MoneyPrinterTurbo Setup: AI Short Videos From One Keyword

MoneyPrinterTurbo is a free, open source tool that turns one topic or keyword into a finished short video: an LLM writes the script, the app pulls matching stock footage, generates a voiceover and burned-in subtitles, and composites everything with ffmpeg. It added roughly 2,306 stars in a single day on the way to 108,475 total, making it the fastest-climbing repository on GitHub trending right now. Setup takes about ten minutes: a portable package on Windows, or one clone and two commands on macOS, Linux and Docker.

• The current release is v1.3.4, published August 12, 2026. Its only prebuilt asset is a 900 MB portable Windows package; every other platform builds from source.
• There are four ways to drive it: a WebUI on port 8501, a REST API on port 8080, a plain CLI, and an agent skill file you can hand to a coding agent.
• You bring your own model key. Kimi, OpenAI, Gemini, DeepSeek, Qwen, Grok, MiniMax and local Ollama are all supported, and the default voice engine, Edge TTS, needs no key at all.
• It is MIT licensed with 16,472 forks and only 30 open issues, which is an unusually clean tracker for a repository this size.

What is MoneyPrinterTurbo and why is it trending?

The repository first appeared in March 2024 as a Chinese-language take on the short-video automation idea, and it has been iterating quietly ever since: v1.3.0 in June 2026, then four point releases through August 12. What it does has not changed much. You type a subject, it asks a language model for a script and a set of stock-footage search terms, fetches clips that match those terms, synthesizes a voiceover, times subtitles against that voiceover, drops in background music and hands ffmpeg the edit list. Ten minutes later you have a vertical or horizontal MP4.

RelatedOpenCut Setup: Self-Host the Free CapCut Alternative

What is climbing is the surface area around that core. Version 1.3.x added one-click cross-platform publishing, an agent skill document that lets a coding agent install and drive the whole thing from a single prompt, and a provider list that reads like a census of the 2026 model market: Kimi, OpenAI, Gemini, DeepSeek, Qwen, Azure OpenAI, VolcEngine Ark, Grok, MiniMax and MiMo directly, plus Ollama, LiteLLM, OneAPI, Groq and Cloudflare AI Gateway. The 16,472 forks tell you most of the audience is not using it as shipped so much as bending it into a pipeline of their own.

How do you install it on Windows?

Windows is the only platform with a prebuilt binary, and it is the fastest route by a wide margin. Download the portable package from the releases page: the v1.3.4 asset is MoneyPrinterTurbo-Portable-Windows-1.3.4.7z and it weighs 900 MB, because it bundles a Python runtime along with the app. Extract it, then run the two batch files in order.

# in the extracted folder, double-click in this order
update.bat
start.bat

The README is explicit that update.bat comes first, because the packaged code lags the repository. After start.bat the browser opens on its own; if the page comes up blank, the project recommends Chrome or Edge. One trap worth respecting: the docs warn against project paths containing non-ASCII characters, special characters or spaces, so extract somewhere short and plain. From a source checkout the launcher is a batch file in the project root instead, and it falls back to uv run streamlit when it cannot find a project Python.

# from a source checkout, in the project root
.\webui.bat

How do you install it on macOS and Linux?

There is no binary for either, so you clone and build. The project moved to pyproject.toml as its primary manifest and ships a uv.lock, so uv is the documented path. Python 3.11 or later is required and 3.11 is what the docs use.

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

If you have not adopted uv, the older path still works; the README keeps requirements.txt around for exactly this and calls it legacy.

python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Then launch the WebUI from the project root. The script picks the virtual environment or uv and finds a free port by itself. To reach it from a phone or another machine, set the host variable first.

sh webui.sh

# reachable from other devices on your LAN
MPT_WEBUI_HOST=0.0.0.0 sh webui.sh

What if you would rather run it in Docker?

Docker is the option to pick on a home server, or anywhere you do not want a Python toolchain on the host. Copy config.example.toml to config.toml before the first start, because that file gets mounted into the containers. The release compose file pulls a prebuilt image from GitHub Container Registry, ghcr.io/harry0703/moneyprinterturbo:latest.

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

The WebUI then answers on http://127.0.0.1:8501 and the API on http://127.0.0.1:8080/docs or /redoc. Plain docker compose up still builds the image from source if you need a local build.

How do you generate a video without a browser?

Two ways. There is a real CLI, which matters on a headless box or in a cron job, and the simplest complete invocation is one flag.

uv run python cli.py --video-subject "How AI is changing everyday life"
uv run python cli.py --help

There is also a standalone API service, for when something else in your stack is the orchestrator.

uv run python main.py

The newest option is stranger and more interesting: the project ships an agent skill document, so if your coding agent can read a skill file and run a terminal, you hand it a URL and a topic and it does the install, the configuration and the generation, then returns a file path. It asks only for API keys that are not already configured. The README says this path currently supports macOS and Windows.

RelatedSuperpowers Setup: Give Your AI Coding Agent a Real Workflow

Use this Skill: https://raw.githubusercontent.com/harry0703/MoneyPrinterTurbo/main/docs/skill/SKILL.md
Create a video with the topic "How AI is changing everyday life."

How does it compare with the hosted AI video generators?

Read that table honestly. The hosted products are more finished and somebody answers the phone. What you get here is control: a swappable model, your own footage, output that never touches a vendor pipeline, and a scriptable interface. For one video, the hosted tools win on effort. For two hundred, or when script generation has to run through one specific model, this is the shape that scales.

What are the gotchas before you rely on it?

Start with ffmpeg. The app normally downloads and detects it for you, but in constrained environments it fails with RuntimeError: No ffmpeg exe could be found. The fix is to install ffmpeg yourself and point the config at the binary, remembering that Windows paths in TOML need doubled backslashes.

[app]
# set to your actual path; Windows separators are doubled
ffmpeg_path = "C:\\Users\\harry\\Downloads\\ffmpeg.exe"

Subtitles are the second thing to know. The default provider is edge, which derives timings from the TTS engine, runs fast and needs no GPU. Switching to whisper gives a more accurate timeline but downloads roughly 3 GB of large-v3 weights from Hugging Face on first use, and that download is one of the most common failures documented in the README. The smaller large-v3-turbo model is about 1.6 GB and is selected in the same block.

[app]
subtitle_provider = "whisper"

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

Three more. On Linux and macOS, heavier runs can hit OSError: [Errno 24] Too many open files; check your limit with ulimit -n and raise it with ulimit -n 10240. The bundled background music in resource/songs comes from YouTube videos, and the README itself says to delete it if there are copyright issues, which is a real consideration if you plan to monetise anything you make. And the cross-platform publishing feature routes through a third-party service called Upload-Post, which means an account, an API key, and that key sitting in plaintext in config.toml next to your model credentials.

The last gotcha is not technical. The name promises something the tool does not deliver: stock footage stitched to a synthetic voiceover is a commodity format, platforms keep tightening their rules on mass-produced repetitive uploads, and the ceiling on that content is low. The tool is genuinely good at what it does. What it does is remove the production bottleneck, not the ideas one.

Our take

The useful way to see MoneyPrinterTurbo is not as a content business in a box but as a well-built orchestration layer that happens to be free. The language model is yours, the footage belongs to Pexels or your hard drive, the voice comes from Edge TTS, and the video work is ffmpeg. What the project contributes is glue, and glue this good is hard to write: keyword extraction that returns footage matching the script, subtitle timing that lands on the syllable, an edit list ffmpeg accepts without arguing.

That also tells you who should install it. If you want a video, use a hosted tool and be done in five minutes. If you want a repeatable pipeline you control, this is the best free option on GitHub today. Just go in understanding that the interesting part was never the rendering. It was always whether you have something worth saying two hundred times.

Original analysis by GenZTech. Tool documentation: harry0703/MoneyPrinterTurbo on GitHub.