Five Trending GitHub Projects: AI Agent Orchestration, Knowledge Compilation, and One-Click Video Download

Xiaoer VideoLab: One‑Click Video Download

The Chrome extension captures the current page URL, sends it to a local Python daemon at 127.0.0.1:7788, which invokes yt-dlp to download the video into ~/Downloads. All processing stays on the local machine, avoiding broad site‑data permissions. It leverages yt‑dlp’s 1800+ site parsers to support YouTube, Bilibili, TikTok, Vimeo, and others, outputting 1080p MP4 on macOS and Windows.

Munder Difflin: Visual Collaboration for Claude Code

Munder Difflin is a local multi‑agent desktop built with Electron, React, and TypeScript. It wraps Claude Code sessions into independent agents with long‑term memory and mailboxes, displayed on a 2D office desk. The architecture has three layers: a GOD orchestration agent, multiple Worker agents (each a Claude Code process), and a shared memory/mailbox layer. Agents communicate via the mailbox, while GOD routes tasks and arbitrates decisions only when needed.

The memory system uses Markdown files with semantic search, enabling millisecond‑level recall across sessions. The prototype supports macOS, Windows, and Linux.

Citadel: Project‑Management Hub for Claude Code and Codex

Citadel adds a project‑management layer to coding agents. It expands /do routing commands to parse natural‑language requests and dispatch them to the most suitable skill or workflow. Project memory is stored in a local .planning/ directory, preserving decisions and context across sessions.

It also supports a Fleet mode that splits large tasks into subtasks executed in parallel Git worktrees, along with lifecycle hooks, file‑protection policies, cost tracking, and telemetry. Installation is a single command in a Claude Code or Codex session: /do setup --express.

Hermes Workspace: Desktop‑Level Multi‑Agent Console

Hermes Workspace is a web‑based console for Hermes Agents, integrating chat, file management, terminal, memory browsing, skill management, and MCP configuration in a single UI. It supports real‑time SSE streaming, browsing/editing of agent memory, and over 2000 skills via the /mcp management page.

The platform offers a Conductor mode that automatically decomposes a high‑level task and distributes it to multiple Worker agents. A Swarm mode uses persistent tmux workers to scale unlimited Hermes Agents, with role presets, Kanban boards, and automated review gates. It is PWA‑enabled and can be accessed via Tailscale on any device.

LLM Wiki Compiler: Compile Scattered Notes into an Interactive Knowledge Base

Inspired by Andrej Karpathy’s “LLM Wiki”, the compiler ingests raw sources (Markdown, code, PDFs) and runs a two‑stage LLM pipeline that produces pages of type concept, entity, comparison, and overview. Each paragraph is annotated with line‑range references to the original files.

Retrieval combines semantic embeddings (incremental updates with hash‑based deduplication) to narrow candidates, followed by BM25 re‑ranking and a wiki‑link graph to assemble the final evidence package.

The tool includes a local web viewer, force‑directed knowledge‑graph visualization, an MCP server compatible with Claude Desktop or Cursor, and an evaluation command llmwiki eval that scores knowledge‑base health, citation coverage, and regression delta.

Installation: npm install -g llm-wiki-compiler; after configuring an API key, run llmwiki quickstart ./notes.md. It supports Anthropic, OpenAI, Ollama, GitHub Copilot, and other model back‑ends, making it suitable for developers who need to organize research notes, paper reviews, or project documentation.

Xiaoer VideoLab — https://github.com/Jane-xiaoer/xiaoer-videolab ⭐374

Munder Difflin — https://github.com/chaitanyagiri/munder-difflin ⭐331

Citadel — https://github.com/SethGammon/Citadel ⭐589

Hermes Workspace — https://github.com/outsourc-e/hermes-workspace ⭐5,416

LLM Wiki Compiler — https://github.com/atomicstrata/llm-wiki-compiler ⭐1,456