eyeroll 0.2.0

AI eyes that roll through video footage — watch, understand, turn into code actions.

eyeroll

AI eyes that roll through video footage — watch, understand, act.

eyeroll is a collection of AI coding agent skills that analyze screen recordings, Loom videos, YouTube links, and screenshots, then turn them into code actions.

Skills

Skill	What it does
watch-video	Watch a video and produce structured notes
video-to-pr	Watch a bug video → diagnose → fix → raise PR
video-to-skill	Watch a tutorial/demo → generate a Claude Code skill

Install

# Install all skills
npx skills add mnvsk97/eyeroll

# Install a specific skill
npx skills add mnvsk97/eyeroll@video-to-pr
npx skills add mnvsk97/eyeroll@video-to-skill
npx skills add mnvsk97/eyeroll@watch-video

Setup

pip install eyeroll
brew install yt-dlp    # for URL downloads (Loom, YouTube)

Option A: Gemini (default, API-based)

eyeroll init           # or: export GEMINI_API_KEY=your-key

Get a free key at aistudio.google.com/apikey. Supports direct video upload, audio transcription. ~$0.15 per analysis.

Option B: Ollama + Qwen3-VL (local, private, free)

brew install ollama    # install Ollama
ollama serve           # start the server

No API key needed. The model is pulled automatically on first use (~6GB for qwen3-vl:8b). Runs entirely on your machine.

Usage

In Claude Code (via skills)

User: fix this bug: https://loom.com/share/abc123
      → video-to-pr skill activates, watches video, finds code, fixes, raises PR

User: watch this tutorial and create a skill from it: ./demo.mp4
      → video-to-skill skill activates, watches video, generates SKILL.md

User: look at this recording, what's going on?
      → watch-video skill activates, produces structured notes

Standalone CLI

# Gemini (default)
eyeroll watch https://loom.com/share/abc123
eyeroll watch ./bug.mp4 --context "checkout broken after PR #432"

# Ollama (local)
eyeroll watch ./bug.mp4 --backend ollama
eyeroll watch screenshot.png -b ollama -m qwen3-vl:2b

# Or set as default
export EYEROLL_BACKEND=ollama
eyeroll watch ./recording.mp4

Backends

Backend	Video	Audio	API Key	Cost	Best for
gemini	Direct upload	Yes	Required	~$0.15/video	Best quality, short videos
ollama	Frame-by-frame	No	None	Free	Privacy, offline, no API limits

Ollama models

Model	Size	Notes
qwen3-vl (default)	6.1GB	Best quality, needs 8GB+ RAM
qwen3-vl:2b	1.9GB	Lighter, works on 8GB machines
qwen3-vl:32b	21GB	Higher quality, needs 32GB+ RAM

How it works

Video (Loom / YouTube / local file / screenshot)
    ↓
eyeroll extracts: frames, audio, on-screen text
    ↓
Backend analyzes: Gemini Flash (API) or Qwen3-VL (local via Ollama)
    ↓
Structured notes → skill decides what to do next
    ↓
video-to-pr:    search codebase → fix → PR
video-to-skill: extract workflow → generate SKILL.md
watch-video:    return notes to the agent

Supported inputs

Input	Notes
Loom URLs	Requires yt-dlp
YouTube URLs	Requires yt-dlp
Local video files (.mp4, .webm, .mov)	Direct analysis
Screenshots (.png, .jpg, .gif)	Single-frame analysis
Any yt-dlp supported URL	1000+ sites

Requirements

• Python 3.11+
• ffmpeg (brew install ffmpeg)
• yt-dlp (brew install yt-dlp) — for URL downloads
• Gemini backend: Gemini API key (free)
• Ollama backend: Ollama installed and running

License

MIT