Turn a JSON shot list into a film-style storyboard sheet and a rough animatic, before you pay a video model to generate anything.
Quick start · Gallery · Write a spec · CLI · Claude Code plugin · Video models · FAQ
Left: shot 3 of rooftop-chase, in the 3d and 2d looks. Right: the top of the storyboard sheet for high-noon.
konte reads a shot list written as JSON and draws two things from it:
- A storyboard sheet, laid out like the boards used on a film production: numbered panels for every shot with their timecodes, and under them the camera in words, the action, the dialogue and who is in the shot.
- An animatic with no text in it: grey sets, one coloured mannequin per character, and the camera you described, cut shot by shot. Each shot is also saved as its own clip, with prompts, for video models that take a reference video, such as Seedance and Kling.
Everything runs locally. There is no model, no account and no hosted service, and no dedicated GPU is needed. The staging is computed, not generated, so the same spec gives the same shots, and a person or an agent can lint it, diff it and regenerate it from a script.
You need Node 22+, Chrome or Chromium, and ffmpeg (ffmpeg only for video; --board-only works without it). On macOS: brew install ffmpeg.
npx @sshworld/konte render rooftop-chaserooftop-chase is one of four bundled examples. To keep the konte command around, install it globally:
npm install -g @sshworld/konte
konte render rooftop-chaseThe files land in ./konte-out/<title>/, where the title is made safe for a folder name (Rooftop Chase becomes Rooftop-Chase). Use -o to choose the folder.
konte-out/Rooftop-Chase/
├── board-3d.png the storyboard sheet: panels, camera, action, dialogue, cast
├── board-2d.png the same sheet in the ink-on-paper look
├── sketch-3d.mp4 the animatic, picture only
├── sketch-2d.mp4
├── shots/
│ ├── 01-3d.mp4 one clip per shot, cut on exact frames
│ ├── 01-3d-first.png its first and last frame, for image-to-video models
│ └── 01-3d-last.png ...
├── prompts.md two prompts per shot for a video model
└── shots.json shot metadata and the list of files written, for scripts
On an M1 Pro laptop with konte 1.0.1, konte render high-noon (35 seconds, seven shots, both looks, default 12 fps) took about 23 seconds end to end; --board-only took about 5 (about 2 per look).
- Storyboard sheets. A header with the title, logline, runtime and a cast legend; one card per shot with up to four panels (1A, 1B, ...), timecodes read from the frames actually drawn, and Camera, Action, Dialogue and In shot rows.
- Text-free video. No titles, captions, name tags or timecodes in the picture, in the animatic, the shot clips or the review copy. Every word is on the sheet and in
prompts.md. - Shots written the way a storyboard writes them.
"size": "FS", "angle": "side", "move": "track". konte solves where the camera stands and swings it round a wall that blocks the view. - Staging checks before drawing.
konte lintreports a subject out of frame, a camera inside a wall, someone running at 15 m/s or sliding while standing. - Two looks from one staging.
3dis a lit greybox;2dis ink and flat tone on paper. - Rough on purpose. Mannequins and boxes, so you judge blocking, framing and timing, not faces or textures.
Every image below was drawn by konte from the JSON in examples/. The animations are excerpts; click a board to see the whole sheet.
|
High Noon · 16:9 · 7 shots · 35 s
|
The Heist · 16:9 · 5 shots · 27 s
|
|
Rooftop Chase · 9:16 · 5 shots · 28 s |
First Contact · 1:1 · 5 shots · 31 s
|
The animatic is only the plan. Here it is next to a finished piece built on top of it: the same blocking, timing and camera, rendered with a character, a pudding and a meadow road.
Left: the konte animatic. Right: a finished 3D render that follows it shot for shot. An 8-second excerpt of a 20-second, 9:16 piece.
A spec is a cast and a list of shots. Save this as standoff.json:
{
"konte": 1,
"title": "Standoff",
"aspect": "16:9",
"cast": {
"stranger": { "label": "The Stranger", "color": "#e8457e" },
"marshal": { "label": "The Marshal", "color": "#3aa0e0" }
},
"scenes": [
{
"title": "Over the shoulder", "seconds": 5,
"note": "Twenty paces of dust between them. The Stranger answers without moving.",
"dialogue": [["stranger", "Just passing through."], ["marshal", "Then keep passing."]],
"actors": [
{ "who": "stranger", "at": [0, 7], "face": 180, "do": "stand+talk", "look": "marshal" },
{ "who": "marshal", "at": [0, -5], "face": 0, "do": "stand", "look": "stranger" }
],
"camera": { "on": "stranger", "over": "marshal" }
}
]
}konte lint standoff.json # check the staging, draw nothing
konte render standoff.json --board-only # the sheets alone, in a few seconds
konte render standoff.json # sheet, animatic, shot clips and prompts- Cast members are people (a mannequin in their colour) or objects (box, ball, cylinder, bottle, shoe). Colour is how you tell them apart across cuts; the sheet's cast legend says who is which.
- Actors stand
ata point or follow a timedpath, anddoactions such aswalk,run,sit,kneel,leap,reach,talkorcarry, alone or combined ("run+carry"). Ajumplifts them along an arc. - Camera is written as a shot:
size(ECUtoEWS),angle,height,move(track,push,pull,orbit,crane-up, ...), over the shoulder withover, or an exact position when you need one. - Words go in
title,noteanddialogue. They appear on the sheet and in the prompts, never in the video.
The full reference, with every key, unit and default, is docs/spec.md.
konte lint <spec.json> check a spec without drawing anything
konte render <spec.json> [options] draw the storyboard sheet and the animatic
konte render <example> draw a bundled example: first-contact, high-noon, rooftop-chase, the-heist
konte --help | --version
| Option | Default | |
|---|---|---|
-o, --out <dir> |
./konte-out/<title> |
Where to write. Specs with the same title share a folder, so give each its own -o if you keep several. |
--look <3d,2d> |
3d,2d |
Which looks to draw: 3d, 2d or both |
--board-only |
Write the storyboard sheet and nothing else (no ffmpeg needed) | |
--review |
Also write review-<look>.mp4: the animatic with each actor's path drawn on the floor. Still no text. |
|
--fps <n> |
spec's fps, else 12 |
Frames drawn per second of story, 1 to 60. Each is held for a whole number of video frames, so the mp4 plays at 24 fps or more. |
--jobs <n> |
half the cores, up to 6 | Browser windows drawing frames at once (at most 12) |
If Chrome is not in a standard place (Windows, a custom install), point CHROME at the executable, for example CHROME="/opt/chrome/chrome" konte render spec.json.
A spec that cannot be read stops before anything is drawn. Staging errors found by lint are printed, the files are still written, and the command exits with status 1.
A full render replaces the files konte wrote in that folder before, and removes clips an earlier render listed in its shots.json but this one did not write, so an old clip cannot pass for a current one. It never deletes files it has no record of writing, but it does overwrite its own file names. --board-only only replaces the board, and says so when older videos are beside it.
konte ships as a Claude Code plugin with a skill, so Claude can storyboard an idea you describe in words:
/plugin marketplace add sshworld/konte
/plugin install konte@konte
Then ask, for example:
Using konte, storyboard a 30-second heist in five shots. Show me the board before rendering the animatic.
The skill has Claude write the spec, lint it, draw the board first, revise until the board tells the story, and only then render the video. See skills/konte/SKILL.md.
Untested. konte's clips and prompts have not yet been tried with Seedance, Kling, Higgsfield or any other video model. The notes below come from the vendors' documentation, read in November 2026.
Some video models accept a reference video and use it to guide camera, pacing and blocking. ByteDance's Seedance 2.5 announcement describes a "clay render" reference built from untextured 3D models; konte's 3d clips are that kind of picture.
konte render spec.json --look 3d- Open
prompts.md. Each shot lists its files inshots/and two prompts. The prompts name the colours of the first look you render. - Attach the shot's clip as the reference video, paste prompt A, and add your own character, product and style references.
| Model | What goes in | What its documentation says |
|---|---|---|
| Seedance 2.5 | The clip, as a reference video | Up to 10 reference videos in one pass; output up to 30 seconds. No limit is given for a reference clip's length. (ByteDance) |
| Seedance 2.0 | The clip, named @Video1 in the prompt |
Up to 3 reference videos, 2 to 15 seconds combined, under 50 MB in total. (fal.ai) |
| Kling 4.0, Omni Reference | The clip, also @Video1 |
Up to 5 videos, 3 to 30 seconds each, 30 seconds in total. (Kling guide) |
| Image-to-video models, such as Seedance 1.5 Pro | -first.png, and -last.png where the model takes an end frame |
No reference video. Restyle the still with an image model first, then use prompt B, which describes the motion in words. |
The Seedance 1.5 row rests on provider pages that list only image and text inputs. Platforms and versions differ; check the one you use. konte lint notes shots longer than 15 or 30 seconds as information only; it does not check file size, resolution or a platform's combined limits.
spec.json ──► src/stage.js ──► konte lint
│
├──► src/render.html (Three.js in headless Chrome) ──► frames ──► ffmpeg ──► sketch, shots/
├──► storyboard sheet (HTML, screenshotted) ──► board-<look>.png
└──► src/prompts.js ──► prompts.md
src/stage.js is a pure model: given a spec and a time, it says where everyone is and where the camera is. Lint, the renderer and the prompt text all run on it, so lint judges the same positions that get drawn and the camera words are measured from the solved camera. The CLI serves the renderer on a local port, drives it in headless Chrome over several windows and pipes the frames to ffmpeg. The renderer page has no text layer; the sheet is a separate page built from the same model.
Can I write the spec in another language?
Yes. Titles, notes, dialogue and labels appear on the sheet exactly as written. On Linux, install a font that covers your script (for Korean, Japanese or Chinese, fonts-noto-cjk).
Why is there no text in the video, not even a shot number? The video is a reference, not a deliverable. Text in a reference clip may be reproduced in the generated output, and it gets in the way when you watch for timing. The sheet holds every word.
Does it make the final video? No. There are no faces, no lighting design and no physics. konte is the pencil sketch that comes before generation or a shoot.
Is there an editor? No. You change the picture by changing the file, which is what makes it easy for an agent to work with.
Where does the name come from? Konte is the everyday word for a storyboard on Korean and Japanese productions.
- For an interactive 3D studio in the browser, with rigged characters, a timeline and an agent panel, use CozyClay. It covers the same greybox-to-video-model idea with far more control.
- For a finished, editable promo drawn in code, use video-shotcraft.
- Choose konte when you want the staging in a text file you can diff, lint and regenerate from a script or an agent, running locally with no account.
MIT. Three.js is bundled under its own MIT license in vendor/.






