DraftFlow

A self-hosted, AI-powered social media scheduling platform. Create, refine, and publish posts to LinkedIn and Twitter with AI assistance, trend tracking, and a weekly digest feature.

Features

• Multi-Platform Publishing — Schedule and publish posts to LinkedIn and Twitter/X
• Idea Board — Capture ideas and convert them into polished posts with AI
• AI Writing Assistant — Generate, improve, and iterate on content using your choice of models through OpenRouter (Anthropic / Gemini / DeepSeek / Kimi / Qwen, etc)
• Trending Topics — Discover what's trending in your industry with web search (Tavily)
• Carousel Builder — Create multi-slide PDF carousels for LinkedIn with AI-generated content and templates
• Inspiration Feed — Curated swipe file aggregating bookmarked articles, top posts, and wiki knowledge
• Visual Builder — Turn any post into a styled infographic image with AI-generated templates
• RSS Feeds — Subscribe to RSS feeds, browse articles, and turn them into posts or ideas
• Weekly Digest — Auto-curate the biggest stories of the week into a short-form post
• Calendar View — Visualize your content schedule at a glance
• Recurring Posts — Set up ideas that auto-generate on a daily, weekly, or monthly schedule
• Multi-Tenant — Support for multiple workspaces and team collaboration
• Markdown Support — Write in **bold** and *italic*, auto-converted to Unicode for LinkedIn
• LLM Wiki — Persistent, compounding knowledge base inspired by Karpathy's LLM Wiki
• AI Agent Pipeline — Mastra supervisor agent orchestrates researcher, writer, and editor agents
• MCP Integration — Extend AI capabilities with Model Context Protocol servers

LLM Wiki — Persistent Knowledge Base

Inspired by Andrej Karpathy's LLM Wiki concept, DraftFlow maintains a compounding knowledge base per workspace that grows smarter with every ingested source.

The problem with typical AI content tools: They start from scratch every time — re-discovering context by searching databases and the web, producing shallow, inconsistent output.

The LLM Wiki approach: Instead of ephemeral retrieval (RAG), DraftFlow maintains structured markdown wiki pages that the AI reads, writes, and updates over time. Feed it a URL, article, or RSS item and the AI extracts knowledge into focused wiki pages, cross-references with existing pages, and maintains an auto-generated index.

wiki/{tenantId}/
  index.md        ← Auto-maintained page catalog
  log.md          ← Chronological activity log
  serverless.md   ← Knowledge: serverless architecture patterns
  ai-agents.md    ← Knowledge: AI agent frameworks
  ...

Three operations:

• Ingest — Feed a URL or text. AI extracts knowledge into 1-5 wiki pages.
• Query — Keyword search across all pages. The AI agent checks the wiki before web search.
• Browse/Edit — Full CRUD UI to read, edit, and manage wiki pages.

How it helps content generation:

• The researcher agent queries the wiki first, using accumulated domain knowledge instead of starting from scratch
• "LLM Wiki" is a content source option alongside trends, ideas, and case studies
• RSS articles can be ingested into the wiki, turning news streams into reusable knowledge
• Existing content (published posts, trends, ideas) can be bulk-hydrated via npx ts-node scripts/hydrate-wiki.ts

As Karpathy puts it: "The tedious part of maintaining a knowledge base is not the reading or the thinking — it's the bookkeeping." The LLM handles the bookkeeping.

Visual Builder

Generate eye-catching infographic images from your post content. Choose from 6 templates — Infographic, Before vs After, Checklist, Quote Card, Stats, and Step-by-Step — and the AI converts your text into a styled visual, ready to attach to your post.

Carousel Builder

Create LinkedIn carousel posts (multi-slide PDFs) from a standalone builder page. Enter a topic or paste content, pick a template, and let AI generate a slide deck rendered as a 1080x1080 PDF.

• 5 templates: Step Guide, Listicle, Tips, Checklist, Stats Deck
• Two generation modes: "Generate" (content only) and "Research & Generate" (Tavily web search + content)
• Branding: Add your name, handle, tagline, and CTA — persisted in settings
• Save & reuse: Save generated carousels for later use
• Post integration: "Create Post" attaches the PDF and navigates to the post editor
• Slide count: Choose 3, 5, 7, or 10 slides per carousel

Inspiration Feed

A curated swipe file that aggregates your best content sources into one browsable page:

• Bookmarked RSS articles — Articles you've bookmarked from the Feeds page
• Favorited posts — Your published posts marked with a star from the dashboard
• Wiki pages — Knowledge pages from your LLM Wiki

Filter by source tab, search across all content, and quickly turn any inspiration into a new post.

Architecture

frontend/          Next.js (React 19) dashboard
backend/           Express + TypeScript API server
  ├── src/
  │   ├── routes/       API endpoints
  │   ├── services/     Business logic (AI, Wiki, LinkedIn, Twitter, Scheduler)
  │   ├── middleware/   Auth middleware
  │   └── db.ts         Sequelize models (SQLite)
  ├── wiki/             LLM Wiki markdown files (per tenant)
  ├── uploads/          File uploads (images, visuals)
  ├── scripts/          Utility scripts (wiki hydration, etc.)
  └── migrations/       SQL migration files

Quick Start (Docker)

git clone https://github.com/AppGambitStudio/DraftFlow.git
cd draftflow

# Copy and configure environment
cp backend/.env.example backend/.env
# Edit backend/.env — set JWT_SECRET (required) and other keys

# Start everything
docker-compose up --build

• Frontend: http://localhost:5003
• Backend API: http://localhost:5002

Local Development

Prerequisites

• Node.js v18+
• npm

Backend

cd backend
npm install
cp .env.example .env
# Edit .env — JWT_SECRET is required
npm run dev

Frontend

cd frontend
npm install
npm run dev

Frontend runs on port 5003, backend on port 5002.

First Run

1. Open http://localhost:5003 and create an account (Sign Up)
2. Go to Settings and configure:
   • OpenRouter API Key — Required for AI features (openrouter.ai)
   • LinkedIn OAuth — Client ID & Secret for LinkedIn publishing
   • Twitter OAuth — Client ID & Secret for Twitter publishing
   • Tavily API Key — Optional, for web search / trending topics (tavily.com)

OAuth Callback URLs

When setting up OAuth apps on LinkedIn/Twitter developer portals:

• LinkedIn: http://localhost:5002/api/auth/linkedin/callback
• Twitter: http://localhost:5002/api/auth/twitter/callback

Database

DraftFlow uses SQLite via Sequelize ORM.

• New tables are auto-created on startup via sequelize.sync()
• Schema changes to existing tables require manual migrations

Running Migrations

Migrations are SQL files in backend/migrations/. Apply them with:

cd backend
sqlite3 dev.db < migrations/YYYYMMDD_HHMMSS_description.sql

Environment Variables

See backend/.env.example for all available variables.

Variable	Required	Description
JWT_SECRET	Yes	Secret key for JWT token signing
DATABASE_URL	No	SQLite path (default: file:./dev.db)
PORT	No	Backend port (default: 5002)
FRONTEND_URL	No	Frontend URL for invite links (default: http://localhost:3000)
LINKEDIN_CLIENT_ID	No	LinkedIn OAuth (can also set in Settings UI)
LINKEDIN_CLIENT_SECRET	No	LinkedIn OAuth secret
TWITTER_CLIENT_ID	No	Twitter OAuth (can also set in Settings UI)
TWITTER_CLIENT_SECRET	No	Twitter OAuth secret

Tech Stack

• Frontend: Next.js 16, React 19, Tailwind CSS, Recharts
• Backend: Express 5, TypeScript, Sequelize, SQLite
• AI: OpenRouter API (Claude, GPT, etc.), Tavily web search
• Agent Framework: Mastra.io with MCP support

License

ISC