AutoReels

Automated short-form video pipeline. One trigger in, an upload-ready 1080×1920 MP4 out. Fully local, fully open-source, zero ongoing cost.

What it does

Press a button. A few minutes later, a finished reel lands in your Telegram. The pipeline discovers a trending topic, researches it, writes a script with a local LLM, generates narration, fetches/generates visuals, assembles a captioned 9:16 video, and delivers it.

Target hardware

Apple M4, 16GB unified memory, macOS 15+. Sequential model execution keeps peak RAM under ~12GB.

Stack

• Server: FastAPI + Uvicorn on port 8420
• DB: SQLite (aiosqlite)
• LLM: Ollama (Llama 3.1 8B Q4)
• TTS: Piper / edge-tts / gtts (fallback chain)
• Timestamps: faster-whisper
• AI video: ComfyUI headless (AnimateDiff+SDXL, Wan 1.3B, LTX 2B) + MLX on Apple Silicon
• Stock: Pexels API (free tier)
• Assembly: MoviePy + ffmpeg + Pillow
• Delivery: Telegram Bot API

Pipeline

Trigger → Discover → Research → Script → Audio → Visuals → Assembly → Validate → Deliver.

Each stage writes output to disk before marking complete, so any run is resumable after a crash or kill. Only one run executes at a time.

Setup

./setup.sh

Installs Homebrew deps (python@3.11, ffmpeg), pulls the Llama model via Ollama, creates a venv, installs Python deps, and initializes the SQLite DB.

For AI video generation on Apple Silicon:

./setup_ai_video.sh

Then configure config.yaml with your Telegram bot token, chat ID, and Pexels API key.

Run

source .venv/bin/activate
python -m server.main

Dashboard at http://localhost:8420.

Trigger

curl -X POST http://localhost:8420/api/trigger \
  -H "Content-Type: application/json" \
  -d '{}'                              # auto-discover trending topic
# or
  -d '{"topic": "india australia critical minerals deal"}'

Works from an iPhone Shortcut on the local network too — same endpoint, POST the same JSON.

API

Method	Path	Description
POST	/api/trigger	Start a new run. Optional { "topic": "..." }.
GET	/api/runs	List all runs.
GET	/api/runs/:id	Run detail with per-step status.
POST	/api/runs/:id/kill	Kill active run.
POST	/api/runs/:id/resume	Resume from last completed step.
GET	/api/runs/:id/output	Download final MP4.
GET	/api/health	Health, model availability, disk space.

Layout

autoreels/
├── server/              FastAPI app, routes, DB, Telegram bot
├── pipeline/            Orchestrator + one module per stage
├── comfyui/workflows/   JSON workflow templates per video model
├── dashboard/           Single-page dashboard (index.html)
├── assets/              SFX, music, fonts, flags, pronunciation.json
├── scripts/             Asset + model setup helpers
├── output/<run_id>/     Per-run artifacts (script.json, narration.wav, segments/, final.mp4)
├── db/autoreels.db      SQLite DB
├── config.yaml          All tunable params
└── setup.sh             One-command setup

Design notes

Audio drives timing: narration is generated first, Whisper extracts word timestamps, visuals cut to match. Stock assets are the backbone; AI generation is the accent — used only when stock doesn't fit. Script quality is the bottleneck, so most prompt-engineering effort lives there.

Status

Work in progress. See AUTOREELS-PRD.md for the full spec.

License

Open source. All dependencies are free and open-source.