MoneyPrinterTurbo: Open-Source AI Video Generator

Where short-video automation stands after Sora left the room

MoneyPrinterTurbo is an open source AI short video generator with a blunt contract: hand it a topic or a keyword, get back a finished HD video, script, footage, voiceover, subtitles and background music included. It offers four ways in (an AI agent, a WebUI, an API and a bare command line), it does not ask for a GPU at the entry level, and the code itself is free under an MIT license. The meter lives elsewhere: in the LLM, voice and stock-footage APIs you plug into it. With Sora exiting the consumer market in 2026, self hosted pipelines of this kind are the obvious next stop.

First, the obvious question: why now

Sora is done, at least the public version of it. The app shut down on April 26, 2026, and the API is scheduled to follow on September 24, 2026, per the Wikipedia entry on Sora. Whatever anyone thought of closed video models, a door closed, and anyone who had a daily hosted-generator habit is shopping for a new one.

That is the context a repository named MoneyPrinterTurbo walks into. The name does its own marketing, and the pitch under it is crude on purpose: one topic in, one TikTok-ready or Shorts-ready video out. No editor, no timeline, no afternoons lost on stock sites. The pipeline is the product.

And the money question is the right one, because the pricing shape here is unusual. The code is free under an MIT license, as the Wavect 2026 review of the project notes, and the same review adds the part the name glosses over: LLM calls, optional voice services, media APIs, compute and storage still carry a meter. Zero license fee, pay per API call. That is the whole business case, and the math beats the hype.

How MoneyPrinterTurbo turns a keyword into a video

The assembly line, in order

The project’s README states the pipeline in one breath: provide a video topic or keyword and it generates the script, matches footage, creates subtitles and background music, and produces an HD short video. Broken into stations, the line runs like this:

Script. A chosen LLM writes it, or you paste a custom one. Multilingual script generation is supported.
Keywords. The model extracts the search terms that drive the footage hunt.
Footage. Stock libraries, your own local assets, or freshly generated clips.
Voice. TTS narration with real-time previews.
Subtitles. Styled and aligned to the voice, with fonts, position, color, size, outline and background all configurable.
Music. Random or custom background music, volume adjustable.
Composition and render. Per the Verdent builder guide, composition runs on MoviePy and rendering on FFmpeg.

Output comes in two shapes: portrait 9:16 at 1080×1920 and landscape 16:9 at 1920×1080. Batch generation makes several videos in one run so you keep the one you like, and a clip-duration setting controls how briskly the footage switches.

Four interfaces, one engine

An AI Agent workflow: an agent that reads the project’s Skill file installs, configures and generates, asking only for API keys it does not already have. Currently macOS and Windows.
A WebUI for the browser crowd, served on port 8501 in the Docker setup.
An API, with documentation on port 8080 under /docs and /redoc.
A pure CLI when there is no browser or port forwarding: uv run python cli.py --video-subject "How AI is changing everyday life" (the project’s own example). The full switch reference sits behind uv run python cli.py --help.

Practical extras: generation settings export and import as a preset file, and every API key can be backed up and restored from the settings dialog. Anyone who has lost a Saturday to re-entering keys recognises the value.

Where the footage comes from, stock or synthetic

Three stock APIs are wired in: Pexels, Pixabay and Coverr, all free-to-use HD sources. The Verdent guide identifies Pexels as the primary source and adds the sensible warning: confirm license terms and attribution before commercial use. Your own local assets slot into the same stage.

MoneyPrinterTurbo: Open-Source AI Video Pipeline

Then there is the second supply line: generated footage. WaveSpeed AI text-to-video models (Seedance by default) mint new visuals from the script’s keywords instead of borrowing them from a library. Stock when stock is good enough, synthetic when the keywords describe something no library carries.

The hardware bill, itemized

Minimum published spec: four CPU cores and 4 GB of RAM. GPU: not required.MoneyPrinterTurbo documentation

Item	Minimum	Recommended	Optimal
CPU	4 cores	6 to 8 cores	8+ cores
RAM	4 GB	8 GB	16+ GB
GPU	Not required	4+ GB VRAM	8+ GB VRAM

Read the GPU row twice, because it is the load-bearing one. No GPU is required to run. A GPU is recommended if you want faster local transcription, faster video processing or smoother batch generation. The documentation’s own guidance splits the cases cleanly: if you lean on cloud LLMs, cloud TTS and online material sources, CPU and RAM matter more than GPU; if you run faster-whisper and heavy batch locally, a GPU lifts throughput noticeably. Recommended platforms are Windows 10+, macOS 11+ or a mainstream Linux distribution, with Python 3.11 or later (3.11 recommended).

Voices, subtitles and the two subtitle modes

The default voice costs nothing: Edge TTS, shown as Azure TTS V1 in the WebUI, asks for no API key. Past that, the roster runs Azure TTS V2 (credentials come from the Azure portal), SiliconFlow TTS, Google Gemini TTS, Xiaomi MiMo TTS, ElevenLabs TTS, a self-hosted Chatterbox TTS, and a no-voice mode for people who subtitle in silence. Everything is selectable with real-time previews, and the WebUI walks you through whichever credentials a provider wants.

Subtitles come in two gears:

edge: built from TTS timestamps. Fast, no GPU, and the default.
whisper: local faster-whisper transcription for a more accurate timeline. The model downloads on first use: roughly 3 GB for the default large-v3, or the smaller and faster large-v3-turbo at roughly 1.6 GB.

[app]
subtitle_provider = "whisper"

[whisper]
model_size = "large-v3-turbo"

The first whisper run pulls the model from Hugging Face. If that automatic download fails, the project’s fix is manual: drop whisper-large-v3 into .\MoneyPrinterTurbo\models so the final path reads .\MoneyPrinterTurbo\models\whisper-large-v3.

Edge TTS

The free default voice, labelled Azure TTS V1 in the WebUI. No key, real-time preview.

faster-whisper

The local transcription engine behind the whisper subtitle mode. Downloads its model once, then works offline.

Upload-Post

The third-party service whose API key turns on one-click publishing to TikTok, Instagram and YouTube Shorts.

The economics: free code, metered pipeline

Back to the money, because with a product named MoneyPrinter someone has to do the arithmetic. The MIT license means the download costs nothing, and the Wavect review lists what does not disappear: LLM calls, optional voice services, media APIs, compute and storage. The bill moves, it does not vanish.

There is one escape route from the cloud meter at the script stage: run a local model through Ollama and skip the cloud API entirely, as described in the Verdent guide. The same guide supplies the honest counterweight: a model small enough to run comfortably on consumer hardware produces weaker scripts than a frontier cloud model, and for short-video scripts the difference is visible. Trade confirmed, choose knowingly.

What the hosted competition charges

For a flat-fee comparison point, OpenShorts positions itself as an open source clip generator with a hosted cloud tier: paid plans start at $12 per month for 100 minutes with no watermark, up to $59 per month, per the OpenShorts site. Different product (it turns long videos into clips rather than scripting from a topic), same shelf in the store. The honest frame: a flat subscription with capped minutes against a bring-your-own-keys pipeline where every video quietly spends API cents. Per-video cost depends entirely on which providers you attach, so run that math once rather than meeting it in the invoice.

Deploying it with Docker

Docker on Windows and Linux

Before the first start, copy config.example.toml to config.toml so it can be mounted into the containers. Then two lines:

cd MoneyPrinterTurbo
docker compose -f docker-compose.release.yml up

The release compose file pulls the prebuilt image ghcr.io/harry0703/moneyprinterturbo:latest; a plain docker compose up builds the image locally instead. Docker Desktop is the prerequisite, and Windows users are pointed at Microsoft’s WSL documentation. Then the browser: the WebUI on http://127.0.0.1:8501, and the API documentation on http://127.0.0.1:8080/docs or http://127.0.0.1:8080/redoc.

The manual path with uv

git clone https://github.com/harry0703/MoneyPrinterTurbo.git
cd MoneyPrinterTurbo
uv python install 3.11
uv sync --frozen

pyproject.toml is the primary dependency manifest, uv.lock pins the resolved environment (hence the frozen sync), and requirements.txt survives only for legacy pip installs where venv + pip still works. On Windows, avoid project paths containing non-ASCII characters, special characters or spaces. Launch the WebUI with webui.bat on Windows (CMD works too) or sh webui.sh on macOS and Linux; the scripts fall back to uv run streamlit when no project Python is found, pick a free local port and open the browser. For LAN access, set MPT_WEBUI_HOST=0.0.0.0 first. Launch the API with uv run python main.py, or plain python main.py when the environment is already activated.

Windows carries one more shortcut: the one-click package from the project documentation points to GitHub Releases. Extract the package, double-click update.bat to pull the latest code, then start.bat to launch. If the browser opens blank, Chrome or Edge behave better.

The AI agent and Colab shortcuts

If your AI agent can read Skill documents and drive a local terminal, the project’s agent prompt hands it the Skill file and a topic; the agent installs, configures and returns the video file path, asking only for required API keys that are missing. macOS and Windows for now. And when installing anything sounds like work, the project ships a Google Colab notebook: open the notebook in Colab and run the pipeline in someone else’s machine room.

The fine print nobody reads until it matters

Music. Tracks live in resource/songs, and the project page carries a warning that deserves to survive every copy: the current project includes some default music from YouTube videos, and if there are copyright issues, please delete them. Straight from the source, and the only correct response is to build your own folder before monetising anything. Subtitle fonts live in resource/fonts, and you can add your own.

Two operational landmines get their own entries on the project page. The RuntimeError: No ffmpeg exe could be found case (ffmpeg normally auto-downloads; when it cannot, set ffmpeg_path or the IMAGEIO_FFMPEG_EXE environment variable), and the OSError: [Errno 24] Too many open files case, solved by raising the system limit with ulimit -n 10240.

Credits, because the project names them and accuracy should too: the page thanks Kimi (Moonshot AI), BytePlus ModelArk, CCSub, Infistar.ai, RecCloud and Picwish as sponsors. RecCloud earns a practical mention: it runs a free online AI Video Generator built on this project, so people allergic to deployment get a hosted front door.

So, does the printer actually print

The promise is an assembly line, and on the documentation it holds: one keyword in, an HD vertical video out, with batch mode for volume. The caveats are stamped just as clearly: API meters, weaker scripts when the LLM runs locally, stock licenses that need reading before commercial use, and a bundled soundtrack that arrives with a delete-if-copyright note attached. What the tool does honestly is move the bill from a license line to a usage line. Anyway. If it prints for you, good luck to it. At least this printer shows the meter.

📊 What to remember before you commit a weekend

Full pipeline: one keyword produces script, footage, voiceover, subtitles and music, at 9:16 (1080×1920) or 16:9 (1920×1080), with batch generation and a clip-duration control.
Cost shape: MIT-licensed code costs nothing; the meter lives in LLM, voice and media APIs, and local Ollama models trade money for visibly weaker scripts, per the Wavect review and the Verdent guide respectively.
Hardware: minimum four cores and 4 GB of RAM, no GPU required; a GPU starts paying once whisper subtitles and batch generation enter the picture.
Deployment: Docker in two commands, WebUI on 8501, API docs on 8080; one-click Windows package, uv on macOS and Linux, plus a Colab notebook for the zero-install trial.
Fine print: the default songs are YouTube-sourced with a copyright caveat attached, and stock footage still needs a license check before commercial use.

Additional Resources

This article includes content created with AI.

MoneyPrinterTurbo: Open-Source AI Video Pipeline