gem-hunter
Language: TypeScript
AI-powered podcast & video generator.
If you want to try our beta version, follow the instruction in the release note below.
Note: The video version of "What is MulmoCast?" is available on X, which is created by MulmoCast itself.
MulmoCast is a next-generation presentation platform, purpose-built for a world where AI and humans collaborate to create and share ideas.
Unlike traditional tools like PowerPoint or Keynote—both designed in the pre-AI era for human creators—MulmoCast is AI-native: reimagined from the ground up to work hand-in-hand with generative AI that understands and produces natural language, images, audio, and video.
As AI takes on an increasingly central role in content creation, it needs more than raw capability—it needs an environment designed for its strengths. MulmoCast provides that environment, empowering AI to generate rich, coherent presentations across multiple formats, while keeping humans in the creative loop.
At the core of MulmoCast is MulmoScript, a JSON-based intermediate language that functions like a screenplay or web markup. Generated by large language models, MulmoScript defines everything from narrative structure to visuals, making it easy to produce presentations that adapt seamlessly across formats.
flowchart TD
S_TextEditor["Creator"] <--> S_AIChat2("LLM")
S_AIChat2 --> MulmoScript["MulmoScript"]
MulmoScript --> MulmoCast["MulmoCast"]
MulmoCast --> O_Video("Video") & O_Podcast("Podcast") & O_Slideshow("Slideshow") & O_PDF("PDF") & O_Manga("Manga") & O_SwipeAnime("Swipe Anime")
MulmoScript@{ shape: div-proc}
MulmoCast@{ shape: trap-t}
S_TextEditor:::StoryInput
S_AIChat2:::ScriptInput
O_Video:::ScriptOutput
O_Podcast:::ScriptOutput
O_Slideshow:::ScriptOutput
O_PDF:::ScriptOutput
O_Manga:::ScriptOutput
O_SwipeAnime:::ScriptOutput
classDef StoryInput stroke-width:1px, stroke-dasharray:none, stroke:#46EDC8, fill:#DEFFF8, color:#378E7A
classDef ScriptInput stroke-width:1px, stroke-dasharray:none, stroke:#374D7C, fill:#E2EBFF, color:#374D7C
classDef ScriptOutput stroke-width:1px, stroke-dasharray:none, stroke:#FBB35A, fill:#FFEFDB, color:#8F632D
In today’s world, people consume information through many channels—on Zoom calls, in Slack threads, via videos on the go, or through podcasts during commutes. Presentations can no longer be confined to a single format or device. That’s why MulmoCast is built to be truly multi-modal, enabling output not only as slide decks but also as videos, podcasts, manga-style comics, and PDF documents.
Whether you're building a podcast from a research report or turning a business pitch into a narrated video, MulmoCast transforms source material into dynamic, multi-format content with the help of state-of-the-art AI like ChatGPT and Claude.
MulmoCast isn't just a tool. It’s a new foundation for communication—one where humans and AI co-create across every mode of expression.
MulmoScript is a simple JSON/YAML format for describing multi-modal content.
You can define speakers, text, images, and layout — all in one script.
Here is the "Hello World" in MulmoScript.
{
"$mulmocast": {
"version": "1.1"
},
"beats": [
{ "text": "Hello World" }
]
}See MulmoScript Format for details on the structure.
npm install -g mulmocast
You'll also need to install ffmpeg:
# For macOS with Homebrew brew install ffmpeg # For other platforms # Visit https://ffmpeg.org/download.html
You can also use Dockerfile which helps you install the pre-requisits.
docker build -t mulmo-cli .
You can use the Docker image like this:
docker run -e OPENAI_API_KEY=<your_openai_api_key> -it mulmo-cli mulmo tool scripting -i -t children_book -o ./ -s story
Create a .env file in your project directory with the following API keys:
OPENAI_API_KEY=your_openai_api_key
GEMINI_API_KEY=your_google_gemini_api_key
# For Anthropic Claude (htmlPrompt feature)
ANTHROPIC_API_TOKEN=your_anthropic_api_tokenFor htmlPrompt configuration, see docs/image.md.
REPLICATE_API_TOKEN=your_replicate_api_key
# For ElevenLabs TTS
ELEVENLABS_API_KEY=your_elevenlabs_api_keyBROWSERLESS_API_TOKEN=your_browserless_api_token # to access web in mulmo toolMULMOCAST_DUMP_USAGE=1 # print JSON usage dump to stdout after each CLI action MULMOCAST_DUMP_USAGE=/tmp/usage.json # ...or write it to a file
Token / per-second / per-char usage per provider:model. See docs/api.md for the full reference.
For enterprise/production environments, or to use Veo movie models that are only available on Vertex AI, use Vertex AI with Application Default Credentials (ADC):
# Install gcloud CLI and authenticate
gcloud auth application-default loginConfigure in MulmoScript:
{
"imageParams": {
"provider": "google",
"model": "gemini-2.5-flash-image",
"vertexai_project": "your-project-id",
"vertexai_location": "us-central1"
}
}| Parameter | Description | Default |
|---|---|---|
vertexai_project |
Google Cloud Project ID | None (enables Vertex AI mode when set) |
vertexai_location |
Region | us-central1 |
For detailed setup instructions, see Vertex AI Setup Guide.
To use Azure OpenAI instead of OpenAI API:
# For image generation IMAGE_OPENAI_API_KEY=<your-azure-openai-api-key> IMAGE_OPENAI_BASE_URL=https://<resource-name>.openai.azure.com/ # For TTS (Text-to-Speech) TTS_OPENAI_API_KEY=<your-azure-openai-api-key> TTS_OPENAI_BASE_URL=https://<resource-name>.openai.azure.com/ # For LLM (translate, scripting) LLM_OPENAI_API_KEY=<your-azure-openai-api-key> LLM_OPENAI_BASE_URL=https://<resource-name>.openai.azure.com/ LLM_OPENAI_API_VERSION=2025-04-01-preview # optional
MulmoScript configuration (same as OpenAI):
{
"imageParams": {
"provider": "openai",
"model": "gpt-image-1.5"
},
"speechParams": {
"speakers": {
"Presenter": {
"provider": "openai",
"voiceId": "alloy",
"model": "tts"
}
}
}
}Important: Azure deployment names must match model names exactly (e.g., deployment name gpt-image-1.5 for model gpt-image-1.5).
For detailed setup and region availability, see Azure OpenAI Usage Guide.
Create a mulmo.config.json file to set project-wide defaults. The CLI searches for it in CWD first, then home directory (~/).
{
"speechParams": {
"provider": "elevenlabs"
},
"imageParams": {
"provider": "google"
},
"audioParams": {
"bgm": { "kind": "path", "path": "assets/bgm.mp3" },
"bgmVolume": 0.15
}
}Top-level keys are applied as defaults (script values take precedence). Use the override key to force values over scripts — useful for enterprise branding or TTS provider enforcement:
{
"speechParams": {
"provider": "elevenlabs"
},
"override": {
"speechParams": {
"provider": "elevenlabs",
"model": "eleven_multilingual_v2",
"speakers": {
"Presenter": { "provider": "elevenlabs", "voiceId": "Rachel" }
}
}
}
}Priority chain: config (defaults) < template/style < script < config.override < presentationStyle (-p)
Note:
kind: "path"entries inmulmo.config.jsonare resolved relative to the script file directory, not the config file location. This is consistent with all other path resolution in MulmoScript.
Verify the merged result with:
mulmo tool info merged --script <script.json>
mulmo tool scriptingmulmo audiomulmo imagesmulmo movieStep 1-1: Run the script generation command
mulmo tool scripting -i -t children_book -o ./ -s story
This will initiate the script creation process.
Step 1-2: Interactive conversation with AI
After running the command, you'll engage in an interactive conversation with the AI to create your story script. The AI will guide you through the process and help shape your content based on your inputs. Once completed, a JSON file (like story-1746600802426.json) will be generated.
mulmo movie {generated_script_file}Replace {generated_script_file} with the output file from STEP 1, such as story-1746600802426.json.
Optionally, you can specify __clipboard as the script file name to paste the script from the clipboard, which is generated by other interactive environment such as ChatGPT or Claude.
Click the image above to watch an example of what you can create
Verify your .env file contains:
OPENAI_API_KEY=your_openai_api_key
Note: Ensure your OpenAI organization is verified to access the GPT Image models. Visit https://platform.openai.com/settings/organization/general and complete the "Verifications" section.
mulmo tool scripting -i -t ghibli_comic -o ./ -s story
This will initiate an interactive conversation with the AI to create your Ghibli-inspired story. Once completed, a JSON file (e.g., story-1747834931950.json) will be generated.
mulmo translate {generated_script_file_from_step_2}mulmo movie {generated_script_file_from_step_2} -c jaThe -c ja flag adds Japanese subtitles to your video.
When the process completes, the CLI will display the path to your generated video file:
Video created successfully! 14.939 sec
writing: /Users/username/path/to/output/story-1747834931950__ja.mp4# Generate script from web content (requires Browserless API KEY) mulmo tool scripting -u https://example.com # Generate script from local file mulmo tool scripting --input-file story.txt # Generate script with interactive mode mulmo tool scripting -i
Note:
You can create a minimal JSON with just beats and complete it with mulmo tool complete:
{
"beats": [{ "text": "Hello, welcome to MulmoCast!" }, { "text": "This is a simple example." }]
}Save this as my_beats.json, then complete it with a style:
# Complete with built-in style mulmo tool complete my_beats.json -s ani # Complete with custom style file mulmo tool complete my_beats.json -s ./my_style.json # Output to specific file mulmo tool complete my_beats.json -s ani -o my_script.json
This generates a complete MulmoScript with all required fields (canvasSize, speechParams, imageParams, etc.) from the style.
Mulmo provides several commands to handle different aspects of content creation:
# Generate audio from script mulmo audio script.json # Generate images for script mulmo images script.json # Generate both audio and images, then combine into video mulmo movie script.json # Translate script to Japanese mulmo translate script.json
Any of the commands above can dump a JSON breakdown of upstream AI consumption when MULMOCAST_DUMP_USAGE is set:
# print JSON to stdout (pipe-friendly: `... | jq .`) MULMOCAST_DUMP_USAGE=1 mulmo audio script.json # ...or write it to a file MULMOCAST_DUMP_USAGE=/tmp/usage.json mulmo movie script.json
Payload shape:
{
"records": 3,
"byModel": [
{ "provider": "openai", "model": "gpt-4o-mini-tts", "records": 1, "inputChars": 20 },
{ "provider": "gemini", "model": "gemini-3.8-flash-tts", "records": 1, "inputTokens": 9, "outputTokens": 59, "totalTokens": 68 },
{ "provider": "elevenlabs", "model": "eleven_multilingual_v2", "records": 1, "inputChars": 18 }
],
"snapshot": [/* per-call UsageRecord[] */]
}byModel groups by provider:model (different models have different rate cards, so summing across them isn't meaningful). The full per-call snapshot is also included so a billing layer can apply its own grouping. See docs/api.md § Usage tracking for the per-agent field-population matrix and the programmatic (library) API.
Every generation command accepts --estimate: it prints what the command would consume — scoped to that command's pipeline — and exits without generating:
mulmo movie script.json --estimate # table per provider:model + total cost mulmo audio script.json --estimate --json # raw UsageEstimate[] records (pipe-friendly)
mulmo audio --estimate covers TTS (plus translation when -l differs from the script language); mulmo images / mulmo pdf cover image, htmlImage, movie, sound-effect, lip-sync and reference-media generation; mulmo movie covers everything; mulmo translate only translation. Values marked ~ are heuristic estimates; unmarked values are exact for the given script.
The library API can estimate the same thing — per beat, per process — without calling any API:
import { estimateUsage } from "mulmocast"; const estimates = estimateUsage(script, { targetLangs: ["ja"] }); // [{ process: "tts", beatIndex: 0, provider: "openai", model: "gpt-4o-mini-tts", // inputChars: { value: 50, precision: "exact" }, outputTokens: { value: 150, precision: "estimated" }, // costUSD: 0.0019, pricingAsOf: "2026-07-03" }, ...]
Each metric carries a precision flag: "exact" for values that are deterministic from the script (TTS character counts, tokenized OpenAI prompts, fixed image-output token tables) and "estimated" for heuristics (LLM output length, speech duration derived from text). costUSD appears when pricing data exists in modelPricing (src/types/provider2agent.ts), where every price records the date it was last verified. See docs/api.md § Pre-run estimation.
HTML-based beats (html_tailwind, markdown, textSlide, chart, mermaid, slide, captions, HTML prompts) and PDF output are rendered in a headless browser that normally has full network access. For a script from an untrusted source, add --strict-network (or set MULMO_STRICT_NETWORK=1; library callers set context.strictNetwork):
mulmo images untrusted.json --strict-network
The renderer may then load only files in the script's folder and mulmocast's output folders, data: / blob: URLs, and the CDNs mulmocast's own templates use (cdn.tailwindcss.com, fonts.googleapis.com, fonts.gstatic.com, and on cdn.jsdelivr.net only the mermaid and Chart.js files the templates load). Anything else — other local files, remote images in the script, fetch, WebSockets, WebRTC, popups, preconnect hints — is blocked, and each blocked request is logged (strict network: blocked …). Files next to the script in the same folder remain readable, so keep an untrusted script in a folder of its own (a script loaded by URL gets the output folders only). Requests to the allowed CDNs still reach those providers, so treat what a page sends there as visible to them. Strict mode guards what the browser loads; files the script names with kind: "path" (images, movies, logos, mermaid or remotion code) are read by mulmocast itself and are not restricted yet (#1622). Remotion beats are always guarded this way. The vision beat renders in its own browser inside mulmocast-vision, which the guard cannot reach, so in strict mode a vision beat stops with an error instead of rendering unguarded.
When running the same mulmo command multiple times, previously generated files are treated as cache. For example, audio or image files will not be regenerated if they already exist.
To force regeneration, delete the old files — including temporary files — under the output directory before re-running the command or use -f (force) option.
If you modify the text or instruction fields in a MulmoScript, mulmo will automatically detect the changes and regenerate the corresponding audio content upon re-run.
MulmoCast includes a powerful Slide DSL (type: "slide") for creating structured presentation slides with JSON. Slides are rendered via Tailwind CSS + Puppeteer into images.
chartjs-chart-sankey) and treemap (chartjs-chart-treemap) auto-loaded by chart typeSet a theme once with slideParams.theme, then use "type": "slide" in each beat:
{
"$mulmocast": { "version": "1.1" },
"slideParams": {
"theme": {
"colors": {
"bg": "0F172A",
"bgCard": "1E293B",
"bgCardAlt": "334155",
"text": "F8FAFC",
"textMuted": "CBD5E1",
"textDim": "64748B",
"primary": "3B82F6",
"accent": "8B5CF6",
"success": "22C55E",
"warning": "F59E0B",
"danger": "EF4444",
"info": "14B8A6",
"highlight": "EC4899"
},
"fonts": { "title": "Georgia", "body": "Calibri", "mono": "Consolas" }
}
},
"beats": [
{
"text": "Welcome to the presentation",
"image": {
"type": "slide",
"slide": { "layout": "title", "title": "Main Title", "subtitle": "Subtitle" }
}
}
]
}Or use a preset presentation style:
mulmo tool complete beats.json -s slide_dark -o presentation.jsonAvailable preset styles: slide_dark, slide_pop, slide_warm, slide_creative, slide_minimal, slide_corporate
For detailed layout specifications and content block reference, see the Slide DSL documentation.
MulmoCast includes 100 pre-designed CSS styles for markdown slides, organized in 10 categories:
| Category | Description | Example Styles |
|---|---|---|
| business | Professional corporate designs | corporate-blue, executive-gray, finance-green |
| tech | Technology and developer themes | cyber-neon, matrix-green, terminal-dark |
| creative | Artistic and expressive styles | artistic-splash, watercolor-soft, graffiti |
| minimalist | Clean and simple designs | clean-white, zen-beige, nordic-light |
| nature | Nature-inspired themes | forest-green, ocean-blue, sunset-orange |
| dark | Dark mode and elegant themes | charcoal-elegant, midnight-blue, obsidian |
| colorful | Vibrant and bold colors | vibrant-pink, candy-colors, aurora |
| vintage | Retro and classic styles | retro-70s, art-deco, typewriter |
| japanese | Japanese aesthetic designs | washi-paper, sakura-pink, zen-garden |
| geometric | Pattern-based designs | hexagon-pattern, isometric, bauhaus |
Add the style property to a markdown beat:
{
"beats": [
{
"text": "Corporate presentation slide",
"image": {
"type": "markdown",
"markdown": "# Quarterly Report\n\n## Key Highlights\n\n- Revenue up 15%\n- New market expansion\n- Team growth",
"style": "corporate-blue"
}
}
]
}Use the mulmo tool info command to see all available styles:
mulmo tool info styles
MulmoScript is a JSON format to define podcast or video scripts: Schema definition: schema.ts
https://github.com/receptron/mulmocast-cli/tree/main/scripts
| Directory | Description |
|---|---|
output/ |
Artifacts generated by commands — e.g., .json, audio, and video files |
output/audio/ |
Temporary audio fragments used in the final audio assembly |
output/images/ |
Image files generated by the images command |
output/cache/ |
Internal cache files for various processing steps |
These directories are automatically created as needed by the
mulmocommands.
CLI Usage
mulmo <command> [options]
Commands:
mulmo translate <file> Translate Mulmo script
mulmo audio <file> Generate audio files
mulmo images <file> Generate image files
mulmo movie <file> Generate movie file
mulmo pdf <file> Generate PDF files
mulmo markdown <file> Generate markdown files
mulmo bundle <file> Generate bundle files
mulmo html <file> Generate html files
mulmo tool <command> Generate Mulmo script and other tools
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
mulmo translate <file>
Translate Mulmo script
Positionals:
file Mulmo Script File [string] [required]
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
-o, --outdir output dir [string]
-b, --basedir base dir [string]
-l, --lang target language
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
-f, --force Force regenerate [boolean] [default: false]
-g, --grouped Output all files under output/<basename>/ directory
[boolean] [default: false]
--backup create backup media file [boolean] [default: false]
--strict-network Let HTML renderers reach only local files and the
CDNs mulmocast's templates use (also
MULMO_STRICT_NETWORK=1) [boolean] [default: false]
-p, --presentationStyle Presentation Style [string]
mulmo audio <file>
Generate audio files
Positionals:
file Mulmo Script File [string] [required]
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
-o, --outdir output dir [string]
-b, --basedir base dir [string]
-l, --lang target language
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
-f, --force Force regenerate [boolean] [default: false]
-g, --grouped Output all files under output/<basename>/ directory
[boolean] [default: false]
--backup create backup media file [boolean] [default: false]
--strict-network Let HTML renderers reach only local files and the
CDNs mulmocast's templates use (also
MULMO_STRICT_NETWORK=1) [boolean] [default: false]
-p, --presentationStyle Presentation Style [string]
-a, --audiodir Audio output directory [string]
mulmo images <file>
Generate image files
Positionals:
file Mulmo Script File [string] [required]
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
-o, --outdir output dir [string]
-b, --basedir base dir [string]
-l, --lang target language
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
-f, --force Force regenerate [boolean] [default: false]
-g, --grouped Output all files under output/<basename>/ directory
[boolean] [default: false]
--backup create backup media file [boolean] [default: false]
--strict-network Let HTML renderers reach only local files and the
CDNs mulmocast's templates use (also
MULMO_STRICT_NETWORK=1) [boolean] [default: false]
-p, --presentationStyle Presentation Style [string]
-i, --imagedir Image output directory [string]
mulmo movie <file>
Generate movie file
Positionals:
file Mulmo Script File [string] [required]
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
-o, --outdir output dir [string]
-b, --basedir base dir [string]
-l, --lang target language
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
-f, --force Force regenerate [boolean] [default: false]
-g, --grouped Output all files under output/<basename>/ directory
[boolean] [default: false]
--backup create backup media file [boolean] [default: false]
--strict-network Let HTML renderers reach only local files and the
CDNs mulmocast's templates use (also
MULMO_STRICT_NETWORK=1) [boolean] [default: false]
-p, --presentationStyle Presentation Style [string]
-a, --audiodir Audio output directory [string]
-i, --imagedir Image output directory [string]
-c, --caption Video captions
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
mulmo pdf <file>
Generate PDF files
Positionals:
file Mulmo Script File [string] [required]
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
-o, --outdir output dir [string]
-b, --basedir base dir [string]
-l, --lang target language
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
-f, --force Force regenerate [boolean] [default: false]
-g, --grouped Output all files under output/<basename>/ directory
[boolean] [default: false]
--backup create backup media file [boolean] [default: false]
--strict-network Let HTML renderers reach only local files and the
CDNs mulmocast's templates use (also
MULMO_STRICT_NETWORK=1) [boolean] [default: false]
-p, --presentationStyle Presentation Style [string]
-i, --imagedir Image output directory [string]
--pdf_mode PDF mode
[string] [choices: "slide", "talk", "handout"] [default: "slide"]
--pdf_size PDF paper size (default: letter)
[choices: "letter", "a4"] [default: "letter"]
mulmo markdown <file>
Generate markdown files
Positionals:
file Mulmo Script File [string] [required]
Options:
--version Show version number [boolean]
-v, --verbose verbose log [boolean] [required] [default: false]
-h, --help Show help [boolean]
-o, --outdir output dir [string]
-b, --basedir base dir [string]
-l, --lang target language
[string] [choices: "en", "ja", "fr", "es", "de", "zh-CN", "zh-TW", "ko", "it",
"pt", "ar", "hi"]
-f, --force Force regenerate [boolean] [default: false]
-g, --grouped Output all files under output/<basename>/ directory
[boolean] [default: false]
--backup create backup media file [boolean] [default: false]
--strict-network Let HTML renderers reach only local f