This edition of GitHub Trending features 35 promising projects, covering a wide range of technologies, from artificial intelligence and desktop application development to tools for data manipulation and multimedia content creation.

Artificial Intelligence and Agents

Decider offers AI agents a fast way to make structured decisions. The system processes a situation and given questions (e.g., "Which team is handling this?", "Is this urgent?", "How serious is the problem?"), returning probabilities for valid answers in a single model pass without needing to parse free text. The project offers both small models for laptops and larger options for GPUs, as well as an API for integration into applications.

Whirl combines multiple AI models into a single chat application, allowing users to switch between them within a conversation. The app remembers preferences between chats, connects tools via MCP, and saves generated documents, diagrams, and interactive pages in an editable format alongside the conversation. Web search, voice input, and image generation are also available. Whirl's full source code is available under an open license with a guide to setting up your own instance.

Livvenerf explores the hypothesis that AI model performance degrades after launch. The project runs the same carefully selected questions daily through Claude Opus 5.5 and Claude Code models, logging prompts and CLI versions. Each response is scored and compared against an early baseline according to published rules. A 30-day test is in progress.

Comma is a personal AI agent that runs 24/7 and manages requests from start to finish. The user formulates a request once, which is then converted into tasks on a board, processed, and paused for user decisions at key stages. Autonomous loops monitor email, repositories, and feeds, activating the agent only when tasks are present. The agent can be accessed via a Mac app, browser, and Signal and Telegram messengers.

Explain allows users to request an explainer video from a coding agent, which is then generated as a narrated MP4. The agent writes the script and creates scenes in JavaScript, managing narration, word syncing, animation, music, and export. The project also saves preview frames and checks text for clipping or mispronunciations, allowing the agent to catch errors before final export.

Open Dots is a self-hosted platform for creating AI assistants with different roles. One "dot" can perform research tasks, another can write copy, each with its own browser and files. Work results are displayed next to the chat, and a review card allows approval of drafts before saving. The template also includes capabilities for voice calls and Slack integration, which require separate setup.

Clear AAI is a plugin for Deep Seek Harness designed for research tasks. It asks deeper questions than simple task completion by an agent: it determines what might disprove a claim, runs a test, logs evidence, and outputs a result with limited output. Supported findings form a knowledge graph. Disproven claims also remain visible. When contradictions arise, the system flags them for human decision-making.

AON brings modern AI chat to a 2007 Nokia phone. The user types a question on the keypad, selects a model (Claude, ChatGPT, Gemini, or Grok), and reads the response on the small screen. The system supports web search, photo handling, and voice-to-text conversion. A small Go server handles communication between the old phone and modern model APIs. The project has been tested on Nokia 6300 and E63 models.

LinkedIn agent skill provides Claude with 11 skills for managing LinkedIn content, including drafting posts, planning the week, suggesting comments and replies, profile review, and transforming long-form content into short posts. A shared voice file helps make drafts sound like the user's style, while local scripts identify common AI phrases and awkward constructions. Content publishing remains the user's responsibility.

Remach Studio allows users to create product promo videos by chatting with a coding agent. The user describes an idea, watches a real-time preview, and can then point to a line of text or a frame to make changes. The agent creates the animation design as a project that can be edited and exported. It's a macOS app built around a user-owned project folder.

Vibe Wise turns Claude Code into a coding partner that helps users think through the development process. Before writing a function, the agent asks the user for approaches to complex problems, helps weigh trade-offs, and shows a design for approval. It then writes and tests the code, explaining the changes made. The frequency of stops can be adjusted based on user experience and available time.

Television provides a visual space for a personal AI agent to place its work. The agent can create charts, documents, web pages, or small applications on a separate screen where the user can view and refine them. The system runs in parallel with a normal chat interface and supports agents with file and skill access. Currently, the agent must run on Linux or macOS.

Muse Gadget SDK allows users to turn an ESP32 board, Raspberry Pi, or Linux device into a custom Muse gadget. A screen, buttons, audio, or sensors can be added, and the device can then be paired with the Muse mobile app. The repository includes open-source firmware and an SDK for ESP32 and Linux, allowing users to start with off-the-shelf hardware and define gadget functionality.

Audio CN provides React developers with ready-made components for building audio applications that would typically be developed from scratch: level meters, oscilloscopes, faders, knobs, channel strips, and a full-featured mixer. Components are added via the Shad CN CLI, and their source code is placed in the project for further styling or modification. Web audio hooks are also included to connect controls to microphones, playback, and live audio signals.

Lipflow allows users to input text without speaking words. By holding a key, the user silently speaks a sentence to their webcam, and the system inserts the most likely text into the cursor. The app runs on macOS, Windows, and Linux, and can be trained on a user's own face and vocabulary. An optional post-processing model and a whisper mode, combining lip and audio analysis, are available for increased accuracy.

Strands decider handles the small decisions an AI agent makes throughout the day: choosing a tool, routing a request, or needing to double-check a response. The system accepts options or a rating scale and returns a choice with a confidence score instead of a long answer. It's a compact model that can be run locally, with examples of integration into Strands agents.

Backburner uses an iPhone to accelerate local models with 27 billion parameters on a Mac. When connected via USB-C, the phone helps process long prompts or store context that doesn't fit into the Mac's memory. In project tests on a MacBook M4 Pro and iPhone 17 Pro Max, prompt reading times for large inputs were reduced. The project is in a pre-release stage and has specific hardware requirements.

SCM (screen memories) makes photos and videos in Mac folders searchable by content. A natural language scene description allows finding a relevant frame, even in the middle of a video. Searching by on-screen text or exact dialogue is also possible. Watched folders keep the library up-to-date, and models run on the Mac after the initial download.

Mesh Avatar Studio turns a single illustration into a 2D character that blinks, talks, tilts its head, and moves its hair. A coding agent separates the image into layers and creates a rig. The user then opens a local editor, moves points to adjust, and sees the pose update in real-time. Eye and mouth shapes can be tested before using the avatar for more natural animation.

Omnirush is a desktop coding agent for macOS and Linux. When a project is opened, the user selects a model and detail level, and the agent handles code changes and Git tasks in a unified interface. MCP skills and tools are supported, along with workflows for branches, worktrees, commits, and pull requests. An Omnirush account can be connected, or custom provider keys can be added.

Neon Plan 3D transforms a smart home dashboard into an interactive 3D model of a house. The user draws rooms, walls, doors, and furniture in a built-in editor, then links them to devices. Lights display in true colors, curtains move, doors open, and cameras show what they see. The home can be controlled from the 3D view or by placing it on a wall-mounted tablet.

Feluca adds a new set of sounds to the MVE FM1 synthesizer. The custom firmware includes nine sound engines, from analog and FM synthesis to chiptune, samples, and granular textures. Four tracks, a 64-step sequencer, effects, and a browser-editor for tweaking sounds and patterns are available. Installation is done via USB using Chrome or Edge without additional hardware.

Jevbox is a document library that gives teams more than just folders and a search bar. Users can upload files, browse them in a visual search, inspect broken-down sections, and ask questions with citations linking back to the source. The system can automatically distribute new documents to relevant library sections and respects each member's access rights in search and chat. It's a full-featured starting point for self-hosting.

Hairline adds 19 miniature isometric scenes to websites that react to the cursor. Hovering over a column field makes them rise, and moving across a "disassembled" laptop causes its layers to spread apart. Components can be embedded in React or any page with a DOM, with customizable colors and motion. A skill is also included to help a coding agent draw new figures in the same style.

Pipod allows users to run PI coding agent sessions in isolated environments on a controlled server. A sub-session is launched for the current project from the terminal, and then can be connected to via CLI or a mobile app. The server manages the creation and connection of sub-sessions, and a single Compose configuration ensures the self-hosted stack runs. The solution is for those who want remote access to their coding agent without needing to keep it on their laptop.

Mobile Dev for Codeex integrates an iOS simulator or Android device directly into the Codeex chat. Users can observe app behavior, interact with screens, direct Codeex to a specific element, and output native and JavaScript logs to the conversation. CPU, memory, and frame rate performance metrics are also recorded, allowing for visual demonstration of performance issues rather than just describing them.

Papermorph transforms PDF documents into interactive web books. Given a chapter and target audience, Claude Code creates a plan, storyboards, narrative, animations, and quizzes before generating the web experience. The project includes a live shelf with examples. It is currently intended for English-language content using Claude Opus 5.5.

OMASVM allows users to run Omari's Linux desktop on a Mac with Apple Silicon via a virtual machine. A single setup command creates an image for OMAC VM, UTM, VMware Fusion, or Parallels applications. Mac features like trackpad gestures, media keys, Wi-Fi, audio, and a shared clipboard are then integrated so the Linux environment feels native on the hardware. Features can be modified later, and a check can be run if issues arise.

Cartcha converts bank transaction text into an expense log accessible on Android. The system can read messages from approved senders, extract the amount and merchant name, and keep the original SMS alongside each entry for correction. Budgets can be set, and spending patterns can be viewed without needing to create an account. The log is stored on-device with optional encrypted backups.

Jumper is an open-source crab robot that is designed and trained by describing desired behavior in a single sentence. An AI assistant creates the robot's appearance as a skin file, trains movement, or builds a scene, and then implements it. The project includes guides for reproducing and evaluating movements, as well as packaging a trained controller for the robot.