtitled 1080p MP4. Rendering runs locally through Chrome, Edge, or Chromium, while model and voice services use the user's API keys; supported providers include Anthropic, OpenAI, Google, OpenRouter, and compatible endpoints, with ElevenLabs or Volcengine TTS for voices. It is available as an Apple Silicon macOS app or from source, and is licensed under AGPL-3.0.
gitai is an open-source, MIT-licensed version-control system designed for AI coding agents. It moves repository operations onto a server: clients fetch individual objects and sparse trees, while the server handles history queries, blame, path history, ancestry, code search, and derived-data caching. Changes land by rebasing onto the current branch tip and using compare-and-swap rather than a merge queue; concurrent lands can be batched, and accepted reference updates are recorded in an ordered changeset log. Conflicts can be resolved through a versioned, vendor-neutral protocol using an LLM provider, a deterministic tool, or no resolver. The server uses normal Git repositories underneath, with SQLite or PostgreSQL for changesets and references and file or cloud object storage for content-addressed packs. It includes a graft CLI and Go SDK, optional cache daemons, agent tokens and branch policies, semantic review rules, post-landing verification and reversion, and agents that monitor changesets. The repository describes it as an independent implementation inspired by Anthropic's internally developed Scion design, not affiliated with or endorsed by Anthropic.
Harness² is an open-source framework for recursively optimizing the system around an AI model while keeping the model weights fixed. At the task level, it uses a contrastive Propose–Probe–Compose process: it proposes edits to an execution harness, probes their effects on the same task, and composes a refined harness from behavioral differences between executions without ground-truth feedback. A second recursion level maintains an improvement harness containing adopted harnesses and evidence-supported lessons for future tasks. Its editable components include system prompts, guardrails, memory, tool guidance, subagents, skills, scripts, and plugins. The repository supports parallel and sequential recursion, OpenCode and Codex executors, and benchmarks including JobBench, LAB, and WorkBuddy-Bench. It runs locally with Linux, Python 3.12, and uv, and is licensed under Apache 2.0; the repository states that it is not an officially supported Google product.
Hy Memory is a cognitive memory engine and Python SDK from Tencent Hunyuan for LLM agents. It routes data into three task-specific memory scenes: hierarchical long-term memory for chat, two-layer session-fact and reusable-experience memory for productivity work, and the same two layers plus a version-evolution diff chain for coding tasks. The chat pipeline organizes conversations into layers including raw dialogue, atomic facts, summaries, schemas, and prospective intentions. Working and coding pipelines extract what a session did separately from lessons that may be reused; the coding pipeline also attaches the actual diffs and user-described revisions for the code touched by each fact. The SDK accepts plain text or OpenAI-format messages, supports automatic scene routing, source isolation, CRUD operations, and configurable vector, graph, and metadata stores. The repository also includes MemHub, a local-first desktop and CLI product that combines conversation history and coding traces from authorized AI applications into a managed local runtime. Hy Memory is distributed under the Apache-2.0 license and can use OpenAI-compatible LLM and embedding endpoints, with ChromaDB as a default vector store and optional Qdrant, FAISS, Kuzu, Neo4j, and MySQL backends.
leanrea is an 8 KB skill file for coding agents that guides reverse engineering of compiled applications using ordinary command-line tools instead of an MCP server. It routes targets such as Electron and ASAR files, Mach-O binaries, native disassembly, Go binaries, Android APKs, WebAssembly, SQLite databases, network captures, firmware, and web applications to appropriate first commands; its procedures use strings, imports, extracted files, and call sequences to narrow the investigation. The skill also instructs agents to work on copies, batch shell probes, search hidden extracted directories, and label claims as seen, inferred, or unknown. It can be installed for supported agents with the skills CLI or manually by copying its plain-text SKILL.md file, and is MIT licensed. The repository reports a benchmark in which leanrea matched REA's accuracy with lower wall time, cost, turns, and token usage, while noting that the comparison used one model, one machine, and three repetitions per cell.
Lucian UX/UI Prompt is a library and live gallery of 400 interactive UI elements: 200 interface parts such as buttons, inputs, charts, and workflow screens, plus 200 website sections and animations. Each element supports light and dark themes, pointer, touch, and keyboard input, reduced motion, and isolated HTML, CSS, and JavaScript without dependencies. Users can try an element in the gallery, copy a prompt containing its working code and porting instructions for an AI coding tool, copy the raw code, or download a standalone HTML file. The repository also provides a skill that lets an AI coding agent search the library and port elements into a project, and it can be built locally with Node.js. The code is licensed under MIT.
Mayoi is a small code editor written in Go by sorrycc. It renders a native user interface directly on the GPU, without a webview or Electron, and provides tabs, columns, syntax highlighting, file navigation, search, Markdown and SVG previews, Vim modes, and a command palette. Its Git interface supports staging, commits, branches, blame, commit graphs, and gutter-based staging or reverting of individual changes. A built-in terminal uses Ghostty's libghostty-vt emulator. Remote SSH opens folders on another machine using the system SSH client; the remote host runs the Explorer, search, Git, and terminals while ports are forwarded locally. Mayoi supports macOS 13 or later and Linux with GTK 3, also builds for Windows, and is licensed under the MIT License.
mcp-for-blender-pro is a Blender add-on and MCP server that provides an AI assistant for inspecting scenes, creating and editing objects, assigning materials, searching and importing assets, running Blender API operations, and exporting scenes. It combines visual chat, scene tools, viewport snapshots, approvals, and a task board in a companion workspace, and can use OpenAI, Anthropic Claude, Google Gemini, OpenRouter, OpenCode, Ollama, or compatible custom endpoints. Single-assistant workflows handle direct requests, while team workflows assign sequential work to managers, workers, and reviewers; the built-in MCP server also supports authenticated Streamable HTTP connections from compatible clients. The add-on supports operation approval, cancellation after the current step, named undo steps, and Python tool validation, although the validation is not a security sandbox. The release bundle targets Windows x64 with Blender 4.2 or later and is released under the MIT license.
Morrow is a free, open-source, local-first desktop workspace for notes, databases, and ideas. Each page is stored as a plain Markdown file in a folder selected by the user, with support for wiki links, callouts, tags, properties, mathematics, canvases, page history, search, daily notes, templates, flashcards, and embedded pages. Its databases provide table, board, calendar, list, and gallery views with filters, sorting, grouping, relations, and formula properties; canvases are stored as open JSON Canvas files. Morrow can open Obsidian vaults without modifying untouched files and import Notion exports, including databases. It works without an account, cloud service, telemetry, or AI features, and includes encrypted peer-to-peer sync between computers on the same network, encrypted pages, backups, PDF and Markdown export, and static-site publishing. The repository reports Windows, macOS, and Linux builds, with the macOS and Linux builds not yet tested on real machines and builds not yet code-signed. Morrow is licensed under Apache License 2.0.
Motion Prompt is a library and agent skill for identifying and requesting visual techniques in motion graphics, camera movement, UI components, landing-page and web effects, and photographic styles. Its gallery pairs each named technique with a moving sample, a one-line meaning, and an example prompt; the repository currently organizes these into five packs of 24 names each, available in English and Japanese. The skill can open the gallery in a browser, let users select samples, and send the selected numbers and notes to Claude Code, Codex, or another supported coding agent. The agent can then retrieve the selected technique's specification and reference code and adapt it to a project using Canvas, Remotion, CSS, After Effects, or prompts for a video-generation model. Its command-line script supports picking, searching, listing, retrieving specifications and code, generating prompts, and viewing the shared code for a pack. The samples are implemented as deterministic Canvas 2D functions of time, with pack-level kits containing shared colors, sizes, cycle lengths, and helpers. The skill uses a standard-library Node.js script and fetches catalog, specification, prompt, and kit files from its published data sources; downloaded data is cached locally, and nothing from the user's project is uploaded. The repository is distributed under the MIT license, with commercial use allowed.
net-merge is a free, open-source networking tool for Windows, macOS, and Linux that combines Wi-Fi, Ethernet, phone hotspots, and USB-tethered connections. It creates a virtual network adapter using TUN and Wintun, assigning each new TCP or UDP connection to one network link and balancing more connections toward faster links; links that lose internet access receive no new connections until they recover. This can increase aggregate speeds for applications that open multiple connections, such as browsers, game and software downloads, cloud uploads, download managers, and torrents, but a single connection remains limited to one link. It requires administrator or network capabilities, supports IPv4 and IPv6, and is distributed under the MIT license.
opendesk is an open-source helpdesk that combines support email, live website chat, and a knowledge base in a shared inbox with live updates for team members. It provides ticket status, priority, assignment, tags, SLAs, notes, automations, reports, a public help center, integrations, a REST API, signed webhooks, realtime sockets, and MCP support. Its built-in AI assistant, Ollie, drafts and summarizes replies, triages tickets, translates content, and answers chat questions from the knowledge base using Anthropic- or OpenAI-compatible providers, including local models. The system uses a Go binary for the API, background jobs, and event relay; an Elixir realtime gateway; and PostgreSQL. Changes are written as events in the same transaction as the underlying update, then published through NATS to the gateway, browsers, and webhooks. The hosted service and self-hosted installation use the same code. Self-hosting is available on Linux through a Docker-based installer, and the project is licensed under Apache License 2.0.
Open Screen Sharing is a macOS, iOS, and iPadOS screen-sharing client and server project by nekohasekai, based on the private Apple VNC protocol used by macOS Screen Sharing. It provides clients optimized for local and public networks, including Tailscale peer-to-peer connections, and adds exclusive mouse control, a debug HUD, fixed language switching, and interactions adapted for touch devices. Its iOS and iPadOS server implementation is based on Apple's iOS developer debugging protocol and can be accessed by the native macOS Screen Sharing app or its own client; the project also documents file dragging and server features that require sideloading because of App Store restrictions. macOS packages, iOS App Store and sideloadable builds, and nightly releases are provided. The project is licensed under the GNU General Public License v3.
Parin is a 2D game engine for the D programming language, positioned between a small library such as raylib and a full engine such as Godot. It uses a code-driven design without an engine-mandated architecture and lets developers choose between garbage-collected and manual memory management. Built largely on the portable Joka utility library, it includes pixel-perfect physics, dialogue, texture-atlas animation, tile maps, path finding, screen shake, and a simple UI library. Projects define a function called every frame and can be started with DUB templates; the engine supports WebAssembly, Windows, Linux, and macOS, with optional bindings or integrations for Box2D, Steamworks, and ENet.
Reason Network is a peer-to-peer coordination tool for teams of coding agents. A small daemon runs on each computer, discovers the team's agent sessions and reports their owner, repository, branch, status, and current work. Agents can use the MCP tools `list_team`, `send_message`, and `check_inbox` to contact a teammate's agent, while the command-line interface supports listing sessions, sending messages, notifications, and terminal replies. Computers connect directly over local networks or the internet using iroh; when a direct connection is unavailable, an encrypted relay carries the traffic. Offline messages are queued until the recipient returns, and the daemon installs agent integrations and hooks for tools including Claude Code, Codex, Cursor, Gemini CLI, Copilot CLI, opencode, pi, and Reason. The repository distributes macOS, Linux, and Windows installers and is licensed under MIT OR Apache-2.0.
Sill is a free, open-source remote-desktop app for showing one window or the whole Mac desktop on an iPhone or iPad. It captures the selected content with ScreenCaptureKit, encodes it as HEVC with VideoToolbox, and sends it over an encrypted direct connection; taps, clicks, scrolling, keyboard input, trackpad gestures, and Apple Pencil input are sent back to the Mac. Devices can connect over a local network, USB, peer-to-peer Wi-Fi, or a user-owned VPN such as Tailscale, with one-time pairing by scanned or entered code and no account or intermediary server. Sill for Mac requires Apple silicon and macOS 14 or later, while the iPhone and iPad app requires iOS or iPadOS 17 or later. It is distributed as a notarized Mac download, an App Store app, Homebrew cask, or source code, and is licensed under Apache License 2.0.
Snitch is a free, open-source, cross-platform real-time network-traffic visualizer from aixisstudio. It captures packets through libpcap or Npcap, parses IPv4/IPv6, TCP, UDP, DNS and TLS-SNI traffic, and presents connections in a live force graph, an interactive geolocated world map, bandwidth timelines and per-process views. It attributes traffic to applications, detects patterns such as port scans and beaconing, passively discovers LAN devices through the system ARP table, identifies trackers from editable suffix lists, and provides persisted alert-suppression and filtering rules. The backend uses FastAPI, WebSockets and SQLite; the interface uses React, D3.js and Electron. Traffic records remain local in SQLite, geolocation is offline using bundled DB-IP Lite data or optional MaxMind databases, and the API is loopback-bound with token authentication by default. Snitch visualizes and explains traffic rather than blocking it, so blocking requires a separate firewall. It is distributed for macOS, Linux and Windows under the AGPL v3 license.
SOMA (Solid Object Modeling with Agents) is a desktop app for designing parametric 3D models in CadQuery with a coding agent. It turns an agent's Python code into a model, rebuilds the model after changes, and lets users inspect it by clicking, measuring, or sketching in the view before sending corrections back to the agent. SOMA can work with CLI agents such as Claude Code, OpenCode, Codex CLI, and Gemini CLI; it does not provide an agent itself. A .soma file stores a model's code, result, and history. The app supports adjustable parameters, timeline-based preview and restoration, agent rule sets for processes such as 3D printing and CNC, overhang and wall-thickness checks, FEM strength analysis, animations, and exports including STEP, STL, 3MF, DXF, and PDF technical drawings. It can also import and export plain CadQuery scripts. SOMA is an early prototype distributed for Windows, Linux, and macOS Apple Silicon. Its software is licensed under the GNU GPL v3.0 or later, while exported STEP and STL files are not subject to the GPL; the project notes that file formats and its API may change during the 0.x releases.
The Agentic Scientist is an open-source AI research pipeline from Sakana AI that uses coding agents to turn a task description into experiments, ablations, evaluations, analyses, and a paper. Its six main phases are task drafting, investigation, ablations, sealed final evaluation, post-hoc analysis, and paper generation. During investigation, agents propose ideas, implement and run experiments, evaluate results against a task rubric, and iterate; later phases compare baselines and ablated variants, rescore the winner in a sealed environment, analyze run artifacts, and write and compile the manuscript. Runs are stored in a structured results directory, can be resumed or inspected through a local web interface, and can drive several agent CLIs and models. The recommended setup runs agent phases in Docker; the repository warns that agents receive full tool permissions, can execute arbitrary commands, and inherit supplied API keys, so runs should be isolated and tasks and sources should be trusted.
Timestral is a family of time-series foundation models from Mistral AI for multivariate probabilistic forecasting with past-only, known-future, and categorical covariates. Its standalone inference package accepts NumPy arrays or PyTorch tensors, forecasts multiple variates jointly, and returns 21 quantiles for each time step rather than only a point estimate. It supports missing observations, categorical covariate encoding, contexts of up to 32,768 steps, and CPU or GPU inference. The models use a causal patch-based transformer trained with Decoupled Horizon Forecasting. Separate prediction tokens read the causally encoded context; covariates condition target variables without reading them, and only known-future covariates extend into the forecast horizon. A quantile head produces non-crossing quantiles, while forecast patches are denormalized using context statistics and earlier predicted medians. The repository provides inference code and downloads model weights from Mistral AI's Hugging Face organization. The code is licensed under Apache 2.0; the model weights are distributed separately under a Modified MIT license.
Tracely is a trace-native CI/CD platform for AI agents that turns production failures into regression tests. It ingests agent traces over OTLP, promotes agent and conversation metadata to indexed fields, applies structural checks and LLM-based evaluators, and clusters related failures into issues. A failing trace can be promoted to a hermetic regression case containing its input, tool outputs, and model outputs as fixtures; the case can then be replayed deterministically and offline in CI, with a fail-to-pass contract and a gate that returns a non-zero status and can block a pull request. The platform also provides trace exploration, multi-turn scenarios, alerts, trends, judge calibration, an SDK and CLI, a GitHub Action, and an MCP endpoint. It can be self-hosted with Docker Compose or deployed with the documented Railway and Kubernetes configurations. The repository is licensed under MIT.
trim-1 is a small model, Go library, and command-line tool that selects the lines of a context an AI agent needs for a query. It accepts tool output, files, logs, or chats and returns the selected lines in their original order, inserting a marker where lines were removed rather than rewriting or summarizing them. The embedded 1.2 MB model uses 712 gradient-boosted trees rather than a neural network. It runs deterministically in milliseconds on one CPU core, with no GPU, network call, or external dependencies. The CLI supports conservative, balanced, and aggressive cutting modes; the Go API also exposes per-line keep probabilities through trim.Score. trim-1 is implemented in pure Go using the standard library and requires Go 1.23 or newer. It fails closed for blank or unsuitable contexts, including contexts below 2,000 bytes, above 8 MiB or 200,000 lines, contexts already containing a removal marker, or cuts that would remove less than 10% of the bytes. The repository is licensed under Apache-2.0, with copied Go math.Exp code separately covered by the BSD license noted in LICENSE-go.
tv-vibehack is an open-source toolkit for debloating Android TV and Google TV devices over Wi-Fi using ADB, while replacing the stock home screen with a lightweight, tvOS-inspired launcher. A setup prompt guides a coding assistant through identifying the TV, backing up packages and settings, asking which services to retain, disabling manufacturer telemetry, advertisements, content rows, and other preinstalled packages with `pm disable-user`, and installing the launcher; nothing is uninstalled or rooted, and restore scripts re-enable the disabled packages and return to the stock launcher. The included native Android TV launcher is written in Java with AndroidX RecyclerView and is designed for low-memory televisions. It provides a four-app dock, an app grid, HDMI and TV input controls, Wi-Fi and settings buttons, a clock, and an optional drifting clock screensaver. The repository also includes a pure-Python ADB client, batch and file-transfer tools, shell and PowerShell debloat/restore scripts, TV-specific debloat examples, and build-and-install scripts. It supports Android TV and Google TV devices but not Samsung Tizen, LG webOS, or Amazon Fire TV. The repository is licensed under the MIT License, with the included Outfit typeface under the SIL Open Font License.
web-ocean is a real-time ocean renderer written from scratch for WebGPU, without a game engine or rendering library. It simulates waves with six cascaded Tessendorf FFT fields, a fetch-limited JONSWAP wind spectrum, directional spreading, swell, capillary dispersion, and choppy displacement; a 17-level CDLOD quadtree provides the surface mesh. Shading includes whitecaps derived from surface folds, persistent camera-centered foam maps, anisotropic sun glitter, turbid water, sky reflections, and a physically based atmosphere with two cloud layers. The scene is rendered in HDR with 4x MSAA, bloom, AgX tone mapping, and dithering, and is viewed through a free camera with adjustable wind, swell, choppiness, foam, sun, haze, clouds, exposure, and bloom settings. The project can be opened directly from a self-contained HTML file in a recent Chrome browser with WebGPU, or run from source with npm; it is licensed under MIT.
yomado is a native macOS application that creates named software virtual displays for visual AI agents. It lets an agent such as Codex work with ordinary desktop applications—including Blender and Marvelous Designer—by placing their windows on a separate display while existing computer-use tools provide screenshots, mouse input, and keyboard input. Its bundled skill guides an agent through creating and naming a work screen, moving selected application windows into it, checking visual access, and returning them afterward. The app supports multiple virtual displays, configurable resolutions and optional HiDPI, menu-bar management, application selection through Finder, window-placement checkpoints, and restoration when a display is closed. Virtual displays remain in the existing macOS login session and share focus, keyboard, and pointer input; yomado is not an agent or a remote-desktop server. The project uses system frameworks and private macOS virtual-display APIs, requires macOS and a native Apple Silicon build is provided. It is an experimental prototype whose compatibility may change with macOS updates; physical-monitor standby, power savings, unattended operation, and reliable agent operation after a monitor is disabled are not validated. The source code is released under the MIT license.
Searchable transcript of GitHub Trending Weekly #54: git-ai, sill, Filee, Tracely-ai, Snitch, web-ocean, earshot, net-merge — Github Awesome (15:13). 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 54. 35 trending open source projects on GitHub right now. Let's go. GI moves the heavy parts of version control onto a server so coding agents can work on the same repository without each cloning everything. Agents fetch only the files they need while history queries and landing changes happen centrally.
00:24 You can also plug in your own conflict resolver, including an AI model. It's an independent open- source implementation inspired by Anthropic Scion design. Sill puts any Mac window or your whole desktop on your iPhone or iPad. Pair a device once by scanning a code. Then pick windows from a bar and control them with touch, a trackpad, a keyboard, or Apple Pencil.
00:48 It works over Wi-Fi or a USB cable and away from home through your own VPN like tail scale. On the folding iPhone Duo, the layout adjusts as you open and close it. Fleet puts a file converter right under your cursor. Hold a key, drag your files and drop them onto a format in the menu that appears. It converts locally and saves results beside the originals so your documents never need uploading.
01:13 You can batch conversions or watch a folder for incoming files. Some formats need optional engines which FY downloads when needed. Tracely turns failed AI agent runs into regression tests you can run before merging another change. It grades production traces, groups similar failures into issues, and lets you freeze a bad run with its recorded model responses and tool outputs.
01:38 Those fixtures support repeatable tests without fresh model calls, provided your agent uses its replay integration. Snitch shows where your computer's network traffic goes and which apps are responsible. Connections appear in a live graph and on a world map with per app bandwidth history to help explain what's happening in the background. Traffic records stay on your machine and even IP geoloccation works offline.
02:07 It visualizes connections and flags suspicious patterns. Blocking traffic still needs a separate firewall. Open screen sharing is a better client for Apple's built-in Mac screen sharing written in Go and Swift on top of the protocol Apple's own screen sharing app uses. It runs on Mac OS, iPhone, and iPad, and it can also connect over RDP. It's tuned for both your local network and the internet, including direct tail scale connections and adds exclusive mouse control and a debug HUD.
02:39 Web Ocean renders a moving ocean in your browser using web GPU with no game engine or rendering library. Simulated waves form white caps, sunlight glitters across the surface, and a free camera lets you fly over the water. You can adjust wind, swell, and sunset lighting as it runs. The whole scene opens from one self-contained HTML file, though you'll need Chrome with web GPU and a capable GPU.
03:05 Earshot gives coding agents audio feedback they can reason about. It describes sound files, finds moments matching a description, and checks how effects line up with video cuts. An agent can also synthesize effects from code, then revise them using measurements and audio model scores. The core tools run locally on CPU through a CLI or MCP server, though those scores still need a human listening check.
03:29 Netmerge combines several internet connections like Wi-Fi, Ethernet, and a phone hotspot, so your apps use all of them at once. It's written in Go and creates a virtual network adapter with no account, no server, and noVPN. Each new connection goes to one of your links, and faster links get more of them. If a link drops, it stops getting new connections until it comes back.
03:56 Lean RIA gives coding agents a compact guide to inspecting compiled apps with ordinary command line tools. It routes each file type to a useful first command, then narrows the search through strings, imports, or extracted code. You can trace an electron feature or check a go binary dependencies without an MCP server. It also tells the agent to distinguish what it found from what it inferred.
04:22 TV Vibe hack replaces your Android TV's cluttered home screen with a lightweight launcher built around your apps. Give it setup prompt to a coding assistant. Choose your favorites and it backs things up before disabling unwanted packages over ADB. You get a four app doc, an app grid and navigation designed for the remote. Packages stay installed but disabled and a restore script brings the stock launcher back.
04:49 Harness Squared improves the setup around an AI agent while keeping the model's weights fixed. It proposes changes to prompts, memory or tool guidance, tries them on the current task, and combines useful changes by comparing how the runs behaved. Lessons carry forward to later tasks. You can explore alternatives in parallel or refine them sequentially with saved execution evidence to inspect alongside the adopted configuration.
05:15 Motion prompt helps when you can picture an animation but don't know what to call it. It pairs visual samples with names, explanations, and example prompts across motion graphics, camera moves, UI, web effects, and photography. Browse the gallery, pick what you want, and pass those references to your coding AI. There's also a skill for claude code and codecs that brings the specs and reference code into your project.
05:42 Timestraw is Mistl's family of models for forecasting data over time with related signals predicted together. You feed it historical observations, add information you already know about the future, and get a range of possible outcomes rather than one confident looking number. The inference package runs on CPU or GPU, so you can try it without a dedicated graphics card.
06:07 Reason network lets your coding agent message a teammate's agent directly, even when they use different tools. A small background service on each computer shows who's working in which repository and branch, then routes messages into the right session. It supports tools including clawed code, codecs, and cursor with encrypted connections and cued messages for offline teammates.
06:32 Trim 1 is a tiny model that decides which lines of a context an AI agent actually needs for its current question. It's written in pure go and runs on 712 gradient boosted trees in a file of about a megabyte, not a neural network. Give it a log, a file or a tool result along with a query and it keeps the useful lines in their original order. It marks wherever something was cut.
06:57 Yumato is a native Mac OS app that creates virtual displays for visual AI agents built only on Apple's own frameworks. Give a task its own named virtual screen, then drag an app like Blender onto it. An agent such as Codeex can work there with screenshots, mouse, and keyboard while your main screen stays free. A bundled skill walks the agent through creating the screen, moving the windows, and putting them back afterward.
07:24 High memory gives AI agents a memory system that separates what happened from what's worth reusing. Its working mode stores session facts alongside lessons for future tasks. for coding. It also keeps a chain of actual code changes, including revisions that summaries can lose. Tenscent includes a Python SDK and Memhub, a desktop and command line tool for recalling history across authorized AI apps.
07:50 Openesk brings support email and website chat into one shared inbox, so your team can assign conversations and see updates live. Its built-in assistant, Ali, drafts replies and answers questions from your knowledge base with support for local models. You can publish a help center alongside it. The self-hosted edition runs the same code as the hosted service, so you can keep support conversations on your own infrastructure.
08:18 64 HW Simulator lets you break a virtual Commodore 64 one pin at a time and watch what happens. pull a chip, ground a signal, or disconnect a resistor, then inspect the screen and oscilloscope. The symptoms emerge from simulated board connections and chip behavior. You can explore repair hypotheses without touching a soldering iron, though a matching symptom is only a clue, and some electrical behavior remains approximate.
08:45 Binny connects an AI model to reverse engineering tools and turns binary analysis into a report you can review. It reads decompiled functions, strings, and imports, records findings, and saves progress after each step. Local GIDra support keeps decompilation on your machine, while analysis scripts run in a restricted Docker sandbox without network access.
09:09 FrameTell turns your notes, PDFs, or a single idea into an animated explainer video. Its director agent drafts a storyboard for your approval, then writes narration, generates a voice over, and builds the animated pages. You can point at a page and ask for changes before exporting a subtitled MP4. Rendering runs locally, while the model and voice services use your own API keys.
09:33 ESP32 SIM runs ESP32 firmware on your computer or in a browser, so you can test supported projects without plugging in a board. It boots real ROMs and boot loaders with models for displays, buttons, and other peripherals. The desktop version can connect firmware to a virtual Wi-Fi network and real services. Hardware coverage varies by chip, though, and the browser version has no network pass through.
10:00 MCP for Blender Pro lets an AI assistant inspect and edit your Blender scene through chat. Ask it to arrange a product display or change materials. Then inspect the result in a viewport preview. You can connect cloud models or local Alama models and expose the same scene tools to an MCP client. Built-in chat includes operation approvals and named undo steps so you can review changes as they happen.
10:24 Lucian UX UI prompt lets you try an interface before asking your coding agent to build it. Browse interactive buttons, charts, or animated website sections. Then copy a prompt containing the working code and instructions for adapting it to your project. You can also take the raw HTML, CSS, and JavaScript yourself. The elements include light and dark themes, keyboard controls, and reduced motion support with no dependencies.
10:54 Mayo is a code editor written in Go that draws its interface directly on the GPU. You can split files into columns, edit with Vim motions, and stage individual changes from the gutter. A built-in terminal uses Ghosty's emulator. Its remote SSH mode is a useful detail. The file explorer, Git, and Search run on the remote host through your systems SSH with nothing extra installed there.
11:18 Enclose is a Mac OS workbench for CAD projects that are built from code powered by Claude. Open a model in 3D. Click the part you want to change and describe the change in plain words. Claude edits the project's code and rebuilds the model so you see the new version right away. It reads step files and runs on both Apple Silicon and Intel Max. Singing 3D parts by talking to a coding agent built on CAD query.
11:45 Describe a case bracket or mount and an agent like clawed code or codeex writes the model as Python. S rebuilds it on every change. You can click, measure or sketch on the view to point out a problem and send that back to the agent. It checks overhangs and wall thickness, runs strength tests, and exports to STL or step. Anan iPhone use is ESP32S3 firmware that turns a small M5 stack stopwatch into an MCP server for controlling an iPhone.
12:16 The board plugs into the phone over USBC and shows up as a mouse and keyboard with assist of touch handling the pointer. Your AI client watches the screen through screen sharing, decides the next step, and calls simple tools like tap, drag, and type. Co-op anything gives AI coding agents a playbook for adding co-op multiplayer to single player PC games.
12:38 An MCP server works out a game's engine from its install folder and checks a directory of more than 230 co-op mods. From there, the agent either finishes an existing mod or reverse engineers the game and builds co-op from scratch. Bots test every build with real game clients over a Steam relay so friends can join with no port forwarding. Casera starts as a flat architectural floor plan, then builds a house around you as you scroll.
13:07 The drawing is a slice through the actual 3D model. Scrolling raises the cut, reveals the walls, and takes the camera through the rooms. Materials appear as you enter each space. It uses 3JS with geometry built-in code and falls back to images and video on devices that can't render it. Marorrow puts notes and databases in a desktop workspace where each page stays a markdown file in your own folder.
13:32 You can open an Obsidian vault or import a notion export, then organize pages into tables, boards, or calendars. Database values remain readable file properties, so they're accessible outside the app, too. It works offline without an account with built-in encrypted sync between computers on the same network. The Agentic Scientist version three of Sakana AI's AI scientist turns a short task description into experiments, ablations, and a written paper, all driven by coding agents.
14:03 The agents propose ideas, run experiments in Docker containers, and score the results against the tasks rubric, then try again. After that come baselines, a sealed final evaluation, analysis, and the paper itself. Most peer-to-peer file tools make you trade one-time codes every single time you want to move a file. Flip-flop pairs your devices once and keeps a persistent trusted tunnel open between them.
14:29 It routes transfers over Bluetooth, local Wi-Fi, or straight across the internet with automaticnat hole punching. You can push files or clipboard text between Linux, Mac OS, Windows, and Android whenever you want. No repairing required. Parin is a simple 2D game engine for the D programming language. It aims for the space between a small library and a full engine like god do.
14:54 You write one function that runs every frame, and you're free to structure the rest of your game however you like. It comes with pixel perfect physics, a dialogue system, animation from texture atlases and tile maps.