This GitHub Trending issue features 35 popular open-source projects.

Hindsight: A Memory System for Agents

Hindsight is a memory system for agents by Vectorize, designed to help agents learn from their actions, going beyond simple recall of past dialogues. The system categorizes stored information into: world facts, the agent's own experiences, observations reinforced by multiple memories, and mental models that are updated as the agent learns. Retrieval occurs in parallel, utilizing semantic, keyword, graph, and temporal search, followed by ranking of results.

RelaDraw: A Language for Diagram Creation

RelaDraw is a language for creating text-based diagrams, occupying an intermediate position between the automatic layout of Mermaid and manual drag-and-drop. Users can describe relationships between objects, such as "to the left of the hub" or "between desktop one and laptop one." A resolver interprets these descriptions as minimum distances and finds the most compact layout. The system also reports overlaps, crossed edges, and text overflow without requiring rendering.

Whiteboard: Collaborative Software Development with Agents

Whiteboard is a desktop canvas designed for collaborative software development between a user and their coding agent. The coding agent (codec) draws sequence and entity diagrams on the canvas. Clicking on a diagram element takes the user to the corresponding code. An integrated diff viewer can collapse large functions into pseudocode and hide tests. Agents also log their decisions, allowing the development process to be tracked.

NeoHorse: Models for Tool Coding and Usage

NeoHorse is a family of open-weight models trained on top of Quen 3.5 for coding and tool usage tasks. The training system distributes tasks to the models, records their actions and outcomes, and then uses this feedback to form the next training batch. The release includes 4B and 9B language models, as well as a 4B Jev variant that evaluates predefined decisions. The repository describes recursive self-improvement as one of the project's goals.

Company-Brain: A Slack Bot for Team Memory

Company-Brain is a Slack bot from Super Memory that remembers information discussed by the team and can act upon it. The bot only has access to channels that the requesting user has access to. Thus, a query in a public channel uses shared memory, while a direct message (DM) can also access private channels. The bot can create issues and read pull requests (PRs) using the user's credentials. It can be deployed locally on Cloudflare Workers, but memory is still processed via the Super Memory API.

CLM: A Contrastive Language Model for Decision Making

CLM (Contrastive Language Model) makes agent decisions by matching embeddings instead of generating text. The model encodes the state and each possible action separately based on a Quen 38B backbone with small trainable heads, and then selects the action that best matches the state. CLM is provided via a Typesafe-compatible API and, according to the authors, achieves parity with Jev at up to nine times lower latency.

Cogsend: Social Media Post Scheduler

Cogsend is a self-hosted social media post scheduler that allows you to publish content to Mastodon, Blue Sky, LinkedIn, Threads, and X from a single draft. The user writes the text once, adjusts it for each platform, and publishes immediately or on a schedule. The system runs on Cloudflare Workers using D1 and R2, and cron jobs that run every minute send scheduled posts. Failed posts are retried up to five times before being moved to a failed list. Credentials are encrypted at rest, and the administrator account has two-factor authentication (2FA).

Disk Tree: Visualizing Disk Space Usage

Disk Tree displays disk contents as a map of nested blocks, with their size corresponding to the actual space used by files. Users can zoom into folders, identify large caches or old builds, and mark files for deletion. A full list and information about potentially freed space are provided before deletion. The preferred deletion method is moving to the Recycle Bin, and irreversible deletion requires confirmation. The program runs on Linux, macOS, and Windows.

InkWave: A Browser Game of Territory Painting

InkWave is a browser game where victory is achieved by painting a larger territory than the opposing team. Players engage in 3v4 matches against bots, switching between seven weapon types and a squid mode that allows movement through ink and climbing painted walls. The game is built using 3JS and runs without a build step. Characters, animations, textures, sound effects, and music are procedurally generated.

Decision Models: Local Execution of Decision Models

Decision Models allows for the local execution of decision models. Similar to Olama, which enables running language models, this system takes a message, email, or application and answers questions like "Is this urgent?" or "What is the client asking for?". Instead of generating text, it returns scores and probabilities. Models can be loaded by name, provided via API, and connected to coding agents through MCP.

Jive: A Terminal Coding Agent

Jive is a terminal coding agent that transforms a plan into an executable graph. Instead of querying a large language model for the next action after each tool call, Jive can execute related steps and use smaller decision-making models along the way. This makes it useful for repetitive tasks, such as checking a large number of records or tracking down bugs in a codebase, where constant model calls increase execution time.

MW: An Agent with a Dedicated Judge

MW is an agent that uses a small "judge" to handle routine decisions, while the main model focuses on code. The judge determines which tool output to include in the context, whether an action aligns with user instructions, and when a task is effectively complete. Each decision is logged for review. The judge can be run in "shadow mode" to see its decisions before they impact the agent's operation.

Clifact Shortcuts: Automating Agent Commands

Clifact Shortcuts analyzes your coding agent's command history to identify frequently used prompts. It groups similar prompts, highlights recurring ones, and allows you to convert useful prompts into short project commands. For example, the prompt "did the email reach all recipients?" can be converted into a /sent command with the appropriate query and reporting steps. The history extractor runs locally, and the user chooses which shortcuts to create.

Anadoodle: Generating Drawn Art with Code

Anadoodle creates drawn art with code, enabling the generation of illustrations, sped-up drawing process videos, or animated films from the same source code. The tool includes 31 styles, ranging from charcoal and watercolor to woodcut and pixel art. Music composed by code can also be added. A key feature is consistency: when re-rendered at a different size, elements remain in place.

Keel: A Unified Workspace for Coding Agents

Keel unifies the coding agents you use into a single, native macOS workspace. Users can track sessions, view changes, use the terminal, and see the agents and files they are working with as a live graph. For new tasks, the local Lia model can help select the right agent. Keel validates this choice and records the decision, allowing for task routing to be tracked.

Logo Design Skill: Creating Logos with Claude

Logo Design Skill assists Claude in the logo creation process, from the initial brand brief to testable SVG concepts. The system checks how each element appears at small sizes, in single color against various backgrounds, and alongside competitors. It also features a database of over 1400 real-world logos for studying design patterns. Once a direction is chosen, the system can generate a set of icons and a brand guide.

Ship Video: Creating a Promo Video from a Web URL or Prompt

Ship Video creates a promo video from a website URL or a short text prompt. An AI agent writes the animation as a single HTML page, checks its appearance at different points in time, then renders each frame in a headless browser and encodes an MP4 using ffmpeg. The advantage is that the video is generated from editable web code, providing a clear source for text, colors, and animations.

Tidewater: A Browser-Based Fishing Game

Tidewater is a browser-based fishing game set on a tropical island that can be explored on foot, underwater, or by boat. Fish bite in different locations and at different times, and players must manage line tension to catch them. Catches can be sold, gear can be upgraded, and then players can venture further into the open sea. The game's ocean and sky are powered by a custom WebGPU engine, requiring a compatible browser and GPU.

UE2 Studio: Generating Music with Vocals on PC

UE2 Studio generates complete songs with vocals on your Windows PC. The user specifies the style and lyrics, after which UE2 writes the melody and chords in musical notation, and then performs the track. The musical notation can be edited, and then the same composition can be rendered with a different sound. The system runs on UA2.cpp, C++, and, according to the author, generates 338 songs in approximately 46 seconds on an RTX 4090.

Rooms: Project Window Management on Mac

Rooms turns each project on your Mac into a saved set of windows and their arrangement. When you press Option+Space and type a room name, its windows are restored, and everything else is hidden. The system can adapt window arrangements for a laptop or an external monitor, and the user can save their own arrangements. Windows remain open, and room data is stored on your Mac.

Jevy: A Browser Agent for Task Automation

Jevy is a desktop browser agent that plans tasks using a language model and selects each action with Jev from Typesafe. The agent analyzes the page, converts available controls into a list of options, from which Jev selects one. This ensures the agent does not create non-existent selectors or write its own JavaScript. Redirection is possible mid-task, and 12 tasks have been tested in a web arena. The author reports 10 successful completions.

Phantomat: A Hyperland Plugin for Infinite Canvas

Phantomat is a Hyperland plugin that replaces workspaces with a single infinite canvas. Each window resides on the same scalable plane, and multiple monitors display adjacent parts of the canvas, panning synchronously as a unified desktop. Pressing Super+Control+G zooms out to show all content. Then, by typing a portion of a window's title, you can instantly jump to it. The system remembers the arrangement of all windows across restarts.

Product Film: Creating Promo Videos and Landing Pages

Product Film is a coding skill (clawed code) that creates promo videos and looping animations for landing pages from real product components using Remotion. It first reads design tokens and logos, then generates a brand kit and a bit sheet. A Python script maps the track's bit grid to music so that transitions align with the rhythm. The user approves static images and contact sheets before the final rendering, which outputs a master file at 240 fps with motion blur.

Omnibin: Nix Package Management Without Installation

Omnibin adds tens of thousands of Nix package commands to your PATH without requiring them to be installed beforehand. A virtual filesystem determines which package provides each command and only downloads files from the Nix cache when the command is used. You can even invoke a specific version, such as an older version of Python. This is convenient for one-off Linux environments, although the first run of a tool requires a download.

Changelog.earth: Planetary News in Patch Note Format

Changelog.earth presents news in a patch note format for the planet. The system collects reports, selects several stories, and transforms their headlines into concise updates on discoveries, conservation successes, and unresolved issues. Each note includes information about the original publisher, date, and a link to the source, allowing for verification of the primary source. Previous issues are also archived, and an RSS feed is published.

Flux: Sync Your Desktop and Android Phone

Flux connects your desktop computer with your Android phone over a local network. Once paired, you can transfer files and clipboard text, view phone notifications, control media, and even use your phone as a webcam or microphone. There's a desktop app for everyday use and a CLI for scripting. Pairing displays a verification code on both devices, allowing you to ensure the connection is correct.

Interference Search: Investigating Reasoning via States

Interference Search is a research project investigating reasoning via explicit states. It advances several possible moves simultaneously, merges branches that reach the same state, and uses a small judge to prune dead ends. This approach has shown good results on the Countdown numbers game, while improvements on coding tasks were significantly smaller. The repository contains code, model weights, raw results, and records of failed experiments.

Bridge Clip: Creating Short Clips from Long Videos

Bridge Clip transforms long videos, for which you have permission to use, into short, subtitled clips. The user uploads a file, the system transcribes the speech, selects promising moments, reformats them for vertical or widescreen video, and renders word-by-word subtitles. The final video is assembled on the user's computer using ffmpeg. Transcription and clip selection use your OpenAI account, so audio and transcription data are sent to that provider.

Omari Meeting Recorder: Record and Transcribe Calls Locally

Omari Meeting Recorder records your calls without a bot joining them, and the audio never leaves your device. The system captures audio from your microphone and computer audio as two separate tracks, then locally transcribes them using whisper.cpp and tags who said what. Your encoding agent adds chapters, and the only thing it sees is the transcription text. The user receives a player with a waveform and an editable transcript.

RetraData: Command-Line Disk Image Analysis

RetraData is a Rust-based command-line tool for analyzing disk images without modifying them. It reads MBR and GPT partition tables, verifies header and record checksums, flags overlapping partitions, and identifies files by their signatures. It can also generate empty or intentionally corrupted images for testing. Everything is output in JSON format, and the tool operates entirely offline.

Halo Battery: Displaying Wireless Device Battery Levels

Halo Battery displays the battery level of each wireless peripheral in the Windows system tray, allowing you to close Synapse. Each device receives its own ring icon, which fills clockwise, turns amber when the battery is low, and glows green while charging. The system directly interacts with the Razer Black Shark V2 Pro, WL mouse Beastex Max, and Game Seir G7 Pro, and also reads battery levels from Bluetooth devices via Windows.

Jev Use: Control Your Mac with Voice or Text Commands

Jev Use allows you to speak or type a command and have it executed on your Mac. The system reads the accessibility controls of the current application, prompts Jev to choose the next action, and then re-checks the screen after execution. This means it can work without sending screenshots. While your command and the details of visible controls are sent to Typesafe, it requires a Typesafe API key and macOS accessibility permission.

Omoto: A Layered Image Editor for Linux

Omoto brings the Compositor's layered image editing experience to Linux, with a special focus on Omari. It even follows the desktop theme. Users can work with masks, selections, retouching tools, adjustment layers, and offline background removal. The program also opens and saves Compositor projects, allowing for the transfer of layered edits between Linux and Mac without flattening them into a single image.

Jevp: Assisting Coding Agents in Finding Files

Jevp helps coding agents find the right files in an unfamiliar codebase. When asked "where is authentication checked?", the system traverses the repository, identifies relevant files, and returns source code snippets with line numbers in a single response. An accompanying skill tells agents when to use Jevp. The relevant source content is sent to a provider you configure, so you should check what data you are willing to send.