Description

An open source automated video creation tool for generating short-form video content. Short Video Maker combines text-to-speech, automatic captions, background videos, and music to create engaging short videos from simple text inputs.

This project is meant to provide a free alternative to heavy GPU-power hungry video generation (and a free alternative to expensive, third-party API calls). It doesn't generate a video from scratch based on an image or an image prompt.

The repository was open-sourced by the AI Agents A-Z Youtube Channel. We encourage you to check out the channel for more AI-related content and tutorials.

The server exposes an MCP and a REST server.

While the MCP server can be used with an AI Agent (like n8n) the REST endpoints provide more flexibility for video generation.

You can find example n8n workflows created with the REST/MCP server in this repository.

TOC

Getting started

• Requirements
• How to run the server
• Web UI
• Tutorial
• Examples

Usage

• Environment variables
• REST API
• Configuration options
• MCP

Info

• Features
• How it works
• Limitations
• Concepts
• Troubleshooting
• Deploying in the cloud
• FAQ
• Dependencies
• Contributing
• License
• Acknowledgements

Tutorial with n8n

Examples

Features

• Generate complete short videos from text prompts
• Text-to-speech conversion
• Automatic caption generation and styling
• Background video search and selection via Pexels
• Background music with genre/mood selection
• Serve as both REST API and Model Context Protocol (MCP) server

How It Works

Shorts Creator takes simple text inputs and search terms, then:

1. Converts text to speech using Kokoro TTS
2. Generates accurate captions via Whisper
3. Finds relevant background videos from Pexels
4. Composes all elements with Remotion
5. Renders a professional-looking short video with perfectly timed captions

Limitations

• The project only capable generating videos with English voiceover (kokoro-js doesn’t support other languages at the moment)
• The background videos are sourced from Pexels

General Requirements

• internet
• free pexels api key
• ≥ 3 gb free RAM, my recommendation is 4gb RAM
• ≥ 2 vCPU
• ≥ 5gb disc space

NPM

While Docker is the recommended way to run the project, you can run it with npm or npx. On top of the general requirements, the following are necessary to run the server.

Supported platforms

• Ubuntu ≥ 22.04 (libc 2.5 for Whisper.cpp)
  • Required packages: git wget cmake ffmpeg curl make libsdl2-dev libnss3 libdbus-1-3 libatk1.0-0 libgbm-dev libasound2 libxrandr2 libxkbcommon-dev libxfixes3 libxcomposite1 libxdamage1 libatk-bridge2.0-0 libpango-1.0-0 libcairo2 libcups2
• Mac OS
  • ffmpeg (brew install ffmpeg)
  • node.js (tested on 22+)

Windows is NOT supported at the moment (whisper.cpp installation fails occasionally).

Concepts

Scene

Each video is assembled from multiple scenes. These scenes consists of

1. Text: Narration, the text the TTS will read and create captions from.
2. Search terms: The keywords the server should use to find videos from Pexels API. If none can be found, joker terms are being used (nature, globe, space, ocean)

Getting started

Docker (recommended)

There are three docker images, for three different use cases. Generally speaking, most of the time you want to spin up the tiny one.

Tiny

• Uses the tiny.en whisper.cpp model
• Uses the q4 quantized kokoro model
• CONCURRENCY=1 to overcome OOM errors coming from Remotion with limited resources
• VIDEO_CACHE_SIZE_IN_BYTES=2097152000 (2gb) to overcome OOM errors coming from Remotion with limited resources

docker run -it --rm --name short-video-maker -p 3123:3123 -e LOG_LEVEL=debug -e PEXELS_API_KEY= gyoridavid/short-video-maker:latest-tiny

Normal

• Uses the base.en whisper.cpp model
• Uses the fp32 kokoro model
• CONCURRENCY=1 to overcome OOM errors coming from Remotion with limited resources
• VIDEO_CACHE_SIZE_IN_BYTES=2097152000 (2gb) to overcome OOM errors coming from Remotion with limited resources

docker run -it --rm --name short-video-maker -p 3123:3123 -e LOG_LEVEL=debug -e PEXELS_API_KEY= gyoridavid/short-video-maker:latest

Cuda

If you own an Nvidia GPU and you want use a larger whisper model with GPU acceleration, you can use the CUDA optimised Docker image.

• Uses the medium.en whisper.cpp model (with GPU acceleration)
• Uses fp32 kokoro model
• CONCURRENCY=1 to overcome OOM errors coming from Remotion with limited resources
• VIDEO_CACHE_SIZE_IN_BYTES=2097152000 (2gb) to overcome OOM errors coming from Remotion with limited resources

docker run -it --rm --name short-video-maker -p 3123:3123 -e LOG_LEVEL=debug -e PEXELS_API_KEY= --gpus=all gyoridavid/short-video-maker:latest-cuda

Docker compose

You might use Docker Compose to run n8n or other services, and you want to combine them. Make sure you add the shared network to the service configuration.

version: "3"

services:
  short-video-maker:
    image: gyoridavid/short-video-maker:latest-tiny
    environment:
      - LOG_LEVEL=debug
      - PEXELS_API_KEY=
    ports:
      - "3123:3123"
    volumes:
	    - ./videos:/app/data/videos # expose the generated videos

If you are using the Self-hosted AI starter kit you want to add networks: ['demo'] to the** short-video-maker service so you can reach it with http://short-video-maker:3123 in n8n.

Web UI

@mushitori made a Web UI to generate the videos from your browser.

You can load it on http://localhost:3123

Environment variables

🟢 Configuration

key	description	default
PEXELS_API_KEY	your (free) Pexels API key
LOG_LEVEL	pino log level	info
WHISPER_VERBOSE	whether the output of whisper.cpp should be forwarded to stdout	false
PORT	the port the server will listen on	3123

⚙️ System configuration

key	description	default
KOKORO_MODEL_PRECISION	The size of the Kokoro model to use. Valid options are fp32, fp16, q8, q4, q4f16	depends, see the descriptions of the docker images above ^^
CONCURRENCY	concurrency refers to how many browser tabs are opened in parallel during a render. Each Chrome tab renders web content and then screenshots it.. Tweaking this value helps with running the project with limited resources.	depends, see the descriptions of the docker images above ^^
VIDEO_CACHE_SIZE_IN_BYTES	Cache for  frames in Remotion. Tweaking this value helps with running the project with limited resources.	depends, see the descriptions of the docker images above ^^

⚠️ Danger zone

key	description	default
WHISPER_MODEL	Which whisper.cpp model to use. Valid options are tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large-v1, large-v2, large-v3, large-v3-turbo	Depends, see the descriptions of the docker images above. For npm, the default option is medium.en
DATA_DIR_PATH	the data directory of the project	~/.ai-agents-az-video-generator with npm, /app/data in the Docker images
DOCKER	whether the project is running in a Docker container	true for the docker images, otherwise false
DEV	guess! :)	false

Configuration options

key	description	default
paddingBack	The end screen, for how long the video should keep playing after the narration has finished (in milliseconds).	0
music	The mood of the background music. Get the available options from the GET /api/music-tags endpoint.	random
captionPosition	The position where the captions should be rendered. Possible options: top, center, bottom. Default value	bottom
captionBackgroundColor	The background color of the active caption item.	blue
voice	The Kokoro voice.	af_heart
orientation	The video orientation. Possible options are portrait and landscape	portrait
musicVolume	Set the volume of the background music. Possible options are low medium high and muted	high

Usage

MCP server

Server URLs

/mcp/sse

/mcp/messages

Available tools

• create-short-video Creates a short video - the LLM will figure out the right configuration. If you want to use specific configuration, you need to specify those in you prompt.
• get-video-status Somewhat useless, it’s meant for checking the status of the video, but since the AI agents aren’t really good with the concept of time, you’ll probably will end up using the REST API for that anyway.

REST API

GET /health

Healthcheck endpoint

curl --location 'localhost:3123/health'

{
    "status": "ok"
}

POST /api/short-video

curl --location 'localhost:3123/api/short-video' \
--header 'Content-Type: application/json' \
--data '{
    "scenes": [
      {
        "text": "Hello world!",
        "searchTerms": ["river"]
      }
    ],
    "config": {
      "paddingBack": 1500,
      "music": "chill"
    }
}'

{
    "videoId": "cma9sjly700020jo25vwzfnv9"
}

GET /api/short-video/{id}/status

curl --location 'localhost:3123/api/short-video/cm9ekme790000hysi5h4odlt1/status'

{
    "status": "ready"
}

GET /api/short-video/{id}

curl --location 'localhost:3123/api/short-video/cm9ekme790000hysi5h4odlt1'

Response: the binary data of the video.

GET /api/short-videos

curl --location 'localhost:3123/api/short-videos'

{
    "videos": [
        {
            "id": "cma9wcwfc0000brsi60ur4lib",
            "status": "processing"
        }
    ]
}

DELETE /api/short-video/{id}

curl --location --request DELETE 'localhost:3123/api/short-video/cma9wcwfc0000brsi60ur4lib'

{
    "success": true
}

GET /api/voices

curl --location 'localhost:3123/api/voices'

[
    "af_heart",
    "af_alloy",
    "af_aoede",
    "af_bella",
    "af_jessica",
    "af_kore",
    "af_nicole",
    "af_nova",
    "af_river",
    "af_sarah",
    "af_sky",
    "am_adam",
    "am_echo",
    "am_eric",
    "am_fenrir",
    "am_liam",
    "am_michael",
    "am_onyx",
    "am_puck",
    "am_santa",
    "bf_emma",
    "bf_isabella",
    "bm_george",
    "bm_lewis",
    "bf_alice",
    "bf_lily",
    "bm_daniel",
    "bm_fable"
]

GET /api/music-tags

curl --location 'localhost:3123/api/music-tags'

[
    "sad",
    "melancholic",
    "happy",
    "euphoric/high",
    "excited",
    "chill",
    "uneasy",
    "angry",
    "dark",
    "hopeful",
    "contemplative",
    "funny/quirky"
]

Troubleshooting

Docker

The server needs at least 3gb free memory. Make sure to allocate enough RAM to Docker.

If you are running the server from Windows and via wsl2, you need to set the resource limits from the wsl utility 2 - otherwise set it from Docker Desktop. (Ubuntu is not restricting the resources unless specified with the run command).

NPM

Make sure all the necessary packages are installed.

n8n

Setting up the MCP (or REST) server depends on how you run n8n and the server. Please follow the examples from the matrix below.

	n8n is running locally, using n8n start	n8n is running locally using Docker	n8n is running in the cloud
short-video-maker is running in Docker, locally	http://localhost:3123	It depends. You can technically use http://host.docker.internal:3123 as it points to the host, but you could configure to use the same network and use the service name to communicate like http://short-video-maker:3123	won’t work - deploy short-video-maker to the cloud
short-video-maker is running with npm/npx	http://localhost:3123	http://host.docker.internal:3123	won’t work - deploy short-video-maker to the cloud
short-video-maker is running in the cloud	You should use your IP address http://{YOUR_IP}:3123	You should use your IP address http://{YOUR_IP}:3123	You should use your IP address http://{YOUR_IP}:3123

Deploying to the cloud

While each VPS provider is different, and it’s impossible to provide configuration to all of them, here are some tips.

• Use Ubuntu ≥ 22.04
• Have ≥ 4gb RAM, ≥ 2vCPUs and ≥5gb storage
• Use pm2 to run/manage the server
• Put the environment variables to the .bashrc file (or similar)

FAQ

Can I use other languages? (French, German etc.)

Unfortunately, it’s not possible at the moment. Kokoro-js only supports English.

Can I pass in images and videos and can it stitch it together

No

Should I run the project with npm or docker?

Docker is the recommended way to run the project.

How much GPU is being used for the video generation?

Honestly, not a lot - only whisper.cpp can be accelerated.

Remotion is CPU-heavy, and Kokoro-js runs on the CPU.

Is there a UI that I can use to generate the videos

No (t yet)

Can I select different source for the videos than Pexels, or provide my own video

No

Can the project generate videos from images?

No

Dependencies for the video generation

Dependency	Version	License	Purpose
Remotion	^4.0.286	Remotion License	Video composition and rendering
Whisper CPP	v1.5.5	MIT	Speech-to-text for captions
FFmpeg	^2.1.3	LGPL/GPL	Audio/video manipulation
Kokoro.js	^1.2.0	MIT	Text-to-speech generation
Pexels API	N/A	Pexels Terms	Background videos

How to contribute?

PRs are welcome. See the CONTRIBUTING.md file for instructions on setting up a local development environment.

License

This project is licensed under the MIT License.

Acknowledgments

• ❤️ Remotion for programmatic video generation
• ❤️ Whisper for speech-to-text
• ❤️ Pexels for video content
• ❤️ FFmpeg for audio/video processing
• ❤️ Kokoro for TTS