https://github.com/claude-world/claude-world-studio
AI-powered content pipeline: trend discovery → research → social publishing. Built with Claude Agent SDK + MCP.
https://github.com/claude-world/claude-world-studio
ai-agent claude claude-code content-pipeline mcp notebooklm social-media threads trend-analysis
Last synced: 2 months ago
JSON representation
AI-powered content pipeline: trend discovery → research → social publishing. Built with Claude Agent SDK + MCP.
- Host: GitHub
- URL: https://github.com/claude-world/claude-world-studio
- Owner: claude-world
- Created: 2026-03-15T05:16:18.000Z (3 months ago)
- Default Branch: main
- Last Pushed: 2026-03-26T14:29:42.000Z (2 months ago)
- Last Synced: 2026-03-27T13:34:03.429Z (2 months ago)
- Topics: ai-agent, claude, claude-code, content-pipeline, mcp, notebooklm, social-media, threads, trend-analysis
- Language: TypeScript
- Size: 9.51 MB
- Stars: 51
- Watchers: 0
- Forks: 8
- Open Issues: 10
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- Codeowners: .github/CODEOWNERS
- Security: SECURITY.md
Awesome Lists containing this project
README
# Claude World Studio
AI-powered content pipeline: from trend discovery to social publishing.
Built with **Claude Agent SDK + MCP** (Model Context Protocol), featuring 3 integrated MCP servers for real-time trends, web scraping, and research automation.
## Features
### One-Stop Pipeline
Three clear entry points — no complex menus, just pick and go:
| Card | Action | What happens |
| ------------------ | --------------------- | ------------------------------------------------------------------------------ |
| **Freestyle** | One click, zero input | Auto: trends → read source → verify timeline → patent score → visual → publish |
| **Custom Topic** | Type your topic | Deep research → score → visual → publish |
| **Custom + Media** | Type your topic | Research → NotebookLM slides/video/podcast → publish |
### Live Agent Execution
Watch Claude work in real-time — tool calls shown as collapsible blocks with MCP server badges:
- **trend-pulse** (green) — 20 real-time trend sources, zero auth
- **cf-browser** (blue) — Cloudflare Browser Rendering for JS pages
- **notebooklm** (purple) — Research + artifact generation (podcast/slides/video)
- **studio** — Built-in publishing + history tools
### Rich Markdown Responses
Full markdown rendering with syntax highlighting, clickable file path previews, cost/duration tracking, and inline images:
### Multi-Language (i18n)
Full support for Traditional Chinese, English, and Japanese — all UI, system prompts, and pipeline cards adapt:
English
Japanese
### Social Publishing (Native TypeScript)
Built-in Threads/Instagram publishing via native `fetch()` — no Python dependency. Every post is checked against Meta's patent-based 5-dimension scoring:
| Dimension | Patent | What it checks |
| ----------------------- | ----------- | ------------------------------------------------ |
| Hook Power | EdgeRank | First line has number or contrast? (10-45 chars) |
| Engagement Trigger | Dear Algo | CTA anyone can answer? |
| Conversation Durability | 72hr window | Has both sides / contrast? |
| Velocity Potential | Andromeda | Short enough? Timely? |
| Format Score | Multi-modal | Mobile-scannable? |
Quality gates: Overall >= 70, Conversation Durability >= 55.
**Supported post types**: text, image, video, carousel (2-20 items), poll, GIF, link preview, text attachment, spoiler (media blur + text), ghost (24hr ephemeral), quote post, reply control, topic tag, alt text, link-comment auto-reply.
### Multi-Account & Persona
- Multiple Threads/Instagram accounts with independent personas
- Per-account style (e.g., "tech-educator", "futurist")
- Per-account persona prompt for tone adaptation
- Matrix publishing: same topic → unique content per account
### Scheduled Tasks
Cron-based task scheduling with per-account targeting:
- Define prompt templates with cron schedules
- Target specific social accounts
- Quality gate enforcement (min score threshold)
- Auto-publish or manual review mode
- Execution history with cost/duration tracking
- Reduced tool set for unattended safety (no Bash/Read/Write)
### Visual Content Generation
NotebookLM integration for auto-generating visuals:
- **Image cards** — slides PDF as single-page visual cards
- **Carousel** — multi-page slides split into carousel images
- **Podcast** — AI-generated audio discussion
- **Video** — slides + podcast combined via ffmpeg
- **Mind maps, reports, flashcards, study guides**
All downloads stay within the workspace directory.
### Workspace File Browser
Browse and preview workspace files directly in the UI:
- Directory tree with depth control
- Text file preview (syntax-highlighted)
- Binary file preview (images, PDFs, audio, video)
- Clickable file paths in chat responses
- Workspace-contained — no access outside project folder
### Source Verification & Timeline Rules
Content integrity enforced at the system level:
- **Read original sources** — Never write from titles/metadata alone. 1+ primary source per topic, 2+ for controversial.
- **Timeline verification** — Every fact gets a verified timestamp. Time words mapped by age.
- **No AI filler** — System prompt blocks generic phrases ("in today's world" / "it's worth noting").
### Session Management
- Multiple concurrent sessions with independent workspaces
- Session resume on app restart (conversation history preserved)
- Idle session eviction (30min TTL)
- Query cancellation via interrupt button or CLI
### Dark Mode
System/light/dark theme with persistent preference across sessions.
### IME Support
Full CJK input method support — Enter during Chinese/Japanese character composition does not trigger send. Shift+Enter for newline.
### Workspace Security
- AI agent operates within session workspace only
- No access to ~/Downloads, ~/Desktop, ~/Documents, or system paths
- Credentials stored in local SQLite — never searched from filesystem
- Publishing via built-in MCP tool only (no external scripts)
- Session workspace validated at creation (blocks system paths)
- File API restricted to workspace root (symlink-safe realpath check)
## Architecture
```
┌─────────────┐ ┌──────────────────┐ ┌─────────────────┐
│ React UI │────▶│ Express + WS │────▶│ Claude Agent │
│ (Vite) │◀────│ Server │◀────│ SDK │
└─────────────┘ └──────────────────┘ └────────┬────────┘
│
┌──────────────────────────┤
│ │ │
┌──────▼──┐ ┌──────▼───┐ ┌─────▼──────┐
│ trend- │ │ cf- │ │ notebooklm │
│ pulse │ │ browser │ │ │
│ (MCP) │ │ (MCP) │ │ (MCP) │
└─────────┘ └──────────┘ └────────────┘
20 sources Cloudflare Podcast/Slides
zero auth Browser /Video/Report
```
## Tech Stack
| Layer | Tech |
| ----------- | --------------------------------------------- |
| Frontend | React 18 + Vite + Tailwind CSS |
| Backend | Express + WebSocket + Claude Agent SDK |
| AI Model | Claude Sonnet 4.6 / Opus 4.6 |
| MCP Servers | trend-pulse, cf-browser, notebooklm |
| Publishing | Native TypeScript fetch() → Threads Graph API |
| Database | SQLite (better-sqlite3, WAL mode) |
| Desktop | Electron 41 |
| Markdown | react-markdown + rehype-sanitize |
## Prerequisites
- **Node.js** >= 18
- **Claude Code CLI** installed and authenticated (`npm install -g @anthropic-ai/claude-code`)
- **Python** 3.10+ (for MCP servers — or use `uvx` for zero-config setup)
---
## Installation
Five ways to install — pick the one that fits your workflow:
| Method | Best for | MCP Servers | Web UI | Desktop App | CLI |
| ------------------------------------------------------- | ----------------- | ----------- | -------- | ----------- | --- |
| [A. Mac Desktop](#option-a-mac-desktop-app-download) | Quickest start | via uvx | Built-in | Yes | Yes |
| [B. npm](#option-b-npm-global) | Most users | via uvx | Yes | — | Yes |
| [C. Source](#option-c-from-source) | Contributors | venv or uvx | Yes | — | Yes |
| [D. Build Desktop](#option-d-build-desktop-from-source) | Custom builds | via uvx | Built-in | Yes | Yes |
| [E. MCP-only](#option-e-mcp-only-for-claude-code-cli) | Claude Code users | via uvx | — | — | — |
### Option A: Mac Desktop App (Download)
Download the latest `.dmg` from [GitHub Releases](https://github.com/claude-world/claude-world-studio/releases/latest), open it, and drag to Applications.
- Double-click to launch — server starts automatically
- No terminal, no Node.js install needed
- MCP servers auto-detected via uvx (install [uv](https://docs.astral.sh/uv/) first)
- All features included: Web UI + CLI + MCP + Publishing
> **Requires**: macOS (Apple Silicon). Intel Mac build available on request.
### Option B: npm global
```bash
npm install -g @claude-world/studio
# Set up MCP servers (one-time, auto-cached via uvx)
npx @claude-world/studio setup-mcp
# Start
studio serve
# Web UI: http://localhost:5173
# API: http://127.0.0.1:3001
```
After install, both `studio` and `claude-world-studio` commands are available globally.
### Option C: From source
```bash
git clone https://github.com/claude-world/claude-world-studio.git
cd claude-world-studio
npm install
cp .env.example .env
# Set up MCP servers
# Option 1: uvx (preferred — no clone, no path config)
npx @claude-world/studio setup-mcp
# Option 2: venv (legacy — clones repos, requires path config)
npx @claude-world/studio setup-mcp --venv
# Start development
npm run dev
# Frontend: http://localhost:5173
# Backend: http://127.0.0.1:3001
```
Use `node bin/cli.js ` or `npm link` to register the `studio` command.
### Option D: Build Desktop from source
Build the native macOS app yourself:
```bash
git clone https://github.com/claude-world/claude-world-studio.git
cd claude-world-studio
npm install
# Development mode (quick test)
npm run electron:dev
# Production build (creates .app + .dmg)
npm run electron:build
# Output: dist/mac-arm64/Claude World Studio.app
```
The Electron app:
- Spawns the Express server automatically on launch
- Loads your login shell PATH (nvm/homebrew compatible)
- Rebuilds native modules (better-sqlite3) for system Node ABI
- Supports all MCP servers (auto-detected via uvx or .env)
### Option E: MCP-only (for Claude Code CLI)
If you already use Claude Code CLI and just want the MCP tools (trend-pulse, cf-browser, notebooklm) without the Studio UI:
```bash
# Install MCP servers via uvx (one-time)
uvx --from 'trend-pulse[mcp]' trend-pulse-server --help
uvx --from cf-browser-mcp cf-browser-mcp --help
uvx --from notebooklm-skill notebooklm-mcp --help
```
Add to your project's `.mcp.json`:
```json
{
"mcpServers": {
"trend-pulse": {
"type": "stdio",
"command": "uvx",
"args": ["--from", "trend-pulse[mcp]", "trend-pulse-server"]
},
"cf-browser": {
"type": "stdio",
"command": "uvx",
"args": ["--from", "cf-browser-mcp", "cf-browser-mcp"]
},
"notebooklm": {
"type": "stdio",
"command": "uvx",
"args": ["--from", "notebooklm-skill", "notebooklm-mcp"]
}
}
}
```
Or add to Claude Desktop App config (`~/Library/Application Support/Claude/claude_desktop_config.json`) with the same format.
---
## MCP Server Setup
The app uses 3 MCP servers — **all optional**, the app works with any combination.
### Quick Setup (uvx — recommended)
```bash
npx @claude-world/studio setup-mcp
```
This pre-caches all 3 servers via uvx. No cloning, no path config. Studio auto-detects uvx at runtime.
To update: `npx @claude-world/studio setup-mcp --update`
### Manual Setup (venv)
See [detailed MCP setup guide](docs/content-pipeline-skill.md#mcp-servers-setup) for per-server installation with local venvs.
### MCP Auto-Detection
Studio resolves MCP servers in this order:
1. **Settings UI** — paths configured via Settings page (stored in SQLite)
2. **Environment variables** — `.env` file or exported vars
3. **uvx fallback** — if `uvx` is on PATH, uses cached packages automatically
Use `studio settings detect` (CLI) or **Settings > Scan System** (UI) to check what's available.
---
## Production Pipeline: End-to-End
### Step 1: Install & Configure
```bash
npm install -g @claude-world/studio
npx @claude-world/studio setup-mcp
studio serve
```
### Step 2: Add Social Accounts
Via UI: **Settings > Social Accounts**
Via CLI:
```bash
studio account create \
--name "My Brand" --handle "@mybrand" \
--platform threads --token YOUR_TOKEN
```
### Step 3: Run the Pipeline
**Web UI**: Click a pipeline card (Freestyle / Custom Topic / Custom + Media)
**CLI (headless)**:
```bash
studio chat --message "Find trending topics, research the best, write and publish" --json
```
**Scheduled**: Settings > Scheduled Tasks — cron + per-account targeting + quality gates
### Step 4: Monitor
```bash
studio status # Server health
studio session list --json # Sessions
studio history --limit 20 --json # Publish history
```
---
## CLI
Full CLI with 23 commands. All commands support `--json` for programmatic use.
```bash
# Server
studio serve # Start web UI
studio status # Check if running
# Sessions
studio session list
studio session create --title "Research" --workspace /path
studio session get
studio session rename --title "New"
studio session delete
studio session messages --limit 20
# Chat (WebSocket streaming)
studio chat --message "Find trends" --json
studio chat --session --message "Publish the best"
echo "What's trending?" | studio chat --json
studio interrupt
# Accounts
studio account list
studio account create --name "Main" --handle "@me" --platform threads --token TOKEN
studio account update --token NEW_TOKEN
studio account delete
# Settings
studio settings get
studio settings detect # Auto-detect MCP tools
studio settings apply # Apply detected values
studio settings set --language en
# Publishing (all Threads post types)
studio publish --account --text "Hello!" --score 85
studio publish --account --text "Vote!" --poll "A|B|C"
studio publish --account --text "Look!" --image URL
studio publish --account --text "Watch" --video URL
studio publish --account --text "Slides" --carousel URL1 URL2 URL3
studio publish --account --text "Post" --link-comment URL
studio history --limit 10
# Files
studio file list --depth 2
studio file read src/index.ts
```
Global flags: `--json`, `--port N` (env: `STUDIO_PORT`), `--host H` (env: `STUDIO_HOST`)
MCP setup wizard: `npx @claude-world/studio setup-mcp`
---
## Security
This is a **local-only development tool**. It runs Claude with full tool access on your machine.
- Binds to `127.0.0.1` only (not exposed to network)
- WebSocket origin verification (exact port whitelist)
- CORS restricted to localhost dev ports
- Workspace containment: AI agent restricted to session workspace
- File API: realpath + workspace-only containment (symlink-safe)
- Path traversal rejected on both client and server
- Session workspace validated at creation (system paths blocked)
- Publishing via native fetch — no Python/shell dependency
- XSS protection via `rehype-sanitize`
- Session isolation: WS messages filtered by sessionId
- Idle session eviction (30min TTL)
- Credentials stored in local SQLite only
> **Warning**: Do NOT expose this server to the internet. The AI agent has `Bash`, `Read`, `Write`, and `Edit` access within the workspace.
## Contributing
Issues and PRs welcome. Please open an issue first to discuss major changes.
## License
MIT