Optical
Optical
A Really Good Work Inc. project

One scene in. Every asset your brand needs out.

Write an asset once, as JSON. Optical renders it as a PNG, GIF, MP4 or PDF, in your brand, fast enough for an agent to try ten versions.

01 The problem

Brand work still gets made one size at a time.

A launch needs a story, a square, a banner, a share card, a GIF and a PDF. Someone builds the same thing six times. Then the copy changes and they do it again.

Agents can write the copy now. They can’t make it look like you. That part still lands on a designer’s desk, one file at a time.

02 How it works

Write the scene. Get the files.

  1. 01Describe the asset as layers: text, images, video, shapes, even real HTML. Each layer gets timing and motion.
  2. 02Point it at your brand. Colours, fonts and type styles come from tokens, not hex codes.
  3. 03Optical draws every frame in Rust. No headless browser.
  4. 04Out come the formats and sizes you asked for, all in one call.
Code 01A scene · JSON
{
  "size": "instagram-story",
  "duration": 6,
  "layers": [
    { "type": "video", "src": "launch.mp4", "trim": [2, 8] },
    { "id": "headline", "type": "text", "text": "Summon the crew.",
      "textStyle": "display", "position": "bottom-left", "width": "80%",
      "enter": "rise" },
    { "id": "logo", "type": "image", "src": "$assets.logo",
      "position": "top-left", "width": "30%",
      "enter": { "preset": "fade", "delay": 0.4 } }
  ]
}
Positions and widths, not pixels. The same scene reflows to a square or a banner.

Render it as ["mp4", "gif", "png"] in ["instagram-story", "square", "og"] and you get nine files back.

03 The reels

Three sizzle reels. No footage.

Made for Demons, a test brand with a crew of mascots. Each reel is one JSON file cut to a 120 BPM beat: kinetic type, colour wipes, bobbing characters, ticking counters.

Fig. 01Demons · three sizes
A · Reel 1080×1920 · 12 s
B · Square 1080×1080 · 10 s
C · Landscape 1920×1080 · 9 s
Rendered on a 2-core server in about 11 seconds each. 1.3 to 1.9 MB.
Beat cuts
Shots, wipes and word slams land on every half second.
Real CSS
The conveyor and the counters are HTML with @keyframes, driven by the scene’s clock.
Brand rules
No mascot on its own colour. Three at most. Caps only on stickers. The validator checks contrast on every shot.
04 Formats

One scene, every format.

Stills, loops, video and print come from the same file. Change a word once and every version follows.

Stills
PNG, JPEG, WebP, SVG
Loops
GIF, APNG, animated WebP
Video
MP4 and WebM, with music and fades
Print
PDF, one page per size. Typst for long documents.
Sizes
24 presets, from story and square to leaderboard ads and letter pages.
Media in
Video with trim, speed and loop. Images. Music.Canvas, WebGL and WebGPU drawing for anything else.
05 Speed

Fast enough to try ten versions.

An agent learns by looking, so looking has to be cheap. Measured on a small 2-core server:

Share card
12 ms
Square JPEG
33 ms
2-second GIF
144 ms
5-second square MP4
0.84 s
6-second story MP4
1.6 s
Same request again
Under 1 ms, from cache

Why it’s cheap to run.

Most of a brand video doesn’t move. A logo sits still for two seconds. A headline holds. So Optical only draws what changed.

Fig. 02Frame reuse
12 frames · 5 drawn · 7 reused 123456789101112 Solid: drawnDashed: same pixels as the last drawn frame
A typical brand clip draws a quarter of its frames. The rest are copies.
Frame reuse
Each frame gets a fingerprint before it’s drawn. Same fingerprint, same pixels, no work.
Video in the encoder
Background footage is scaled and zoomed by ffmpeg, under a see-through overlay. That alone cut one render’s CPU by 88%.
Drafts
Preview at half size as a contact sheet. Render the finals once.
The last tune-up
28% faster, 32% less server CPU, files 2% smaller.
06 Device or server

Their device when it can. Your server when it should.

The same scene renders in the browser with WebAssembly and WebCodecs, or on Node with ffmpeg. Same pixels either way, so you pick by cost, not by look.

Fig. 03Where a render runs
Scene Fits thisdevice? DeviceWASM · WebCodecs ServerNode · ffmpeg File yesno failed
Small jobs stay on the visitor’s machine. Big ones come to you.
Device
Stills, GIFs and short clips the browser can encode, under a size budget. Costs you nothing.
Server
Long video, missing codecs, agents and scripts. Runs on Vercel or any Node box.
Fallback
If the device can’t finish, the server picks it up. The person waiting never knows.
Long video
Split into chunks, render them in parallel, join them in one pass.
07 For agents

Built for agents to drive.

Models are good at JSON and bad at pixels. So the agent writes the scene, Optical does the drawing, and every call hands back a picture the model can check.

get_brand
Tokens, type styles, assets and sizes. Start here.
preview
The whole timeline as one contact sheet. One image to look at, not a video.
edit_scene
Small changes by layer id. No rewriting the scene to move a logo.
render
Every format and size in one call.
validate
Names the mistake and where it is: a token that doesn’t exist, timing that can’t happen, text you can’t read.
Code 02An edit · JSON
[
  { "op": "set", "target": "headline", "path": "text", "value": "Meet the crew." },
  { "op": "merge", "target": "logo", "value": { "enter": "pop" } },
  { "op": "retime", "target": "logo", "start": 1.2 }
]

Plug it into Claude, Cursor or anything that speaks MCP with optical mcp --brand brand.json. Or call the same tools from your own code.

08 Tested

Checked against Chrome on every pull request.

91 CSS cases render in Optical and in Chrome, then get compared pixel by pixel. The results post to the pull request.

CSS
85 of 91 match.Not yet: 3D perspective, writing-mode, container queries, !important, full-turn rotations, list markers.
Browser APIs
20 of 20.WebAssembly, WebCodecs (H.264, VP9, AV1), OffscreenCanvas, WebGL, WebGL2, WebGPU, audio mixing, workers.
Cost
Free. GitHub’s standard runners and first-party actions only.
09 Try it

Render the reels yourself.

Code 03Terminal
git clone https://github.com/reallygoodwork/optical
cd optical
npm install
npm run demons   # renders the three reels into examples/demons/videos

Node 20 and ffmpeg. To use it in your own project, install it straight from GitHub: npm install github:reallygoodwork/optical. The README covers the rest. Get the code

10 Studio

Made by Really Good Work Inc.

Why “Optical”
Before computers, film titles and effects came off an optical printer. Its operator layered footage, mattes and type into one finished negative, frame by frame, the same way every time. Same job, new machine.
Built on
Takumi for drawing, Mediabunny for video in the browser, ffmpeg on the server.
License
MIT. Use it for anything.Just keep the credit.Read it
Really Good Work Inc.