He Replaced After Effects With 16 GitHub Repositories—and the Video Is Still Rendered by Code

Someone removed After Effects from the motion design workflow.

eigen listed 16 open-source repositories on X and said he now uses Claude Opus 5.5 plus these projects for Motion, VFX, 3D, Sound, and export. He calls this setup a motion studio, and all the repositories are open source.

The video looks like a normal motion piece. Buttons morph, interfaces switch, loaders spin, sound effects hit on cue. The difference is how it is made. The video is computed frame by frame by code, and generative video models are not involved.

On September 22, 2026, Anthropic released Claude Opus 5.5. After that, a wave of motion videos made with coding agents appeared on X. @twoclipping open-sourced a prompt template, and @notdwd showed a 12-second example: 1920x1080, 30fps, one HTML file, all visuals coming from a single seek(t) function, rendered with Playwright plus ffmpeg, with sound also synthesized by code.

This mechanism is worth unpacking.

What Is seek(t)

In ordinary video editing, the timeline moves forward, and each frame depends on the state of the previous frame. Code rendering requires the opposite: given a timestamp, directly compute what that frame looks like.

A 25-second video at 30fps has 750 frames. The time corresponding to frame f is:

t = f / 30

The renderer calls seek(t), and the animation returns the state at that moment. Frame 600 can be calculated on its own; there is no need to render frames 0 through 599 first.

This means the animation cannot depend on the previous frame, cannot contain uncontrolled randomness, and cannot use timers that only run once in the browser. Every visible property must be derived from the timestamp.

eigen’s article breaks the production system into six stages: Brief compilation, scene engine, renderer, quality control, variant generation, and delivery. The core tool is Remotion. It lets you write video with React components; titles, product screenshots, data charts, feature reveals, logo transitions, and CTAs can all be made into reusable components, and then brand data can be passed in.

Canvas, SVG, or HTML can also work. Playwright controls Chromium, waits for fonts and assets to load, calls window.seek(frame / fps), captures the frame, and hands it to FFmpeg for encoding. Remotion also officially provides Agent Skills; the install command is npx skills add remotion-dev/skills. If you are more comfortable describing animation with HTML, CSS, and media elements, you can look at HyperFrames, but there is no need to run two renderers for every project at the same time.

What Each of the 16 Repositories Handles

eigen’s list covers visuals, 3D, sound, recording, and delivery. Organized by responsibility, it is easier for readers to judge which piece they are missing.

Visuals and motion

01 Remotion: React video engine. https://github.com/remotion-dev/remotion

02 Storyboarder: storyboards and animatics. https://github.com/wonderunit/storyboarder

03 Anime.js: animation timelines and easing. https://github.com/juliangarnier/anime

04 Flubber: smooth shape morphing. https://github.com/veltman/flubber

05 Pixelorama: pixel art and sprite animation. https://github.com/Orama-Interactive/Pixelorama

3D, physics, and shaders

06 Three.js: 3D scenes, lighting, and cameras. https://github.com/mrdoob/three.js

07 PlayCanvas: real-time 3D worlds. https://github.com/playcanvas/engine

08 cannon-es: 3D physics and collisions. https://github.com/pmndrs/cannon-es

09 Sverchok: procedural geometry in Blender. https://github.com/nortikin/sverchok

10 Effekseer: particle effects and stylized VFX. https://github.com/effekseer/Effekseer

11 twigl: real-time shader coding and GIF loops. https://github.com/doxas/twigl

Sound

12 Bark: generated speech and sound. https://github.com/suno-ai/bark

13 Pedalboard: reverb, compression, and audio effects. https://github.com/spotify/pedalboard

Recording, editing, and encoding

14 OBS Studio: screen capture and recording. https://github.com/obsproject/obs-studio

15 Shotcut: timeline editing and finishing. https://github.com/mltframework/shotcut

16 FFmpeg: encoding, compression, and delivery. https://github.com/FFmpeg/FFmpeg

The list says free, but for commercial use do not look only at that word. Remotion has a separate license; qualifying individuals and teams of up to three people can use the Free License, while larger teams and certain automation setups have different requirements. Copyright for client logos, fonts, music, and sound effects must also be confirmed separately. Open-source repositories solve the cost of tools, not licensing.

One Brief Becomes 18 Exports

eigen’s goal is very specific: a SaaS founder sends a website, logo, brand colors, screenshots, and product description; the system produces a 25-second product video, 3 opening hooks, 3 aspect ratios 9:16, 1:1, and 16:9, and 2 CTAs. That is 18 exports in total.

The easiest mistake here is cropping a landscape video directly into portrait. Portrait should be re-laid out: interface components may stack vertically, while landscape may place them side by side. What is shared is the narrative and motion timing; layout constraints are calculated separately per aspect ratio. It is similar to responsive web pages, except the output is video.

Another thing that is easy to skip is input validation. Can the fonts load? Will the CTA copy overflow? Do the referenced screenshots exist? Has the client confirmed the product claim of a “240% increase in conversion rate”? A production system should reject bad input, avoiding the generation of a video that looks beautiful but is factually wrong.

Quality Checks Cannot Be Left to “looks good”

In eigen’s article, quality control takes up a large portion. The process has five rounds: first look at the storyboard and judge whether the story is coherent; then look at the frames before and after transitions and check for overlapping text, disappearing elements, and jumps; then shrink the preview to phone size and see whether it is readable on a small screen; next verify branding and facts; finally check the full video and audio sync.

A contact sheet can quickly show the whole video; use FFmpeg to extract one frame per second and stitch them into one image. But fixed intervals are not enough. The broken frame in a transition may happen to fall between two sampled images. Take extra samples before and after every transition.

Also, do not let the agent look only at the code and say “the animation is great.” It must inspect the rendered result and provide the timestamps checked, problems found, fixes made, and remaining risks, with human confirmation at the end.

@mablesjoseph shared a 45-second hand-drawn animated short and mentioned using 163 model calls and hours of work. The final film was not generated in one shot. This example is closer to real production than “one-click video.”

Where Is the Commercial Value?

Tony Dinh said that a year ago he paid more than $1,000 for a similar video, and now he can make it in less than 30 minutes with Opus 5.5. @achxvi also showed a product video completed in 45 minutes. These cases show that output is starting to enter the usable range; they do not prove that everyone can immediately make money from it.

eigen gave a hypothetical quote: $750 for a launch package, including one product, one approved narrative, 3 openings, 3 aspect ratios, 2 CTAs, a final MP4, a cover frame, and a limited number of revision rounds. Two packages for $1,500, before tax, not deducting software, compute, audio licensing, sales, and revision costs. These are illustrative numbers, not the market average.

The metrics more worth tracking are: how long the first design takes, how long each additional variant takes, how much rendering and API costs are, how many revision requests there are, how many delivery errors occur, and whether the client buys again. Automation only makes business sense if it reduces repetitive production work without sacrificing quality.

Someone in the comments asked how audio is synced. In eigen’s article, the approach is to analyze the beat and peaks of the audio track, bind button clicks, chart completions, transitions, and logo reveals to timestamps, and finally package with FFmpeg. Sound must be checked just like visuals.

A More Stable Starting Order

Do not build a platform first. Choose one kind of video, such as a 25-second dynamic product dashboard. Prepare one product with clear licensing, one real logo, a set of screenshots, and a style guide. Make four scenes: a strong opening title, an interface interaction with a cursor, a data or feature visualization, and product identity plus CTA.

Make this one piece genuinely good-looking. Then turn text, images, colors, and timing into configurable data. Then add a second aspect ratio, a third aspect ratio, and opening variants. Only after quality is stable should you add batch jobs and a client interface.

The prerequisite for automation is having something worth automating. Batch-exporting mediocre animation will only produce more mediocre animation.

eigen’s article ends with a judgment: a good video is a portfolio; a reusable motion engine is an asset. The former earns attention; only the latter can become a production process.