GitHub Trending Weekly #50: openmuse, open-glean, unreal-agent, BongoCat, laya-mlx, mini-AGI, flute

Author: Github Awesome · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: GitHub Trending Weekly #50: openmuse, open-glean, unreal-agent, BongoCat, laya-mlx, mini-AGI, flute · Channel: Github Awesome · 14:38 · Watch on YouTube

Topics: AI agents, Task and workflow automation, AI developer tools, Local and open-source models

Summary

This episode surveys 35 trending open-source GitHub projects, spanning AI agents, local decision models, developer tools, creative software, and utilities. Several projects focus on making coding agents more manageable through shared interfaces, task orchestration, browser or desktop control, and safeguards such as logging and verification. Others use small or local models for structured decisions rather than generated prose, with examples for document sorting, form filling, and classification. The lineup also includes creative tools for motion design and product videos, as well as projects for Android decompilation, Mac terminal sessions, and lossless compression.

Key points

Open Muse is a personal AI agent with a persistent browser, optional isolated Linux terminal, and access to Gmail and Google Calendar; users can take over mid-task.
Open Glean searches files, notes, and connected apps, then produces cited answers; its deep research mode runs parallel searches.
Several projects target structured decisions rather than text generation, including Lya, Reflex, Bespoke Nimble, and the Node.js/TypeScript Lya package.
Tools such as Zcode, Magpie, AI Manager, and Mr. Mac Workspace bring coding agents or their configuration into shared desktop, web, or terminal interfaces.
Agent orchestration and safety projects include Unreal Agent’s recoverable operations, FormMerge’s intent conflict checks, and Jev Router’s permission and risk filters.
Browser and desktop automation projects include ArcCUA, Fast Browser Use, Fluid Use, and Jev Voice Browser.
Creative projects include Null Motion for frame-synced motion-design breakdowns, Rizo Windows 8’s procedural films, and Flute’s 3D product shots built from React interfaces.
The episode also features DDC, a Rust Android decompiler; Herder GPUI, a native Mac interface for terminal sessions; and Glide, a lossless compressor.

Chapters

Start	Chapter	Summary
00:00	AI agents and local decision models	The episode introduces projects for visible personal agents, file research, recoverable agent runs, desktop companions, and small models that make structured choices.
04:12	Coding tools and creative applications	The roundup covers video search and coding-agent configuration alongside motion-design analysis, text-to-tool workflows, procedural films, and cinematic product shots.
06:50	Computer-use agents and coding workflows	Projects hand off desktop tasks, detect conflicts between coding agents, automate browser actions and forms, and organize coding CLIs and their sessions.
09:24	Agent orchestration, skills, and interfaces	The video presents multi-model coding orchestration, tool selection with audit logs, Claude skills, animated loading indicators, and voice-controlled browsing.
11:30	Decision systems and document processing	Local decision models and a Node.js package are followed by PDF and document classification, agent decision logging, and an Android decompiler.
14:06	Self-hosted inference, terminal sessions, and compression	The final projects cover a self-hosted decision API, a native Mac window for persistent terminal sessions, and a lossless compressor for structured data and version backups.

Quotes

35 trending open-source projects on GitHub right now. — 00:00

If an agent receives the same event twice, you don’t want it doing the job twice. — 01:04

Flute turns the React components you already have into cinematic 3D product shots. — 06:18

Logging the guess before the result is the part worth copying. — 13:04

Close the window and your work keeps running until you attach again. — 14:06

Concepts and entities

Open Muse, OpenGlean, Unreal Agent, Laya MLX, Mini AGI, Bespoke Nimble, Magpie, Flute, Jev, DDC

Full transcript (original language)

00:00 Welcome back to GitHub Awesome. This is GitHub trending weekly number 50. 35 trending open-source projects on GitHub right now. Let’s go. Open Muse is a personal AI agent from Copilot Kit that does its work where you can see it. It gets a persistent browser, an optional Linux terminal in an isolated container, and read-write access to your Gmail and Google Calendar. And you follow the plan as it runs. If it’s heading the wrong way, you take over the browser or terminal mid-task. It runs on iOS, Android, and

00:33 the web from one Expo app. Search is a Mac browser built to stay out of your way. It uses the WebKit already on your machine, keeping the app around 3 MB. You get tabs, one field for addresses and searches, an ad blocker, and a way to hide clutter that keeps coming back. Passwords live in the macOS keychain. There’s no account or sync, so your browsing history stays on that Mac. Open Glean searches your files, notes, and connected apps through Hydra DB,

01:04 then uses your chosen model to write answers with source citations. It’s deep research mode splits a question into smaller searches, runs independent branches in parallel, and combines their findings. You can host the interface yourself and choose an OpenAI compatible model endpoint. If an agent receives the same event twice, you don’t want it doing the job twice. Unreal Agent is an async-first Go harness built around that problem. It deduplicates inputs, persists sessions

01:35 and operation state, and supports recovery and forks. Tool calls become serializable operations that execute outside the coordinator’s event loop. That separation matters when a run fails halfway through. There’s a record to resume from. Zcode puts the same coding agent runtime behind a desktop app, browser workspace, and terminal interface. You get a desktop app, a web workbench, and a terminal agent built from the same codebase. So, the interface you’re in doesn’t change what the agent knows. It

02:06 works over SSH and WSL for remote projects, and the CLI piece runs standalone if you don’t want electron near your terminal. Bongo Cat puts an animated cat on your desktop that reacts to your keyboard and mouse. Completely unnecessary, pretty charming. This native version uses C SDL3 and OpenGL with support for custom models. It runs on Windows, macOS, and Linux, and the maintainers say your input stays local rather than being recorded or uploaded.

02:36 Your desktop gets a tiny coworker who takes typing far more seriously than you do. Lya MLX handles small decisions that don’t need a chatbot response. Give it a support message and candidate departments, and it returns probabilities for each choice without generating text. This MLX port runs locally on Apple silicon without PyTorch or a cloud API. On an M3 Max, the maintainers report a 13.4 ms median for a short short English decision. MiniAGI is a byte-level language model

03:08 you train from scratch on an 8GB GPU. It reads raw bytes with no tokenizer and pages its experts in from disk 32 at a time. It keeps learning as you feed it new data, and the author reports 99.84% retention of what it learned earlier. The author calls it a toy model despite the name. Continual learning on a laptop GPU is the interesting part. Lya CoreML brings Lya’s typed decisions to Apple’s neural engine. Give it a short question, and it returns

03:39 probabilities for choices or scores without generating a paragraph. On an M3 Max, the maintainer reports a 4.98 ms median for one multilingual decision, and a 2.78-fold improvement in system energy per decision over compiled MLX. Bespoke Nimble is a Qwen 3.5 to 9B fine-tune that answers typed questions about text with no reasoning text. You hand it a schema of multiple choice or yes or no fields and it scores only the candidate answers, returning JSON with a

04:12 probability for each option. On their 324 example test set, the team reports 90.12% agreement up from 66.36 for the base model. Vmodel’s Swift SDK lets iOS and macOS apps search for moments inside video. Upload a clip, then look for a scene, spoken words, or text that appeared on screen. It handles streamed uploads, progress, cancellation, and typed search results with timestamps so your app can jump to the matching moment. Vmodel’s

04:43 back end does the indexing and search and you’ll need an API key. Magpie is a menu bar app that picks the model for all your coding agents from one place. It covers nine of them, including Claude Code, Codex, Gemini CLI, and Cursor, and rewrites each config file in place without wrecking your comments or indentation. A local gateway speaks OpenAI’s and Anthropic’s APIs and forwards to whichever vendor you chose. Profiles snapshot every agent’s settings so you can switch the whole setup at once.

05:16 Null Motion plays a finished ad film at full width and runs black and white motion drafts underneath it, frame synced, so you can see how each section was built. FFmpeg scene detection splits splits the reference to three to eight sections and each draft is HTML and GSAP on a paused timeline driven by the film’s clock. Export renders the whole thing to MP4 in the browser with web codecs. It’s a tear-down view for motion design and it runs entirely local. Shapeshift starts with one text box,

05:48 then turns what you type into the right little tool. Dinner with Priya Friday becomes an event card. A shopping list becomes a checklist. Split 2400 between three calculates each share. Jev can identify the intent while ordinary code handles the dates and math. It also works offline with a built-in classifier and saved cards stay in your browser. Rizo Windows 8 is a collection of procedural films, each living in a single HTML file. One follows a night

06:18 train, another turns a flock of starlings into moving risograph ink. Canvas draws the frames, web audio handles the sound, and the repo includes Claude code skills and rendering tools for making your own. Every frame depends on time alone, so you can jump to any moment and inspect it exactly. Flute turns the React components you already have into cinematic 3D product shots. It installs into a Next.js or Vite app, and your coding agent writes JSON scenes that point at your real UI

06:50 with camera moves and depth of field layered on top. A built-in studio previews each scene and exports MP4 at up to 120 frames per second. Launch videos built from your actual interface instead of screenshots are the part that stands out. ArcCUA lets a computer use agent hand off the click-by-click part of a task. A planner sets the goal, allowed text, constraints, and success check. JV chooses actions from controls the Mac desktop exposes. It uses accessibility

07:21 data and local OCR, then waits for the UI to settle before the next move. If it gets stuck or hits its action budget, it hands control back to the planner. FormMerge catches the merge conflict before anyone writes the code. When parallel coding agents publish what they plan to touch, it compares those intents. So, if one agent plans to replace payment service, while another plans to extend it, you hear about it now instead of at merge time. A verification gate runs your tests on a

07:52 clean commit before anything lands. Browser agents sometimes try to click buttons that aren’t there. Fast browser use scans the controls actually visible on a page, then asks a local model to choose from those options. It can navigate sites and fill forms as a skill for Codex or Claude code, and leaves a trace you can check afterward. It runs on your machine, though the model needs substantial memory. Fluid use uses a surprisingly small AI

08:22 model for a familiar chore, filling out forms. On a Mac, it reads fields in a real app or browser, matches them to details you provide, and types the answers in. The decisions run on your device, and the repo reports about 1 millisecond per field. You handle anything that needs judgment, like essays, uploads, and clicking submit. Mr. Mac Workspace is a Windows desktop app that puts Codex CLI or Claude code in one window and your project files in the other. The right side has folders

08:52 for your inbox, projects, knowledge, and processes with a markdown editor and preview built in. You can pin chats, color code tabs, and see which agent is still working. It ships with 14 project skills, including Blender workflows. AI manager is a desktop app that keeps your AI coding CLIs in one place, forked from CC Switch. It detects 10 tools, including Claude code, Codex, Gemini CLI, and Open Code, and installs, updates, or repairs them across NPM, PNPM, Bun, or Volta. From the same

09:24 window, you point each tool at an API endpoint and test it, add MCP servers, install skills, and search old sessions to pick them back up. Astro Flash Orchestrator splits a coding task between two models. Astra plans the work and reviews the finished patch. A deep seek flash sub agent handles implementation, tests, and routine fixes. The idea is to spend Astra’s attention on decisions while flash does the longer build loop. It includes phase task briefs, verification steps, and reversible setup.

09:56 Jev router decides which tool your agent should reach for and stops there. It gathers models, sub agents, skills, MCP tools, and CLIs into one candidate list, asks type safe’s Jev a type choice question, then filters the answer through permission, availability, and risk checks. It never runs anything itself, and every decision goes into an append-only log. Isis one skills pack is three Claude skills for explaining work, and they pick between themselves based on what you want to end up holding.

10:28 UI 56 sync gives a short plain language answer in chat and never writes a file. The summary skill writes one self-contained HTML page with modes you can stack, accessible for ADHD or dyslexia, artistic, and animated. It reads the actual files before it summarizes instead of working from memory. Thinking Orbs is a Swift UI port of Jaco Bentolila’s Thinking Orbs, animated loading indicators for AI interfaces. You drop in one line, like thinking orb searching, and get a dotted 3D sphere

11:00 for that agent state, with nine to choose from. They pause off screen and respect reduced motion. The port was checked against the original TypeScript engine across 725,149 frames. Jev voice browser lets you drive a Chromium window by talking to it, and it often acts before you finish the sentence. Partial transcripts go to type safe’s Jev with about a dozen typed questions, and it picks the intent and the target instead of generating text. So, URLs and

11:30 search terms come through verbatim. Say no, the other one, and it corrects itself. The author reports about 300 milliseconds from your last word to a decision. Reflex is a local decision model for jobs where you need a choice, not a paragraph. Give it a support ticket, document, or photo and several fixed answer questions. It returns probabilities in one model pass. It asks each question with the options in two different orders to reduce position bias.

12:01 There’s a Jev compatible API, but you should calibrate the confidence numbers on your own data before using them as thresholds. This Lya package runs Convey’s decision model from Node.js and TypeScript with no Python and no PyTorch. It uses ONNX runtime and the author says outputs match the Python reference to four decimal places. You pass a state, like a ticket or an email, plus typed questions and get back per option probabilities, a rubric score, or a calibrated yes or no.

12:33 A single PDF can contain several documents and sorting them by hand gets old fast. DocJev takes PDFs, Word files, or slide decks, classifies them using rules you write, and splits combined packets into page ranges. It flags uncertain boundaries for review and can export the pieces as PDFs. Text extraction runs locally by default, but the extracted text goes to the hosted Jev service for decisions. Jevmin pulls an agent’s decisions out of its text and treats them as data.

13:04 Every call becomes a yes or no with confidence, a pick from a list, or a score on a scale, and code-based rules decide whether the agent gets to act on it. Each decision lands in a hash-chained ledger before the outcome is known, gets graded afterward, and feeds back into calibration. Logging the guess before the result is the part worth copying. DDC is an Android decompiler written in Rust going after the job JADX does. The author reports a 226 megabyte APK

13:34 decompiled in 5 seconds and says all 909,689 output files across seven real APKs parse cleanly under javac. You don’t always need a full decompile either. Over 20 subcommands answer questions like strings, cross-references, and class hierarchies in milliseconds. Jeff is a self-hosted stand-in for Type-Safe’s Jev API. Point the official SDK at your own server and it answers choice, score, and yes or no questions using a 400 million parameter Glyformer model.

14:06 The maintainer estimates about 2.6 per million single question requests on an L4 against 15.6 for Jev. Their benchmark shows the trade-off. Jeff trails Jev on harder reasoning tasks. Herder GPUI gives Herder a native Mac window for its terminal sessions, split panes, workspaces, and agent activity. It connects to a Herder daemon you install separately, then draws the terminal itself using Rust and GPUI. The useful bit is that the daemon owns the sessions. Close the window and your work

14:38 keeps running until you attach again. Glide is a Rust lossless compressor built for logs, data dumps, and version backups. Its record mode groups fields from structured data before compressing them. Its base mode stores a new version against an older one while a companion store can find that older match automatically. The appeal is smaller files and fast parallel reads. The trade-off is slower compression, especially when record mode has to parse the data first.