Top Open-Source GitHub Projects : VoxCPM, ElyOffice, pandacode, Snitch, PageIndex & net-merge #304

Author: ManuAGI - AutoGPT Tutorials · 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: Top Open-Source GitHub Projects : VoxCPM, ElyOffice, pandacode, Snitch, PageIndex & net-merge #304 · Channel: ManuAGI - AutoGPT Tutorials · 15:12 · Watch on YouTube

Topics: AI agents, Software engineering with AI, Local and open-source models, Open source and software ecosystems, Privacy and data protection

Summary

This weekly roundup introduces 20 open-source GitHub projects spanning speech generation, Android development, coding agents, document search, office software, security, networking, and memory systems. The featured tools include agent memory and coordination platforms, local-first apps, and utilities for faster inference and Rust development. Several projects emphasize privacy or offline operation, while others help agents work with visual interfaces, documents, or audio. The video closes with a brief invitation to try the projects and watch the next update.

Key points

VoxCPM is a tokenizer-free text-to-speech model supporting 30 languages, voice design, and voice cloning.
Android Skills provides official Android development guidance for coding agents, while Cursor Plugins extends Cursor with skills, rules, and integrations.
Hindsight and High Memory offer long-term memory capabilities for agents, with High Memory also tracking code diffs and experience.
PageIndex uses hierarchical document trees and LLM reasoning for vectorless retrieval from complex documents.
ElyOffice is an offline-first Rust office suite that keeps files locally and bundles several document and productivity tools.
Panda Code brings multiple coding agents, project workspaces, Git, and approval prompts into one desktop application.
Zero Breach and Snitch provide local security tools for checking spyware indicators and visualizing network connections.
Net Merge distributes new connections across multiple network adapters to make use of Wi-Fi, Ethernet, and tethered connections.
Earshot lets coding agents analyze, evaluate, synthesize, and refine audio using a CPU-only toolkit.
Other projects include tools for virtual screens, Mermaid diagrams, Android TV cleanup, Rust hot reload, and Apple screen sharing.

Chapters

Start	Chapter	Summary
00:00	Introduction and channel announcement	The host introduces the weekly GitHub project roundup and announces a new channel focused on AI agents.
00:31	AI, developer, and document tools	The roundup covers VoxCPM, Android Skills, Cursor Plugins, Hindsight, PageIndex, Yamado, Mermaid renderers, ASCII Sharks, and Eagle 4.
06:24	Office, security, and coding-agent projects	The host presents ElyOffice, Zero Breach, TV Detox, Panda Code, and Undine.
10:20	Rust iteration and network tools	The roundup features Rjit, Snitch, Open Screen Sharing, and Earshot.
13:33	Networking and agent memory	The final projects are Net Merge, for using multiple network connections, and High Memory, for persistent agent memory.

Quotes

It outputs 48 kHz audio and runs through a Python API, CLI, or web demo locally on your GPU. — 01:04

Retain stores facts, recall searches them, and reflect reasons over them. — 03:14

Keeping files in local folders with no account or cloud service. — 06:56

State lives in the host process and survives each swap. — 10:50

Net merge is a free open source go app that uses all of a computer’s network adapters at once — 13:33

Concepts and entities

VoxCPM, Android skills, Cursor plugins, Hindsight, PageIndex, ElyOffice, Panda Code, Snitch, Earshot, NetMerge

Full transcript (original language)

00:00 Every week, developers release new open-source projects on GitHub, and this is our weekly GitHub project update video covering the best of them. Top trending open-source GitHub projects this week, you’ll discover useful and trending developer tools from local office suites to AI agent memory. Without wasting time, let’s get started. >> Before we jump into today’s project updates, here’s a quick announcement for everyone. We’ve launched a brand new YouTube channel called AI Agent Studio, dedicated entirely to AI agent projects,

00:31 tutorials, and tools. So, if you’re interested in staying up to date with the latest AI agent open-source projects, learning how to build your own agents, or exploring cutting-edge agent frameworks, make sure to check it out. Subscribe now to get weekly videos, in-depth guides, and real-time project breakdowns. The link is right there in the description. Don’t miss it. All right, let’s get into today’s video. >> Project number one, Vax CPM, tokenizer-free multilingual speech with voice design and cloning. Vax CPM is an

01:04 open-source Python text-to-speech model that generates speech without a tokenizer using a diffusion autoregressive design. It solves the need for expressive multilingual voices by supporting 30 languages, voice design from a text description, and cloning from a reference clip. It outputs 48 kHz audio and runs through a Python API, CLI, or web demo locally on your GPU. This suits developers building voice apps. The project forbids impersonation and fraud. Install it with pip and generate speech. Project number two,

01:36 skills, official Android guidance packaged for coding agents. Android skills is an open-source library of modular instructions that help coding agents follow Android development best practices. It solves the problem of LLMs underperforming on specific Android workflows by supplying skill.md files with official guidance. Developers install skills through the Android CLI or as a Claude code or Codex plugin, and agents activate them automatically. This suits Android developers using coding agents. Install a skill and ask your

02:09 agent to make your app edge-to-edge. Project number three, Cursor plugins. The official plugin marketplace for Cursor’s coding agent. Cursor plugins is an open-source marketplace repository holding the Cursor plugin specification and official plugins. It solves the need to extend a coding agent by packaging skills, rules, and MCP server definitions into standalone directories, each with its own manifest. Plugins cover developer workflows like code review and parallel agents, plus integrations such as GitHub, Playwright,

02:41 Gmail, and Salesforce. This suits developers customizing Cursor or building their own plugins. Browse the directory and try one in your workflow. Project number four, Vine’s Hindsight. Agent memory that learns instead of only recalling. Hindsight is an open-source agent memory system that helps agents learn over time, not only recall past chats. It exposes three operations, retain, recall, and reflect. Retain stores facts, recall searches them, and reflect reasons over them. It runs self-hosted through Docker, pip, or

03:14 Kubernetes with Postgres and connects to many LLM providers and coding agents through SDKs, MCP, and an LLM wrapper. This suits developers building long-running agents. Start the Docker server and retain your first memory. Project number five, Vinge page index. Vectorless rag that reasons through a document tree. Vecstemd is an open-source Python SDK for vectorless reasoning-based rag. It solves the limits of similarity search on long, complex documents by building a hierarchical tree index, then letting an

03:46 LLM reason through that tree to find the relevant sections with no vector database or chunking. It runs locally with your own LLM key or through page index cloud for scanned documents. This suits developers building document question answering for reports and filings. Install the package and ask your question. Project number six, Yamado virtual screens on a Mac built for AI agents to work in. Yamado is a native macOS app that creates virtual displays for AI agents doing visual computer use where an agent reads

04:17 screenshots and operates desktop apps directly with mouse and keyboard. It solves the lack of a dedicated workspace for agent tasks by letting you drag an app’s windows into a named virtual screen separate from your own with no dummy plug or paid display utility needed. An included skill lets a computer use agent create, name, and manage its own work screen automatically. It runs on Apple silicon and requests real mouse and keyboard control through a visible time-limited lease the person must approve. This suits developers running visual AI agents on apps with out an MCP

04:50 integration like Blender. Download the app and try the skill on your next visual task. Project number seven, Bonds Paley, a small fast library for rendering mermaid diagrams to SVG. Paley is an open-source JavaScript library that renders mermaid diagrams to SVG. It solves the weight of existing renderers by staying small, synchronous, and free of runtime dependencies supporting every diagram type in mermaid 12.1.0. Developers install it from NPM or a CDN,

05:21 theme it with CSS variables, and mount it directly into the page or load diagram types on demand. This suits developers in building diagrams in web apps or docs. Install the package and render your first diagram. Project number eight, Mask ASCII Sharks, swimming ASCII sharks rendered on an HTML canvas. ASCII Sharks is a dependency-free TypeScript library that draws fish and sharks as moving ASCII art on an HTML canvas. It solves the need for a lightweight playful visual with no image or animation libraries by

05:53 giving developers an ocean engine class that takes a canvas and a list of creatures like hammerheads and animates them directly as text characters. This suits developers wanting a fun text-based visual touch for a webpage. Install the package and spawn your first shark. Project number nine, Eagle 4, a drafter that speeds up LLM inference across any hardware. Eagle 4 is source available inference code that speeds up large language models through speculative decoding. It solves the mismatch between fixed draft strategies

06:24 and changing hardware by measuring costs on your setup and adapting how far to speculate for each batch size and workload. The Eagle harness package plugs into SG Lang, vLLM, and Token Speed serving models with Eagle 4 drafts while keeping identical output to the target model. This suits teams serving large language models who want faster generation. Follow the quick start to serve your first model. Project number 10, Ili Office, an offline office suite built natively in Rust. Ili Office is an open-source suite of offline-first

06:56 desktop apps built in Rust and GPUUI. Keeping files in local folders with no account or cloud service. It solves cloud dependency in office software by bundling a word processor, spreadsheet, presentation tool, PDF editor, Git client, and Markdown editor powered by LibreOffice Kit, MuPDF, and Tesseract. It currently builds on macOS with Linux and Windows packaging in progress. This suits users wanting a private local office suite. Build it from source and open your first document. Project number 11, Vinca Zero Breach, a local scanner

07:31 that checks for signs of ad-delivered spyware. Zero Breach is a read-only Python scanner that checks a computer for traces of the AdFlection and CandyRu spyware campaign. It solves the need for a self-service check by searching browser history, caches, auto start entries, and file fingerprints against 338 published indicators, rating each match by strength. It runs offline with no administrator rights and no network requests, and exports indicators as sticks and MISP feeds or DNS block lists. This suits security researchers

08:02 and people checking their own device. Download the script, verify its hash, and run a scan. Project number 12 TV Detox, a tiny ad-free home screen for a slow Android TV. TV Detox is an open-source 30 MB Android TV home screen paired with a prompt that has Claude code or Codex clean up an old TV over ADB. It solves TV bloat by having the coding agent identify safe to disable bloatware, turn off ads and sponsored rows, and install a calm home screen with just your apps, HDMI inputs, and a clock. Apps are disabled, never deleted,

08:38 and undo.sh restores everything. This suits owners of aging Android TV and Google TV devices. Turn on developer mode and paste the prompt into your agent. Project number 13 - Panda Code, a native desktop hub that runs every coding agent side by side. Panda Code is a free open-source Tory desktop workspace that runs coding CLIs like Claude code, Codex, Kimiko, and GitHub Copilot CLI side by side with projects, Git, a terminal, and a browser in one window. It solves the fragmentation of juggling separate agent tools by giving

09:12 each task a shared board, work trees, and approval prompts, and lets an architect coordinate multiple worker agents. It stores everything locally in SQLite with no telemetry or account, and can also run agents in local VMs or cloud sandboxes. This suits developers running several coding agents on real projects. Download it and open a project to start your first task. Project number 14 - Undine, mermaid diagrams compiled ahead of time in Rust. Ondine is an open-source Rust compiler that turns Mermaid diagram syntax into SVG at build

09:45 time instead of rendering it in the browser. It solves the overhead of shipping a full diagram parser by compiling diagrams ahead of time, so pages load precompiled SVG through a 210-byte runtime. It accepts the same syntax as Mermaid and works through a Vite plugin, a CLI, a Node compiler, a Markdown and MDX remark plugin, and optional in-browser WASM modules. This suits developers who embed diagrams in docs or web apps and want smaller, faster pages. Install the CLI or plugin and compile your first diagram. Project

10:20 number 15, Rjit, an experiment that hot reloads Rust code in milliseconds. Rjit is an experimental Rust hot reload tool that uses Rust analyzer as its compiler front end and Cranelift as its code generator, skipping LLVM and a linker entirely. It solves the slow edit-to-run cycle in Rust by translating Rust analyzer’s MIR into machine code on the fly, recompiling only the functions an edit actually changed, and repointing their function slots. State lives in the

10:50 host process and survives each swap. Measured on one machine, a one-file edit swaps in about 7 milliseconds against roughly 169 milliseconds for an incremental Cargo build. It supports only a small subset of Rust today. This suits Rust developers curious about faster iteration loops. Run it with Nix and edit the demo to watch it hot swap. Project number 16, Snash, a local visual map of every connection your computer makes. Snash is an open-source cross-platform network traffic visualizer that shows every connection a

11:23 computer makes as a live node graph and world map. It solves the gap between blunt firewalls and raw packet tools by explaining traffic visually, flagging anomalies like port scans or beaconing, and attributing each connection to the app responsible. It captures packets locally through libpcap, stores history in SQLite, and makes zero outbound calls with offline geolocation by default. This suits privacy-conscious users and security-minded developers. Download it for your platform and watch your traffic appear. Project number 17, Open Screen

11:57 Sharing, an improved client and server for Apple’s screen sharing protocol. Open Screen Sharing is an open-source project building better servers and clients for the private protocol behind macOS’s screen sharing and VNC. It solves the limits of Apple’s own app by optimizing for local and public networks, including smooth operation over Tailscale, and fixing issues like language switching. It runs on macOS and, through iOS’s developer debugging protocol, turns an iPhone or iPad into a VNC server, too, connectable from the

12:28 native macOS app. This suits Apple users wanting a more capable remote screen experience. Install the macOS or iOS build to try it. Project number 18, Earshot, ears for coding agents that hear, judge, and make sound. Earshot is an open-source CPU-only toolkit that lets AI coding agents listen to audio and react to it instead of only generating code blindly. It solves agents’ inability to judge their own sound work by describing what a file sounds like, matching it against a description, finding the best moment in

13:00 a recording, and scoring a film’s sound design against professional launch videos. Agents can also synthesize sounds entirely in code and refine them by listening, or convert a real recording into an editable code recipe. It runs as a CLI, Python API, or MCP server for Claude code. This suits developers building code-made videos, games, or products who need an agent that can audit and craft audio. Install it and run Earshot Describe on a sound to see it listen. Project number 19, net merge. Combine Wi-Fi, ethernet, and a

13:33 phone into one faster link. Net merge is a free open source go app that uses all of a computer’s network adapters at once, spreading new connections across Wi-Fi, ethernet, and USB or hotspot phone tethers. It solves the waste of an idle second connection by giving each new connection to one link. So, apps that open many connections, like downloads or games, get the combined speed. It works on Windows, macOS, and Linux through a tune interface and wintun, with no account, proxy setup, or server involved. This suits anyone with

14:06 more than one working internet connection who wants faster downloads. Install it with the one-line script and plug in a second connection to try it. Project number 20, Osh High Memory. A dual system memory engine with its own coding diff chain. High Memory is a Python SDK and engine that gives LLM agents long-term memory across three task types, chat, working, and coding. It solves shallow recall by organizing chat memory into layers of facts, summaries, and schemas, while coding and working tasks get a two-layer fact and

14:38 experience memory, plus a version evolution chain that stores real code diffs and agent summaries would otherwise lose. It runs with Chroma DB, Qdrant, or Faiss for vectors and Kuzo or Neo4j for graphs, connecting to any OpenAI compatible LLM. A companion desktop app, MemHub, unifies memory across tools like Claude Code, Cursor, and Codex locally. This suits developers building agents that need to remember facts, experiences, and code history over time. Install the package and write your first memory to try it. Thanks for

15:12 watching. See you in the next update.