Skip to content

← The Lab

motlab: an open-source video-as-code engine that encodes inside Chrome, with an agent skill and a notes-on-frames viewer

A new MIT project from developer Stanyslaus Hary takes the code-to-video idea in a different direction. Each scene is a canvas page with one renderAt(t) function, frames are encoded inside Chrome with WebCodecs, and you review by clicking on frames and typing notes your agent works through. The creator's benchmark puts it 2 to 4 times as fast as Remotion and HyperFrames on one scene.

How it works. A motlab project is a project.json listing scene pages and audio files. Each scene is plain HTML with a canvas and a window.renderAt(t) function, so everything you see depends only on time and preview, stills and the final render all draw the same frames. motlab render opens headless Chrome for each scene and encodes frames with the browser's VideoEncoder inside the page. ffmpeg only joins the parts and adds the audio. Footage is decoded in the page with Mediabunny, so scenes can cut and draw over real video. No AI image or video model is involved.

The agent part. Setup installs two skills into your agent (Claude Code by default; also Codex, Cursor, OpenCode, Copilot, Gemini CLI, Antigravity, Hermes and others). You hand the agent a folder of references (logos, screenshots, footage, fonts, music, voice) and a brief, and it plans the shots, writes the scenes and opens the viewer. In the viewer you click a frame, drag a box over part of it, or mark a time range, then type a note. The agent reads all the notes from selection.json and works through them by number. Versions sit side by side, so a second cut doesn't overwrite the first. Export goes up to 4K MP4.

Asset kits. A second skill, motlab-assets, builds reusable kits: icons, illustrations, rigged 2D characters, transitions and 3D objects, output as transparent PNG plus SVG, or GLB/FBX/OBJ with PBR textures (Blender is optional for the 3D path). The point is that every video built from the kit shares one look.

The speed claim, with its caveats. In the creator's benchmark (one 1080p, 300-frame scene on a Windows laptop with an RTX 3060), motlab took 9.5 to 12.6 s, against 27.2 s for HyperFrames 0.9.52 and 30.7 s for Remotion 4.0.543. The README is upfront about the limits: it's one scene on one machine, macOS and Linux weren't measured, and with software-only Chrome at 4K HyperFrames came out faster. The bench/ folder is there if you want to run it yourself.

Requirements and status. Node 24+, Chrome or Chromium, and ffmpeg/ffprobe on PATH. Install with npm install -g @motlab/cli then npx @motlab/cli setup. The repo appeared on Nov 16, reached v0.3.0 on Nov 18 and had about 190 GitHub stars when we looked, so it's very early. The viewer UI is in English or Indonesian.

Signals: the creator's own repo and posts only. No curator, trending or official pickup yet, so it's a promising find, not HOT.

Skills & repos in our directory

More from the Lab

Topics: HyperFrames · Remotion · Motion graphics skills