MDX Video Framework

egaki

AI image and video generation from the terminal. MDX-to-video framework built on Remotion.

egaki is a TypeScript CLI built on the Vercel AI SDK and goke. It supports Google Imagen, Gemini image-capable models, video generation (Veo, Kling, Wan, Bytedance Seedance, xAI Grok), and model discovery with pricing.

It also includes a MDX-to-video framework built on Remotion: write MDX with headings as section boundaries, use built-in animation primitives, easing presets, layout transitions, captions, and TTS. Export to MP4 directly in the browser.

Install

pnpm add -g egaki

Quick start

egaki login
egaki image "a watercolor fox reading a map" -o fox.png
egaki video "a paper boat drifting on a calm lake at sunrise" -o boat.mp4

Quick start with ChatGPT auth

egaki login --provider chatgpt
egaki image "a dreamy studio ghibli style bakery at sunrise" -m gpt-image-1.5 -o bakery.png
egaki image "change the red jacket to a blue jacket" -m gpt-image-1.5 --input portrait.png -o portrait-blue.png

CLI examples

Generate an image from text

egaki image "cinematic mountain village at sunrise" -o village.png
egaki image "isometric floating city, detailed, soft colors" -m imagen-4.0-generate-001

Edit with an input image

egaki image "add a red scarf and make it winter" --input portrait.jpg -o portrait-winter.png
egaki image "turn this into a manga panel" --input https://example.com/photo.jpg -o manga.png
egaki image "change the red square to a blue square" -m gpt-image-1.5 --input input.png -o output.png

Inpainting with a mask

egaki image "replace the sky with dramatic storm clouds" --input landscape.png --mask mask.png -o storm.png

Generate multiple images

egaki image "minimal logo concepts for a cat cafe" -n 4 -o logo.png

Control composition

egaki image "cyberpunk alley at night" --aspect-ratio 16:9
egaki image "polaroid-style travel photo" --aspect-ratio 4:5
egaki image "wide landscape matte painting" -m gpt-image-1.5 --aspect-ratio 3:2 -o wide.png

Use Google Cloud billing via Vertex AI

egaki login --provider vertex --key AIza...
egaki image "editorial sneaker photo on white seamless" -m vertex/imagen-4.0-generate-001 -o sneaker.png
egaki video "storm over mountains" -m vertex/veo-3.1-fast-generate-001 --duration 6 -o storm.mp4

Pipe image output to other tools

egaki image "flat icon of a fox" --stdout | magick - -resize 512x512 fox-icon.png

Generate a video from text

egaki video "a paper boat drifting on a calm lake at sunrise" -o boat.mp4
egaki video "timelapse of a stormy sea, cinematic" -m veo-3.1-generate-001 --duration 8 -o storm.mp4

Generate with a cheap model

# Kling v2.5 Turbo — fast and inexpensive
egaki video "a cat walking on a rooftop at night" -m klingai/kling-v2.5-turbo-t2v --duration 5 -o cat.mp4

Image-to-video

# Animate a still image (model must support i2v)
egaki video "slowly animate the clouds" --input photo.jpg -m klingai/kling-v2.6-i2v -o animated.mp4

Control resolution and aspect ratio

egaki video "aerial drone shot over a city grid" \
  -m veo-3.1-fast-generate-001 \
  --aspect-ratio 16:9 \
  --resolution 1080p \
  --duration 6 \
  -o city.mp4

Generate multiple videos

egaki video "waves crashing on cliffs at golden hour" -n 2 -o waves.mp4
# writes waves.mp4 and waves-1.mp4

Pipe video output to other tools

egaki video "looping rain animation" --stdout | ffmpeg -i pipe:0 -vf fps=12 rain.gif

xAI Grok image generation

# Basic text-to-image
egaki image "a fox wearing armor in a misty forest" -m grok-imagine-image -o fox.png

# High quality + 2K resolution
egaki image "product photo of a ceramic vase on linen" \
  -m grok-imagine-image-quality \
  --quality high \
  --resolution 2k \
  -o vase.png

# Edit an existing image
egaki image "add dramatic storm clouds to the sky" \
  -m grok-imagine-image \
  --input landscape.jpg \
  -o landscape-storm.png

# Generate multiple variations
egaki image "abstract geometric pattern" -m grok-imagine-image -n 3 -o pattern.png
# writes pattern-0.png, pattern-1.png, pattern-2.png

# Control aspect ratio and output format
egaki image "vertical phone wallpaper, aurora borealis" \
  -m grok-imagine-image \
  --aspect-ratio 9:16 \
  --output-format jpeg \
  -o wallpaper.jpg

xAI Grok video generation

# Basic text-to-video
egaki video "a paper airplane gliding through clouds" \
  -m grok-imagine-video \
  --duration 5 \
  -o airplane.mp4

# Text-to-video with resolution and aspect ratio
egaki video "cinematic drone shot over a city at sunset" \
  -m grok-imagine-video \
  --duration 8 \
  --resolution 720p \
  --aspect-ratio 16:9 \
  -o city.mp4

# Image-to-video (animate a still image)
egaki video "slowly pan across the scene with gentle wind" \
  -m grok-imagine-video \
  --input photo.jpg \
  --duration 5 \
  -o animated.mp4

# Video editing (modify an existing video — pass the source via --input)
egaki video "make it look like a watercolor painting" \
  -m grok-imagine-video \
  --mode edit-video \
  --input ./original.mp4 \
  -o edited.mp4

# Video extension (continue from last frame)
egaki video "the camera keeps moving forward" \
  -m grok-imagine-video \
  --mode extend-video \
  --input ./clip.mp4 \
  -o extended.mp4

# Reference-to-video (R2V): generate a video guided by 1-7 reference images.
# The images act as style and content references (not as the first frame).
# Accepts local files or URLs; local files are uploaded automatically.
egaki video "the model walks down a white runway wearing the outfit from the reference" \
  -m grok-imagine-video \
  --mode reference-to-video \
  --reference-images ./model-face.jpg \
  --reference-images ./outfit.jpg \
  --duration 8 \
  -o runway.mp4

# Low-res draft for quick iteration
egaki video "test animation concept" \
  -m grok-imagine-video \
  --duration 3 \
  --resolution 480p \
  -o draft.mp4

xAI Grok auth

# Option 1: direct API key
egaki login --provider xai --key xai-...

# Option 2: Grok Build subscription (browser OAuth)
egaki login --provider xai-oauth

# Check auth status
egaki login --show

Discover models and pricing

egaki models
egaki models --type video
egaki models --type image
egaki models --provider google
egaki models --json

Advanced image examples

# Deterministic result (models that support seed)
egaki image "studio product shot of a ceramic mug" \
  -m imagen-4.0-generate-001 \
  --seed 42 \
  -o mug-seed-42.png

# High quality Gemini image generation with 4K output
egaki image "architectural concept art, brutalist library interior" \
  -m gemini-2.5-flash-image \
  --image-size 4K \
  --aspect-ratio 16:9 \
  -o library-4k.png

# Multi-reference edit (pass --input multiple times)
egaki image "blend style from first image and color palette from second" \
  --input style-reference.jpg \
  --input palette-reference.jpg \
  -m nano-banana-pro-preview \
  -o hybrid-style.png

# JSON mode for automation/pipelines
egaki image "futuristic sneaker concept" \
  -m gpt-image-1.5 \
  --json \
  -o sneaker.png

# Batch generation with indexed output names
egaki image "mascot variations, flat vector look" \
  -m gpt-image-1-mini \
  -n 6 \
  -o mascot.png
# writes mascot-0.png ... mascot-5.png

Model quick reference

Image models

Model ID	Best for	Example command
`imagen-4.0-ultra-generate-001`	High-quality prompt-to-image with seed + ratio	`egaki image "luxury perfume ad on marble" -m imagen-4.0-ultra-generate-001 --aspect-ratio 3:4 --seed 7 -o perfume.png`
`gemini-3.1-flash-image-preview`	Fast, cheap text+image edits with wide aspect-ratio support	`egaki image "turn into manga splash page" -m gemini-3.1-flash-image-preview --input portrait.jpg --aspect-ratio 4:1 -o manga-wide.png`
`nano-banana-pro-preview`	Highest-fidelity Google text+image output	`egaki image "fashion editorial, dramatic rim light" -m nano-banana-pro-preview --input model.jpg --image-size 2K -o editorial.png`
`gpt-image-1.5`	OpenAI image generation with strong editing/inpainting	`egaki image "replace background with neon city" -m gpt-image-1.5 --input product.png --mask bg-mask.png -o product-neon.png`
`fal-ai/flux/schnell`	Very fast low-cost ideation batches	`egaki image "logo sketch, geometric fox" -m fal-ai/flux/schnell -n 8 -o fox-logo.png`

Video models

Model ID	Best for	Example command
`veo-3.1-generate-001`	Highest quality video with audio, up to 4K	`egaki video "rainy Tokyo street at night" -m veo-3.1-generate-001 --duration 8 -o tokyo.mp4`
`veo-3.1-fast-generate-001`	Fast Veo with 720p-4K, good for iteration	`egaki video "abstract paint patterns" -m veo-3.1-fast-generate-001 --duration 5 -o paint.mp4`
`vertex/veo-3.1-generate-001`	Same as above, routed through Vertex AI	`egaki video "rainy Tokyo street" -m vertex/veo-3.1-generate-001 --duration 8 -o tokyo.mp4`
`klingai/kling-v2.5-turbo-t2v`	Cheap, fast Kling text-to-video	`egaki video "a paper boat on a pond" -m klingai/kling-v2.5-turbo-t2v --duration 5 -o boat.mp4`
`bytedance/seedance-v1.5-pro`	Bytedance, audio support, three resolutions	`egaki video "timelapse of clouds above mountains" -m bytedance/seedance-v1.5-pro -o clouds.mp4`
`grok-imagine-video`	xAI video with editing, extension, R2V	`egaki video "a dog catching a frisbee" -m grok-imagine-video --duration 5 -o dog.mp4`

Feature support by model family

Google Imagen (imagen-*): supports --seed, --aspect-ratio, --input, --mask, -n
Google Gemini image models: supports --input, --aspect-ratio, --image-size; usually no --seed
OpenAI image models: strong editing and inpainting; size controls are model-specific
BFL image models (flux-*): Kontext/Pro variants via AI Gateway subscription
Recraft models (recraft-*): v2/v3/v4 families available via AI Gateway subscription
xAI image models (grok-imagine-*): --quality, --resolution, --output-format, --input for editing, -n for batches. Auth via API key or Grok Build OAuth
Vertex models (vertex/*): same models as Google AI Studio, routed through Vertex AI / Google Cloud billing
Google Veo video models: up to 4K, audio optional, duration 4-8s
Kling video models: mode (std/pro), audio on v2.6+, image-to-video support
Bytedance Seedance: 480p-1080p, audio support on v1.5-pro
xAI Grok video (grok-imagine-video): 480p-720p, 1-15s, i2v, video editing, video extension, R2V from reference images

Subscription and usage

egaki supports both authentication modes:

BYOK (bring your own keys): add provider keys with egaki login per provider.
ChatGPT auth: log in with egaki login --provider chatgpt to use the ChatGPT/Codex backend for supported OpenAI image generation and editing flows.
xAI Grok Build auth: log in with egaki login --provider xai-oauth to use your Grok Build subscription for xAI image and video generation.
Egaki subscription: use one egaki_... key to access all supported models without managing keys for each provider.
Google vs Vertex: bare model IDs (e.g. imagen-4.0-generate-001) use Google AI Studio. Prefix with vertex/ (e.g. vertex/imagen-4.0-generate-001) to route through Vertex AI / Google Cloud billing.

# Subscribe and get a checkout URL
egaki subscribe --email user@example.com --plan pro

# Subscribe without email prefill
egaki subscribe --plan pro

# Save your Egaki key after checkout
egaki login --provider egaki --key egaki_...

# BYOK examples (direct provider keys)
egaki login --provider google --key AIza...
egaki login --provider vertex --key AIza...

# ChatGPT OAuth for Codex-backed image generation/editing
egaki login --provider chatgpt
egaki image "turn this product shot into a clay render" -m gpt-image-1.5 --input product.png -o product-clay.png

# Check subscription usage / cancel
egaki usage
egaki unsubscribe

CLI commands

Command	What it does
`egaki dev`	Zero-config MDX video dev server (just pass an .mdx file)
`egaki image`	Generate/edit images (Imagen, Gemini, GPT, Fal, xAI)
`egaki video`	Generate videos (Veo, Kling, Wan, Seedance, xAI)
`egaki speech`	Text-to-speech (OpenAI, ElevenLabs, Cartesia)
`egaki demucs`	Stem separation via fal.ai (no local torch needed)
`egaki bpm`	Detect BPM of audio (local, no API needed)
`egaki loudness`	Measure perceived loudness in LUFS (local, no API needed)
`egaki voice clone`	Clone a voice from audio (Cartesia, ElevenLabs)
`egaki transcribe`	Speech-to-text (OpenAI, ElevenLabs, Deepgram, Groq, Cartesia)
`egaki models`	List available models with pricing

Audio workflow example:

# 1. Separate vocals from a song
egaki demucs song.mp3 --stems vocals -o stems/

# 2. Clone the voice from the isolated vocals
egaki voice clone stems/song-vocals.mp3 --name "Singer" --json

# 3. Generate TTS with the cloned voice
egaki speech "Your text here." --voice <voice-id> -m sonic-3.5 -o output.mp3

# 4. Transcribe audio to get word timestamps
egaki transcribe recording.mp3 -m whisper-1

# 5. Detect BPM for beat-synced videos
egaki bpm soundtrack.mp3 --json

BPM detection

Detect the tempo of any audio file locally. No API call, runs in milliseconds.

egaki bpm song.mp3
# BPM: 120
# Interval: 0.500s

egaki bpm track.wav --json
# { "bpm": 120, "intervalInSeconds": 0.5 }

# Pipe from ffmpeg
ffmpeg -i video.mp4 -f mp3 - | egaki bpm --stdin

Use the detected BPM in your MDX video frontmatter to lock scene cuts to the beat:

---
fps: 30
bpm: 129
---

<Audio src="/soundtrack.mp3" />

# Verse duration=8beats

BPM detection requires node-web-audio-api (optional dependency). If missing, the command tells you how to install it:

pnpm add node-web-audio-api

Loudness measurement

Measure perceived loudness in LUFS using K-weighting and two-pass gating inspired by EBU R128. Useful for checking if your audio matches platform targets before compositing.

egaki loudness narration.mp3
# Integrated: -14.2 LUFS
# Max: -8.1 LUFS
# Min: -28.4 LUFS
# Range: 20.3 LU

egaki loudness track.wav --json

Common targets: YouTube/Spotify -14 LUFS, Apple Music -16 LUFS, podcasts -16 to -18 LUFS.

Same node-web-audio-api optional dependency as BPM detection.

Help

egaki --help
egaki image --help
egaki video --help
egaki models --help

Auth and billing

egaki login stores provider keys in ~/.config/egaki/credentials.json.
egaki subscribe, egaki usage, and egaki unsubscribe manage Egaki plans.
Video costs are tracked per-second based on model, resolution, and duration.

MDX Video Framework

Write MDX with headings as section boundaries; each heading becomes a timed Remotion Series.Sequence. Frontmatter sets global fps and bpm. Export to MP4 directly in the browser via WebCodecs.

Getting started

The fastest way is zero setup: point egaki dev at an .mdx file. No package.json, vite.config.ts, or npm install needed — react, remotion, and vite all come from the egaki CLI's own installation.

egaki dev video.mdx          # serve one file, random free port
egaki dev                    # serve current dir (auto-discovers video.mdx)
egaki dev intro.mdx --port 5199 --open

egaki dev creates a small node_modules in the folder containing symlinks into the CLI's install (marked with a .egaki-shim file, safe to delete). This also makes editor TypeScript resolution work in the scratch folder.

Create video.mdx:

---
fps: 30
bpm: 120
---

# Intro duration=3s

<TranslateX from={-140} to={0} duration={0.5 * FPS}>
  <div style={{ fontSize: 72, fontWeight: 900, color: 'white' }}>
    Hello World
  </div>
</TranslateX>

# Outro duration=2s

<Opacity from={1} to={0} duration={0.5 * FPS} startInFrames={-0.5 * FPS}>
  <div style={{ fontSize: 48, color: 'white' }}>Goodbye</div>
</Opacity>

Run egaki dev video.mdx and open the printed URL to see the player with controls and export.

Full project setup (optional)

For a real project with its own dependency versions, create a package and a vite config instead:

mkdir my-video && cd my-video
pnpm init
pnpm add egaki remotion @remotion/media react react-dom vite

Create vite.config.ts:

import { defineConfig } from 'vite'
import { video } from 'egaki/vite'

export default defineConfig({
  plugins: [video({ entry: './video.mdx' })],
})

Run vite (or a dev script) and open the browser.

How it works

MDX file
  │
  ▼
app.tsx (server, Spiceflow RSC)
  └── passes the raw MDX source string to the client via RSC flight
        │
        ▼
mdx-client.tsx ('use client', runs in the browser)
  ├── safe-mdx parses AST, resolves user imports
  ├── splits into MdxSection[] by headings, parses durations
  ├── renders each section to JSX (client-side React)
  └── renders PlayerPage with the sections
        │
        ▼
player-page.tsx (client)
  ├── wraps sections in Remotion <Series> / <Series.Sequence>
  ├── renders interactive <Player> with controls
  └── "Export MP4" button ► @remotion/web-renderer (WebCodecs, in-browser)

MDX renders fully on the client. Expression props can be functions (easing={x => x}), and user .tsx components don't need a 'use client' directive.

Sections and duration

Each # heading creates a new section. Set duration with a unit suffix:

Unit	Example	Meaning
`s`	`duration=3.5s`	Seconds (multiplied by fps)
`beats` / `beat`	`duration=8beats`	Beats (using frontmatter `bpm`)
`frames` / `frame` / `f`	`duration=90frames`	Raw frames
(bare number)	`duration=90`	Raw frames

---
fps: 30
bpm: 120
---

# Intro duration=3s

# Verse duration=8beats

# Bridge duration=90frames

# Outro duration=2s

Heading durations are parsed from raw text, not MDX expressions. duration={33 * BEAT} does not work; use duration=33beats instead.

Auto-duration from media

When a heading does not set an explicit duration, the section uses the maximum duration of any <Audio> or <Video> element inside it (including positive startInFrames offset). The longest media wins. If no media is present and no duration is set, the section falls back to a default frame count.

Pitfall: if you want a section's duration to match a short audio clip, never place a long-running video or full-length soundtrack <Audio> inside that section. Place long-running media in the preamble (before the first heading) so it plays across all sections without affecting any section's auto-duration.

`FPS` and `BEAT` scope variables

MDX expressions have access to FPS and BEAT as global scope variables, derived from frontmatter fps and bpm. No import needed.

FPS = frames per second (default 30). Use to convert seconds to frames.
BEAT = frames per beat, computed as fps / (bpm / 60). At 120bpm/30fps = 15 frames.

---
fps: 30
bpm: 120
---

# Intro duration=2s

<TranslateX from={-140} to={0} duration={0.5 * FPS} startInFrames={0.3 * FPS}>
  <div style={{ fontSize: 72, color: 'white' }}>Hello</div>
</TranslateX>

# Verse duration={8 * BEAT}

<Opacity from={0} to={1} duration={2 * BEAT}>
  Content appears over 2 beats
</Opacity>

Always use FPS instead of raw frame numbers. Raw frame literals like duration={20} break when the export fps changes (e.g. 30fps preview vs 60fps final render). This applies to duration, startInFrames, and any frame-based value.

<!-- correct -->
<TranslateX from={-140} to={0} duration={0.5 * FPS} cutInMotion={0.1}>

<!-- wrong, breaks at different fps -->
<TranslateX from={-140} to={0} duration={15}>

Beat-synced music

Align the beat grid to the actual track: detect the track's BPM with egaki bpm, find the time of its first downbeat, then trim so the downbeat lands on frame 0.

egaki bpm music.mp3 --json
# { "bpm": 129.2, "intervalInSeconds": 0.464 }

---
fps: 30
bpm: 129.2
---

{/* First downbeat at 0.372s → trim 11 frames so the grid starts on a beat. */}
<Audio src="/music.mp3" trimBefore={11} volume={0.5} />

To tighten pacing without breaking beat sync, raise the frontmatter bpm and speed the music up to match:

---
bpm: 140
---

<Audio src="/music.mp3" trimBefore={11} playbackRate={140 / 129.2} />

Scene cuts stay locked to the musical beat while every scene gets shorter.

Preamble

Content before the first # heading is the preamble. It renders at the Remotion composition level, outside the <Series> that sequences sections. Preamble content persists across all sections for the entire video duration, and renders in the background behind section content.

Use the preamble for:

Soundtracks: <Audio src="/music.mp3" /> plays for the full video
Ambient background video: <Video src="/bg.mp4" /> loops behind all sections
Global background: a <Background> with a shader or static color
Persistent overlays: watermarks, logos

---
fps: 30
bpm: 120
---

<Audio src="/soundtrack.mp3" />
<Video src="/ambient-bg.mp4" objectFit="cover" />

# First Section duration=5s

This content appears on top of the ambient background video.

Animation primitives

Built-in components available in MDX without imports: Opacity, Scale, TranslateX, TranslateY, Blur. Each animates one CSS property from from to to over duration frames.

Enter vs exit is inferred from startInFrames: positive or zero = enter (offset from section start), negative = exit (offset from section end).

# Scene duration=3s

<!-- Enter: slide in from left over 0.5s -->
<TranslateX from={-140} to={0} duration={0.5 * FPS}>
  <div style={{ fontSize: 72, fontWeight: 900, color: 'white' }}>Title</div>
</TranslateX>

<!-- Exit: fade out over last 0.5s of the section -->
<Opacity from={1} to={0} duration={0.5 * FPS} startInFrames={-0.5 * FPS}>
  <div style={{ color: 'white' }}>Subtitle</div>
</Opacity>

`inline` prop

By default, primitives use <Fill> wrapper (full-frame AbsoluteFill). Pass inline to wrap in a plain <div> instead, so the element stays in flow layout (flex, grid, etc.).

Use inline when animating elements inside a layout (cards, list items, flex children). Nest multiple inline primitives to compose animations:

<Scale from={0.8} to={1} duration={30} easing={impulseOvershoot(71)} inline>
  <Opacity from={0} to={0.5} duration={20} easing={(t) => t} inline>
    <div style={{ width: 100 }}>Content</div>
  </Opacity>
</Scale>

Never use % widths/heights on children of inline primitives. The inline wrapper has no intrinsic size, so percentage values resolve to 0. Use px or em instead.

Sequential animations

When an element has two animations at different times (e.g. a shrink-in followed by a pulse), nest two wrappers. Each applies independently:

<Scale from={1.7} to={1} duration={msToFrame(1490, fps)} easing={EASE.smooth} inline>
  <Scale from={1} to={1.1} duration={msToFrame(3572 - 1632, fps)}
         startInFrames={msToFrame(1632, fps)} easing={impulseOvershoot(96)} inline>
    <div style={{ width: '38%', aspectRatio: '675 / 392' }}>Card</div>
  </Scale>
</Scale>

`style` prop for layout

The style prop on inline primitives is essential for flex/grid participation. Use it to set dimensions, overflow, border-radius:

<Scale from={0} to={1} duration={24} easing={EASE.smooth} inline
  style={{ width: '100%', height: '100%', borderRadius: '20%', overflow: 'hidden' }}>
  <img src={src} style={{ width: '100%', height: '100%', objectFit: 'cover' }} />
</Scale>

`cutInMotion`

Use cutInMotion (0-1) to clip the animation at the scene boundary for conveyor-belt transitions. The scene cuts while text is still in motion, and the next scene's text appears already mid-slide.

Pattern:

First scene: enter with no cutInMotion, exit with cutInMotion={0.3}
Middle scenes: enter with cutInMotion={0.1}, exit with cutInMotion={0.2}
Last scene: exit with small cutInMotion={0.1}

# First Scene duration=3s
<TranslateX from={-140} to={0} duration={0.7 * FPS}>
  <TranslateX from={0} to={140} duration={0.7 * FPS} startInFrames={-0.7 * FPS} cutInMotion={0.3}>
    <div style={{ fontSize: 72, fontWeight: 900, color: 'white' }}>Title</div>
  </TranslateX>
</TranslateX>

# Middle Scene duration=3s
<TranslateX from={-140} to={0} duration={0.5 * FPS} cutInMotion={0.1}>
  <TranslateX from={0} to={140} duration={0.6 * FPS} startInFrames={-0.6 * FPS} cutInMotion={0.2}>
    <div style={{ fontSize: 72, fontWeight: 900, color: 'white' }}>Title</div>
  </TranslateX>
</TranslateX>

See example-cut-in-motion/ and example-shader/ for working demos.

Composing enter and exit moves

Combine primitives for entrances and exits with real presence: a scene-level Scale for the camera push, plus TranslateY/TranslateX for the element move. Enter animations should end at the natural position (to={0} or to={1}) so the resting frame is predictable.

# Hook duration=8beats

<Scale from={1} to={1.5} duration={8 * BEAT}>
<Background>
  <WaveGradientShader style={{ width: '100%', height: '100%' }} />
</Background>

<TranslateY from={80} to={0} duration={0.6 * FPS}>
  <TranslateY from={0} to={-80} duration={0.5 * FPS} startInFrames={-0.5 * FPS} cutInMotion={0.3}>
    <div style={{ fontSize: 92, color: 'white' }}>Title</div>
  </TranslateY>
</TranslateY>
</Scale>

Go big. Subtle values read as accidental drift; large values read as intentional motion. Prefer Scale from={1} to={1.5} over 1.05, translate distances of 80-260px over 10-20px, and pair them with short durations (0.4-0.7s) and strong easings. Reserve subtle magnitudes (+0.05 scale) for ambient drift layered under the main move, not for the move itself.

Typewriter text

Reveal text character by character, synced to the voiceover's word timestamps (see "Syncing animations to word timestamps"). Two details make it look right:

Lay out the full text invisibly and reveal characters on top, so centered lines never shift while typing.
Render the caret absolutely positioned (zero width) so it never wraps lines or pushes layout.

function Typewriter({ words, fontSize }: { words: { text: string; startSec: number; endSec: number }[]; fontSize: number }) {
  const frame = useCurrentFrame()
  const { fps } = useVideoConfig()
  const t = frame / fps
  return (
    <div style={{ position: 'relative', fontSize, color: 'white', textAlign: 'center' }}>
      {/* ghost copy keeps the layout stable */}
      <span style={{ visibility: 'hidden' }}>{words.map((w) => w.text).join(' ')}</span>
      <span style={{ position: 'absolute', inset: 0 }}>
        {words.map((w, i) => {
          const chars = w.text.length
          const shown = t < w.startSec ? 0 : t >= w.endSec ? chars
            : Math.ceil(((t - w.startSec) / (w.endSec - w.startSec)) * chars)
          return <span key={i}>{w.text.slice(0, shown)}{i < words.length - 1 && shown === chars ? ' ' : ''}</span>
        })}
        <span style={{ position: 'absolute', width: 0 }}>|</span>
      </span>
    </div>
  )
}

`<Fill>`

A full-frame layer like Remotion's AbsoluteFill but with better defaults for video content. Children stretch horizontally to fill the frame and center vertically. Available in MDX without imports. Accepts style to override alignment. Prefer <Fill> over raw <AbsoluteFill>.

`<AngledScreen>`

WebGL 3D screen with true depth-of-field. Children DOM is captured per frame (HTML-in-canvas) and rendered as a perspective-tilted plane in a fragment shader: parts of the plane further from the camera get progressively more bokeh blur, plus fog toward backgroundColor and subtle film grain. Available in MDX without imports:

# Hero duration=5s

<AngledScreen rotateX={9} rotateY={-17} translateZ={120} backgroundColor="#0a0608">
  <Img src="/screenshot.png" style={{ width: '100%', borderRadius: 14 }} />
</AngledScreen>

Depth of field uses the exact Three.js BokehShader algorithm with the same defaults as the original plugin, so the cinematic look is on by default: blur grows linearly with distance from the focus plane and saturates at maxBlur, sampled with a 41-tap ring kernel in screen space (plane edges melt into the background where out of focus).

Props: perspective (camera distance px, default 1200), rotateX/rotateY/rotateZ (degrees, CSS semantics), translateX (px horizontal shift), translateZ (px toward camera), bokeh (default true), aperture (shallower DOF with bigger values, default 0.5), maxBlur (blur saturation cap in screen-UV units, default 0.12), focus (focus distance as a fraction of perspective; default 0 = auto: near side sharp, blur ramps progressively from mid-image to the far edge), fog (fade to background with depth, default 0.35), grainIntensity (default 0.02), chromaticAberration (radial R/B split, stronger in the bokeh, default 0.45), backgroundColor, width/height (content wrapper size, default 80%/auto), debug (grayscale depth view). All props are editable live in the tweakpane panel.

Bokeh uses a sparse 6-blade hexagonal iris (13 weighted taps) so the aperture shape reads as hard disks instead of a smooth gaussian. Raise chromaticAberration for purple/cyan fringing like a real lens.

The defaults need no tuning: the near half stays sharp and only the far side melts. Raise aperture for an earlier, heavier falloff, or set an explicit focus to place the sharp plane manually (Three.js BokehPass semantics).

Requires Chrome 149+ with chrome://flags/#canvas-draw-element enabled. When unsupported it falls back automatically to BasicAngledScreen, the deprecated CSS-only predecessor (same transform props, directional backdrop-filter blur instead of true depth-of-field). Don't use it directly in new videos; the fallback is automatic.

Easing presets

Import from egaki/video. Always use cubicBezier() from egaki/video instead of Easing.bezier() from remotion, because the egaki version attaches metadata that lets the tweakpane UI show and edit the curve.

`springFromDuration` and `dspring`

Use these instead of raw spring({ config: { damping, stiffness, mass } }).

import { spring } from 'remotion'
import { springFromDuration, dspring, EASE } from 'egaki/video'

// springFromDuration returns a config object for Remotion's spring()
const scale = spring({ frame, fps, config: springFromDuration(0.5, 0.3) })

// dspring is the shorthand that calls spring() internally
const opacity = dspring(frame, fps, 0.6, 0.25)  // 600ms, subtle bounce
const pop = dspring(frame, fps, 0.4)              // 400ms, no bounce

bounce parameter: 0 = no overshoot, 0.25 = subtle Apple-like, 0.5 = playful, 1 = maximum bounce.

`EASE` presets

import { EASE, smoothEasing, bounceEasing } from 'egaki/video'
import { interpolate } from 'remotion'

const x = interpolate(frame, [0, 60], [0, 500], {
  easing: EASE.smooth,
  extrapolateLeft: 'clamp',
  extrapolateRight: 'clamp',
})

Bezier presets (clean curves, no overshoot): smooth, natural, decelerate, accelerate

Spring/bounce presets (values can exceed 0-1): elasticSnap, bounce, bounceAnticipate, bounceThrow, overshoot, overshootElastic, overshootBouncy, decelerateOvershoot, decelerateElastic, naturalThrow, accelerateImpulse, accelerateElastic, impulseSlow, impulseOvershoot

Continuous intensity functions

Each preset also has a continuous function that takes any intensity 0-100:

import { overshoot, naturalThrow } from 'egaki/video'

const scale = interpolate(frame, [0, 30], [0, 1], {
  easing: overshoot(63),
  extrapolateLeft: 'clamp',
  extrapolateRight: 'clamp',
})

All 14 presets: naturalThrow, decelerateOvershoot, decelerateElastic, accelerateImpulse, accelerateElastic, elasticSnap, bounce, bounceAnticipate, bounceThrow, impulseSlow, impulseOvershoot, overshoot, overshootElastic, overshootBouncy.

Curve engine primitives

For building custom curves:

import { cubicBezier, polybezier, pathPreset } from 'egaki/video'

// Analytic cubic-bezier y(x) solver
const ease = cubicBezier(0.5, 0, 0, 1)

// Multi-segment curve from control points
const throwCurve = polybezier([
  { x: 0, y: 0, upper: 0.7 },
  { x: 0.33, y: -0.2, lower: 0.8, upper: 0.8 },
  { x: 0.67, y: 1.3, lower: 0.1, upper: 0.1 },
  { x: 1, y: 1, lower: 0.8 },
])

// Intensity-parameterized preset
const myPreset = pathPreset({
  0: [
    { x: 0, y: 0, upper: 0 },
    { x: 0.4, y: 1.05, lower: 0.8, upper: 0.2 },
    { x: 1, y: 1, lower: 0.7 },
  ],
  100: [
    { x: 0, y: 0, upper: 0 },
    { x: 0.15, y: 1.8, lower: 0.8, upper: 0.1 },
    { x: 1, y: 1, lower: 0.8 },
  ],
})
const easing = myPreset(35)

LayoutTransition

<LayoutTransition id="x"> makes an element animate from its position in the previous section to its new position in the current section. Matching is by id.

# Scene 1 duration=3s

<LayoutTransition id="title">**Hello**</LayoutTransition>

# Scene 2 duration=3s

<LayoutTransition id="title" duration={25} bounce={0.2}>**Hello**</LayoutTransition>
<LayoutTransition id="subtitle">**World**</LayoutTransition>

Props: id (required), duration (frames, default 20), bounce (spring bounce 0-1, default 0.15), easing (custom easing function, overrides bounce), mode ('both' | 'position' | 'size', default 'both').

During the transition, LayoutTransition automatically interpolates border radius, background color, box shadow, and opacity between old and new elements.

Intra-scene transitions with `showFrom` / `showUpTo`

LayoutTransition also works within a single section. Use showFrom and showUpTo props (frame numbers) to create time-windowed instances of the same id. Only one instance is visible at a time; the element FLIP-animates between them.

# Active Item duration=6s

<div style={{ display: 'flex', flexDirection: 'column', gap: 48 }}>
  <div style={{ display: 'flex', alignItems: 'center', gap: 16 }}>
    <LayoutTransition id="dot" showFrom={0} showUpTo={2 * FPS} duration={18} bounce={0.2}>
      <Dot />
    </LayoutTransition>
    <span>First item</span>
  </div>
  <div style={{ display: 'flex', alignItems: 'center', gap: 16 }}>
    <LayoutTransition id="dot" showFrom={2 * FPS} showUpTo={4 * FPS} duration={18} bounce={0.2}>
      <Dot />
    </LayoutTransition>
    <span>Second item</span>
  </div>
</div>

Demo: example-layout-transition/ project.

`<Server>` component slots

<Server> is a reserved MDX element marking a subtree as React Server Components. Its children render server-side (async allowed, fs/API access) and get spliced into the client tree as a slot.

import { AsyncStats } from './async-stats'
import { TextToSpeech } from 'egaki/text-to-speech'

# Analytics duration=6s

<Opacity from={0} to={1} duration={15}>
  <Server>
    <AsyncStats />
    <TextToSpeech text="Analytics that build themselves." />
  </Server>
</Opacity>

Rules:

Files referenced inside <Server> execute in the RSC env. *.server.{ts,tsx} files are excluded from the client bundle automatically.
Each <Server> must start on its own line.
Inside <Server> normal RSC rules apply: no function props into client refs, no Remotion hooks (built-ins like Opacity work as client refs).
Bare specifiers work (import { TextToSpeech } from 'egaki/text-to-speech').

Generated media in `.server.tsx` files

GeneratedImage, GeneratedVideo, and GeneratedSpeech can be used inside .server.tsx files for type-safe composition with custom logic.

Import from egaki/generate-media (not egaki/video). The egaki/video exports are client stubs that return null.

// hero-scene.server.tsx
import { GeneratedImage } from 'egaki/generate-media'
import { Opacity, Fill } from 'egaki/video'

export async function HeroScene() {
  return (
    <Fill>
      <Opacity from={0} to={1} duration={20}>
        <GeneratedImage
          prompt="a magical forest with glowing mushrooms"
          seed={99}
          model="imagen-4.0-generate-001"
          style={{ width: '80%', margin: 'auto', borderRadius: 16 }}
        />
      </Opacity>
    </Fill>
  )
}

import { HeroScene } from './hero-scene.server'

# Scene duration=5s

<HeroScene />

Components imported from .server.* files are automatically wrapped in <Server> by the MDX parser. No manual <Server> block needed.

See generated-media-example/ for a working demo.

Media components

<Video> and <Audio> from @remotion/media are available in MDX. Both accept startInFrames (in frames) to offset playback.

Positive startInFrames: delays playback from section start. <Video src="/clip.mp4" startInFrames={1 * FPS} /> plays after 1 second.
Negative startInFrames: offsets from section end. <Audio src="/sfx.mp3" startInFrames={-2 * FPS} /> plays 2 seconds before end.

Gaps before and after audio

A gap before is a positive startInFrames. A gap after happens automatically when the clip is shorter than the section; use a negative startInFrames to anchor the clip to the cut, or trimAfter to cut the source early.

{/* 0.4s breath after the cut, then the voice starts */}
<GeneratedSpeech text="..." startInFrames={0.4 * FPS} />

{/* whoosh that ends exactly at the scene cut */}
<Audio src="/whoosh.mp3" startInFrames={-1 * FPS} />

{/* play only the first 2s of a longer clip */}
<Audio src="/tail.mp3" trimAfter={2 * FPS} />

`objectFit` for full-frame media

<Video> and <Img> from egaki/video support objectFit as a component prop:

<Fill>
  <Video src="/bg.mp4" muted loop objectFit="cover" />
</Fill>

Value	Behavior
`"cover"`	Fills the container, crops overflow
`"fill"`	Stretches to fill, ignores aspect ratio
`"contain"`	Fits inside with letterboxing
`"none"`	No resizing, centered at native size

Never use raw <img> tags. Always use <Img> from egaki/video, which calls delayRender() while the image loads. A raw <img> renders a blank frame during export.

Import rules

Do NOT import <Audio>, <Video>, or <OffthreadVideo> from remotion. Always import <Audio> and <Video> from @remotion/media.

Component	Import from	Works?
`<Audio>`	`@remotion/media`	Yes
`<Video>`	`@remotion/media`	Yes
`<Audio>` / `<Video>`	`remotion`	No (throws in web-renderer)
`<OffthreadVideo>`	`remotion`	No

`useAbsoluteCurrentFrame`

Remotion's useCurrentFrame() returns the frame relative to the current section (resets to 0 at each boundary). useAbsoluteCurrentFrame() returns the absolute frame across the entire composition.

import { useAbsoluteCurrentFrame } from 'egaki/video'
import { useVideoConfig } from 'remotion'

function GlobalTimer() {
  const absoluteFrame = useAbsoluteCurrentFrame()
  const { fps } = useVideoConfig()
  const elapsedSeconds = (absoluteFrame / fps).toFixed(1)
  return <span>{elapsedSeconds}s</span>
}

Use this for preamble overlays, cross-section sync, or global elapsed time.

`useTweakpane`

Registers tweakable parameters in a shared tweakpane pane (top-right in the player). When the component unmounts, its folder is removed.

import { useTweakpane } from 'egaki/video'

export function MyComponent(props: MyComponentProps) {
  const { children, style } = props

  const tp = useTweakpane('MyComponent', {
    blur: { value: props.blur ?? 12, min: 0, max: 50, step: 0.5 },
    visible: props.visible ?? true,
    label: props.label ?? 'Hello',
    color: props.color ?? '#ff0055',
    offset: props.offset ?? { x: 50, y: 25 },
  })

  return <div style={{ filter: `blur(${tp.blur}px)`, color: tp.color }}>
    {children}
  </div>
}

Pass props as defaults so tweakpane overrides them live. A "Copy changes" button serializes all modified params as structured markdown.

`resolveAssetPath`

Resolves file paths in server components so /image.png maps to {projectRoot}/public/image.png. Import from egaki/generate-media.

`cachedGenerate`

A higher-order function that wraps any async function with filesystem caching, deduplication, stale management, and progress tracking. All built-in generate functions use it. Available for user code via import { cachedGenerate } from 'egaki/cached-generate'.

import { cachedGenerate } from 'egaki/cached-generate'

const cachedExplore = cachedGenerate({
  namespace: 'explore',
  prefixFrom: (p) => p.query,
  generate: async (params) => fetchExploreApi(params.query),
  serialize: (result) => ({ json: result, extension: '.json' }),
  deserialize: ({ filePath }) => JSON.parse(fs.readFileSync(filePath, 'utf-8')),
})

Files go to public/generated/{namespace}/. Errors are not cached; failed calls retry on next invocation.

Captions

Word-by-word captions synced to narration audio. Workflow: transcribe audio to get word timestamps, convert to frame delays, render a Caption component.

Transcription

egaki transcribe audio.mp3 --model whisper-1

Models with word timestamps: whisper-1 (OpenAI), ink-whisper (Cartesia, cheapest), scribe_v1 (ElevenLabs), nova-3 (Deepgram), whisper-large-v3, whisper-large-v3-turbo, distil-whisper-large-v3-en (Groq, fastest).

Frame delays

Use each word's startSecond with FPS:

<Caption words={[
  { word: "Just", delay: 0 },
  { word: "quit", delay: 0.26 * FPS },
  { word: "your", delay: 0.48 * FPS },
]} />

Re-transcribe after regenerating TTS. When you regenerate audio, all word timestamps change. Always re-run egaki transcribe and update delays.

Default caption style

Film-style subtitles at the bottom of the frame:

<AbsoluteFill style={{
  display: 'flex',
  alignItems: 'flex-end',
  justifyContent: 'center',
  padding: '0 80px 120px',
}}>
  <span style={{
    fontSize: 42,
    fontWeight: 400,
    color: '#f5d442',
    fontFamily: '"Georgia", "Times New Roman", serif',
    textAlign: 'center',
    lineHeight: 1.4,
    letterSpacing: '0.01em',
    maxWidth: '70%',
  }}>
    {words.map((w, i) => (
      <span key={i} style={{ opacity: frame >= w.delay ? 1 : 0 }}>
        {i > 0 ? ' ' : ''}{w.word}
      </span>
    ))}
  </span>
</AbsoluteFill>

Style rules:

Georgia serif, fontWeight: 400, soft yellow #f5d442
42px for 1080x1920 vertical; scale proportionally for other resolutions
Bottom of screen via alignItems: 'flex-end' with padding-bottom: 120px
maxWidth: '70%' to prevent captions spanning full width
No text shadow, no uppercase, no fade animation; instant opacity: 0/1
No layout shift: render ALL words always, toggle visibility with opacity, never conditional rendering

Reference examples: example-sun-montage/ (film-style), example-bible-montage/ (editorial cascade with font rotation), example-captions/ (TikTok highlight style using @remotion/captions).

Voice cloning and TTS

Audio separation

egaki demucs song.mp3 --stems vocals,other -o stems/

Voice cloning

# Cartesia (default, instant, free, up to 10s)
egaki voice clone stems/vocals.mp3 --name "my-voice" --language en --json

# ElevenLabs (better for noisy audio)
egaki voice clone stems/vocals.mp3 --provider elevenlabs --name "my-voice" --remove-background-noise --json

TTS generation

egaki speech "Your text here." --voice <voice-id> --model sonic-3.5 -o public/voice.mp3

Speed control: Cartesia supports --speed 0.6 to 1.5. Default is 1.0.

Inserting pauses

Cartesia (sonic-3.5): Use SSML <break> tags.

Playwriter lets you control Chrome from code. <break time="400ms"/> Install the extension.

ElevenLabs (eleven_v3): Use plain-text pause tags.

But there's a catch. [long pause] Introducing Cloud Browsers.

Using TTS in videos

Place audio in public/ and reference with <Audio>. For voiceover over the whole video, put it in the preamble. For section narration, put it inside the section.

<Audio src="/voice-intro.mp3" />

# Scene duration=5s

<Caption words={[...]} />

Voiceover pacing

Give every section its own <GeneratedSpeech> and make sure the speech never overruns the section. Check with ffprobe:

for f in public/generated/audio/*.wav; do
  ffprobe -v error -show_entries format=duration -of csv=p=0 "$f"
done

Overrun → add a beat to the section or shorten the copy.
Dead air → shorten the section or lengthen the copy.
Don't fit VO with the TTS speed option: models treat it as a hint, so output duration is nonlinear and varies per generation. Don't use audio playbackRate either; it shifts pitch.
seed={2} rerolls a generation whose length landed wrong.
startInFrames={0.4 * FPS} adds a breath after a cut so narration doesn't run back to back across scenes.

Syncing animations to word timestamps

Transcribe the generated speech to drive typewriter reveals or per-word highlights:

egaki transcribe public/generated/audio/hook-*.wav -m scribe_v1 -o words.json

The output segments array has startSecond/endSecond per word. Feed those into your component and reveal each word at startSec * FPS. Re-transcribe whenever the VO text, speed, or seed changes.

DimOverlay pattern

When a preamble <Video> plays behind section content, use DimOverlay to darken and blur the video:

# Scene duration=10s

<Background>
  <div style={{ width: '100%', height: '100%', backdropFilter: 'blur(20px)' }}>
    <DimOverlay darkness={0.8} duration={22} />
  </div>
</Background>

See sun-montage-example/components.tsx for the implementation.

Masked word-by-word text reveal

A staggered slide-up animation where each word is clipped by an overflow: hidden container and slides into view with translateY:

function MaskedWordsText({ text, startSec }: { text: string; startSec: number }) {
  const frame = useCurrentFrame()
  const { fps } = useVideoConfig()
  return (
    <>
      {text.split(' ').map((word, i) => {
        const wordStart = (startSec + i * 0.061) * fps
        const progress = interpolate(frame, [wordStart, wordStart + 0.607 * fps], [0, 1], {
          easing: EASE.smooth,
          extrapolateLeft: 'clamp',
          extrapolateRight: 'clamp',
        })
        return (
          <span key={i}>
            <span style={{ display: 'inline-block', overflow: 'hidden', verticalAlign: 'top' }}>
              <span style={{ display: 'inline-block', transform: `translateY(${(1 - progress) * 100}%)` }}>
                {word}
              </span>
            </span>
            {i < text.split(' ').length - 1 ? ' ' : null}
          </span>
        )
      })}
    </>
  )
}

See testimonial-example/components.tsx for a working implementation.

Frosted glass with `backdrop-filter`

Use backdrop-filter: blur() with a semi-transparent backgroundColor. This works because egaki uses HtmlInCanvas which supports all CSS features.

<div style={{
  backdropFilter: 'blur(54.5px)',
  WebkitBackdropFilter: 'blur(54.5px)',
  backgroundColor: 'rgba(255, 255, 255, 0.13)',
  borderRadius: '6%',
}} />

Subpixel jitter prevention

Two rules when animating transform: scale() or translateX/Y():

Add willChange: 'transform' to promote the layer to its own compositor surface.
Round values to 3 decimal places: Math.round(value * 1000) / 1000.

const scale = interpolate(frame, [0, 90], [1, 1.35], {
  extrapolateLeft: 'clamp',
  extrapolateRight: 'clamp',
  easing: EASE.cinematic,
})
const s = Math.round(scale * 1000) / 1000

<div style={{
  transform: `scale(${s})`,
  transformOrigin: '60% 45%',
  willChange: 'transform',
}}>

Never use clamp on slow, continuous animations (scrolling, panning, drift). Clamping freezes the output at the boundary. Only use clamp for bounded animations (opacity 0-1, progress bars).

Light-leak overlay transitions

Overlay a short light-sweep clip (black background, bright streaks) at the end of each scene so the flash lands right at the cut. With mixBlendMode: 'screen' black becomes transparent and only the streaks overlay the scene; a contrast crush removes the grey veil from compression noise.

Find the clip's flash peak once with ffmpeg -i clip.mp4 -vf signalstats -f null - (look for max YAVG), then position the peak ~2 frames before the section end:

function LightSwipe({ src, peakSec }: { src: string; peakSec: number }) {
  const { durationInFrames, fps } = useVideoConfig()
  const from = Math.max(0, durationInFrames - Math.round(peakSec * fps) - 2)
  return (
    <Sequence from={from} layout="none">
      <Fill style={{ pointerEvents: 'none' }}>
        <Video src={src} muted style={{
          width: '100%', height: '100%', objectFit: 'cover',
          mixBlendMode: 'screen', filter: 'brightness(0.75) contrast(2.2)',
        }} />
      </Fill>
    </Sequence>
  )
}

Screen recordings in scenes

Prep recordings with ffmpeg before dropping them in: crop to the exact app window (no desktop background), retime with setpts to fit the section, and strip audio.

ffmpeg -i raw.mov -vf "crop=2048:1152:256:144,setpts=PTS/2.2,scale=1600:900" \
  -an -r 30 -c:v libx264 -crf 18 public/recordings/demo.mp4

Two presentations that work well:

Windowed: rounded corners, spring entrance, and an ambient glow on a black scene (dark drop shadows are invisible on black; use a faint colored boxShadow glow instead).
Full-bleed: objectFit: 'cover' filling the frame, wrapped in a slow scale drift so it never feels static.

Zoom and ambient drift

Zoom at a specific point

const scale = interpolate(frame, [0, 90], [1, 1.2], {
  extrapolateLeft: 'clamp',
  extrapolateRight: 'clamp',
  easing: EASE.cinematic,
})

<div style={{
  transform: `scale(${Math.round(scale * 1000) / 1000})`,
  transformOrigin: '55% 35%',
  willChange: 'transform',
}}>

Ambient drift

No scene should feel static. Layer a slow linear animation over the full scene duration on top of entrance effects. Keep magnitude small: +0.05 to +0.10 for scale, 10-30px for translate. Interpolate from a negative offset to 0 so the final frame is predictable.

// Entrance: fast zoom-in over 1.5s
const zoomIn = interpolate(frame, [0, 45], [1, 1.35], {
  extrapolateLeft: 'clamp',
  extrapolateRight: 'clamp',
  easing: EASE.cinematic,
})

// Ambient drift: slow linear scale over full 3s scene
const drift = interpolate(frame, [0, 90], [-0.08, 0], {
  extrapolateLeft: 'clamp',
  extrapolateRight: 'clamp',
})

const s = Math.round((zoomIn + drift) * 1000) / 1000

Premounting video sequences

When building montages in TSX, use <Series> with premountFor to preload the next clip before the cut:

import { Series } from 'remotion'
import { Video, Fill } from 'egaki/video'

<Series>
  <Series.Sequence durationInFrames={45}>
    <Fill><Video src="/clip-01.mp4" muted loop objectFit="cover" /></Fill>
  </Series.Sequence>
  <Series.Sequence premountFor={60} durationInFrames={45}>
    <Fill><Video src="/clip-02.mp4" muted loop objectFit="cover" /></Fill>
  </Series.Sequence>
</Series>

In MDX sections this is handled automatically (1 second premount).

Framer Motion integration

egaki auto-detects motion in your project and bridges it with Remotion's frame-based rendering. motion.div, springs, variants, staggered children, and keyframes work inside MDX sections. Animations are frame-deterministic and support backward scrubbing.

Limitations:

AnimatePresence / exit animations won't replay on backward seek
layout prop is not seekable
layoutId / shared layout does not work; use egaki's <LayoutTransition>
whileInView, whileHover, drag, tap are not deterministic
Zero overhead when motion is not installed

CSS rules

Never use viewport-relative units (vw, vh, vmin, vmax). Remotion compositions have a fixed pixel size; viewport units resolve against the browser window. Use %, px, or em instead.
All CSS features work including perspective, transform-style: preserve-3d, backdrop-filter, mask-image, mix-blend-mode, filter, clip-path. egaki uses HtmlInCanvas (Chromium's drawElementImage).
Export is Chromium-only. Keep the tab in the foreground during export (background tabs throttle requestAnimationFrame).
Scene-level <Scale> must wrap <Background> too, and must never go below 1 (e.g. use from={1} to={1.2}). Scaling below 1 shrinks the composition and exposes the black background behind it.

MDX LSP autocomplete

Every egaki video project can have MDX LSP support so built-in components get autocomplete with prop types.

@mdx-js/typescript-plugin installed as a devDependency.
tsconfig.json with the plugin registered:

{
  "compilerOptions": {
    "plugins": [{ "name": "@mdx-js/typescript-plugin" }]
  },
  "mdx": { "checkMdx": true },
  "include": ["**/*.ts", "**/*.tsx", "**/*.d.ts", "**/*.mdx"]
}

egaki-env.d.ts in the project root (auto-generated on first run):

import 'egaki/mdx-components'

VS Code: install the MDX extension. Zed: install zed-mdx and enable TypeScript in settings.

Agent skill

This package ships a skill file that teaches AI coding agents how and when to use it. Install it with:

npx -y skills add remorses/egaki

License

MIT

Name		Name	Last commit message	Last commit date
cli		cli
docs		docs
example-acme		example-acme
example-angled-screen		example-angled-screen
example-animated-search-bar		example-animated-search-bar
example-auto-duration		example-auto-duration
example-bible-montage		example-bible-montage
example-browser-new		example-browser-new
example-captions		example-captions
example-claude-fusion-launch		example-claude-fusion-launch
example-code-block		example-code-block
example-cut-in-motion		example-cut-in-motion
example-cylinder-text		example-cylinder-text
example-framer-motion		example-framer-motion
example-framer-spiral		example-framer-spiral
example-generated-media		example-generated-media
example-kaleidoscope		example-kaleidoscope
example-launch-video		example-launch-video
example-layout-transition		example-layout-transition
example-midjourney		example-midjourney
example-mirror		example-mirror
example-modular		example-modular
example-playwriter-cloud		example-playwriter-cloud
example-rajavarman		example-rajavarman
example-release-notes		example-release-notes
example-scrolling-transcript		example-scrolling-transcript
example-shader		example-shader
example-showcase		example-showcase
example-sun-montage		example-sun-montage
example-testimonial		example-testimonial
example-text-grid		example-text-grid
example-video		example-video
gateway		gateway
midjourney		midjourney
website		website
.gitattributes		.gitattributes
.gitignore		.gitignore
.npmrc		.npmrc
AGENTS.md		AGENTS.md
MEMORY.md		MEMORY.md
README.md		README.md
kontext-test.png		kontext-test.png
nano-banana-test.png		nano-banana-test.png
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
xai-test.png		xai-test.png

Latest commit

History

Folders and files

Repository files navigation

egaki

Install

Quick start

Quick start with ChatGPT auth

CLI examples

Generate an image from text

Edit with an input image

Inpainting with a mask

Generate multiple images

Control composition

Use Google Cloud billing via Vertex AI

Pipe image output to other tools

Generate a video from text

Generate with a cheap model

Image-to-video

Control resolution and aspect ratio

Generate multiple videos

Pipe video output to other tools

xAI Grok image generation

xAI Grok video generation

xAI Grok auth

Discover models and pricing

Advanced image examples

Model quick reference

Image models

Video models

Feature support by model family

Subscription and usage

CLI commands

BPM detection

Loudness measurement

Help

Auth and billing

MDX Video Framework

Getting started

Full project setup (optional)

How it works

Sections and duration

Auto-duration from media

FPS and BEAT scope variables

Beat-synced music

Preamble

Animation primitives

inline prop

Sequential animations

style prop for layout

cutInMotion

Composing enter and exit moves

Typewriter text

<Fill>

<AngledScreen>

Easing presets

springFromDuration and dspring

EASE presets

Continuous intensity functions

Curve engine primitives

LayoutTransition

Intra-scene transitions with showFrom / showUpTo

<Server> component slots

Generated media in .server.tsx files

Media components

Gaps before and after audio

objectFit for full-frame media

Import rules

useAbsoluteCurrentFrame

useTweakpane

resolveAssetPath

cachedGenerate

Captions

Transcription

Frame delays

Default caption style

Voice cloning and TTS

Audio separation

Voice cloning

TTS generation

`FPS` and `BEAT` scope variables

`inline` prop

`style` prop for layout

`cutInMotion`

`<Fill>`

`<AngledScreen>`

`springFromDuration` and `dspring`

`EASE` presets

Intra-scene transitions with `showFrom` / `showUpTo`

`<Server>` component slots

Generated media in `.server.tsx` files

`objectFit` for full-frame media

`useAbsoluteCurrentFrame`

`useTweakpane`

`resolveAssetPath`

`cachedGenerate`

Frosted glass with `backdrop-filter`