Your coding agent's new side gig: narrated explainer videos

What it does

You point your coding agent at the repo and tell it what to explain. The agent writes two files — script.md for the narration and scenes.js for the drawings — and explainroo turns them into an MP4: Kokoro reads the script aloud (28 voices, no account needed), Whisper stamps the timing of every word, and headless Chrome renders the frames, which ffmpeg stitches together. It also composes its own background music and ducking, and can build product demos where a fake mouse clicks through rebuilt app screens.

The interesting bit

An agent can’t watch the video it made, so explainroo gives it other senses: stills of each scene, a contact sheet of frames, a layout check for cut-off or overlapping text, and a speech check for mispronounced words. That feedback loop — agent renders, agent inspects, agent fixes — is the actual product; the video is just the output.

Key highlights

• Fully local: voice, timing and rendering all run on your machine; the only optional paid piece is AI illustrations via OpenRouter.
• Word-level sync: drawings can appear on specific words because Whisper knows exactly when each one is spoken.
• Five looks (paper, clean, chalk, blueprint, midnight) and sizes for YouTube, Shorts/TikTok, 4:5 and square, with captions that light up word by word on tall formats.
• A pace setting from 0.7 to 1.6 speeds up voice, pauses, animations and music together.
• Rough.js hand-drawn style plus 1,800 Lucide icons; scenes can also show charts, code or your own screenshots.

Caveats

• Works best with Claude Code and Opus 5.5, per the README — other agents are supported but less proven.
• Developed and tested on Linux; macOS and Windows are “should work” territory.
• First setup downloads ~400 MB of models, and you need Node 20.11+, ffmpeg and Chrome already in place.

Verdict

If you already run a coding agent and want cheap, private, scriptable explainer or demo videos, this is a genuinely clever use of the agent as director. If you want polished motion-graphics output, a canvas renderer with hand-drawn lines won’t fool anyone into thinking you hired a studio.