Reelect

Your saved Instagram Reels — transcribed, structured, and actually useful. Runs entirely on your machine. No cloud. No API keys. No cost per request. Works with any OpenAI-compatible LLM provider (LM Studio, OpenRouter, Ollama, etc.).

---

The problem

You save a reel with a recipe, a useful app, a film recommendation. Three days later it's buried somewhere in your saved feed with no way to find it. Instagram's saved section has no search, no categories, no text — just an endless scroll of thumbnails.

Reelect fixes that.

---

The main idea — Actionable content extraction

Most reels worth saving contain something concrete: a recipe, a step-by-step guide, a film to watch, a tool to try. Reelect listens to the audio, looks at the frames, and extracts that information into structured text you can actually read and use — without opening Instagram at all.

🍳 recipe      → ingredients + step-by-step instructions
📋 guide       → numbered steps, tips, how-to
🎬 recommendation → film / series / book + why it's worth it
🔗 resource    → app / website / tool + what it does

If a reel contains nothing actionable — just entertainment or opinion — it still gets a summary and tags, but no actionable block is generated.

Example

A reel where someone talks about making a burnt basque cheesecake becomes:

{
  "summary": "Рецепт сан-себастьяна без миксера за 40 минут.",
  "category": "cooking",
  "tags": ["cheesecake", "выпечка", "рецепт", "десерт"],
  "actionable": {
    "type": "recipe",
    "title": "Сан-Себастьян чизкейк",
    "content": "Ингредиенты: 600г сливочного сыра, 5 яиц, 200г сахара, 300мл сливок\n\n1. Смешать все ингредиенты венчиком\n2. Вылить в форму, застеленную бумагой\n3. Выпекать 30-40 мин при 220°C"
  }
}

The card in the viewer shows the recipe directly — no need to watch the video.

---

How it works

Instagram Saved or Direct Reel URL
      ↓
  reelect_pipeline.cli
      ↓
  download         →  saved_videos/raw/.../<id>.mp4
                    +  <id>.manifest.json
                    +  <id>.transcript.txt
                    +  <id>.frames/frame_001.jpg
      ↓
  analyze          →  saved_videos/meta/<id>.json
                       { summary, category, tags, actionable }
      ↓
  viewer           →  http://localhost:8000

1. Download — fetches only new videos from your Instagram Saved via gallery-dl
2. Transcribe — audio extracted with ffmpeg, transcribed locally with faster-whisper
3. Analyze — frames sampled and sent alongside the transcript to a local vision LLM via LM Studio
4. Extract — the LLM produces summary, category, tags, and actionable content in one pass
5. View — clean web UI with search, category filter, and live pipeline control

Everything runs on your machine. Audio is transcribed by Whisper locally. Video frames are analyzed by a local LLM via LM Studio. No data leaves your computer, no API keys required, no cost per video — just your hardware doing the work.

---

Getting started

1. Prerequisites

• Docker & Docker Compose
• LM Studio with a vision-capable model loaded, server running on port 1234
• A cookies.txt exported from your browser while logged into Instagram
  • Install Cookie-Editor, open instagram.com, export in Netscape format

Choosing a model

Model selection is the most hardware-dependent part of the setup — there's no universal recommendation. A few guidelines:

Hardware	What works
Apple Silicon (M1–M4)	Qwen2.5-VL-7B, Qwen3-8B, Gemma-3-12B
NVIDIA GPU (8 GB VRAM)	Qwen2.5-VL-7B-Instruct, Llama-3.2-11B-Vision
NVIDIA GPU (16 GB+ VRAM)	Qwen2.5-VL-32B, Llama-3.2-90B-Vision
CPU only	Keep it small: Qwen2.5-VL-3B or 7B with aggressive quantization (Q4)

What to look for:

• The model must support vision input (image frames from the video)
• Bigger isn't always better — a well-quantized 7B often outperforms a bloated 32B that barely fits in VRAM
• If the model supports thinking/reasoning (Qwen3, DeepSeek-R1), set LM_STUDIO_THINKING_BUDGET=512 to cap reasoning tokens and avoid the model spending all its token budget on thinking instead of the JSON output
• Set LM_STUDIO_MAX_TOKENS_METADATA high enough (8192+) for reasoning models — they consume tokens for thinking before writing the response

Practical tip: start with the smallest vision model that runs at a comfortable speed on your machine, then upgrade if the output quality isn't good enough.

2. Install

git clone https://github.com/your-username/reelect.git
cd reelect
cp .env.example .env

3. Configure .env

INSTAGRAM_USERNAME=your_handle

# For LM Studio (local):
LLM_BASE_URL=http://host.docker.internal:1234/v1
LLM_MODEL=qwen2.5-vl-7b-instruct

# For OpenRouter (cloud):
# LLM_BASE_URL=https://openrouter.ai/api/v1
# LLM_MODEL=qwen/qwen2.5-vl-7b-instruct
# LLM_API_KEY=sk-or-...

CRON_SCHEDULE="0 */12 * * *"

See .env.example for all options.

4. Run

The first launch may take a while if you've saved many reels :)

docker compose build
docker compose up -d

Open http://localhost:8000.

The pipeline runs on schedule automatically. You can also trigger it manually from the UI with the ▶ Run pipeline button.

---

Configuration reference

Variable	Default	Description
INSTAGRAM_USERNAME	—	Your Instagram handle
CRON_SCHEDULE	0 */12 * * *	Pipeline run schedule
LLM_BASE_URL	http://host.docker.internal:1234/v1	OpenAI-compatible endpoint (LM Studio, OpenRouter, Ollama, etc.)
LLM_MODEL	—	Model name (must match provider exactly)
LLM_API_KEY	—	Optional API key (only if provider requires auth)
LLM_NATIVE_VIDEO	false	Send full video to model instead of frames
LLM_THINKING_BUDGET	512	Reasoning token limit for thinking models
LLM_MAX_TOKENS_VISUAL	4096	Max tokens for visual description
LLM_MAX_TOKENS_METADATA	8192	Max tokens for metadata + actionable
MAX_WORKERS	3	Parallel whisper + ffmpeg workers
LLM_CONCURRENCY	1	Concurrent LLM requests

---

Notes

• cookies.txt and .env are gitignored and never committed.
• The Whisper model is baked into the Docker image — transcription is fully offline.