Type a Topic, Get a Video: Inside a 140,000-Star Automation

MoneyPrinterTurbo has 140,000 GitHub stars, a Chinese-language README, and a promise that fits in one sentence: give it a topic or a keyword, and it writes the script, matches footage, adds subtitles and background music, and renders a high-definition short video.
The result is a complete production pipeline, and for a certain kind of channel it is the whole job. The project was built by harry0703, a Chinese developer, and it has become one of the most-starred repositories in the Chinese open source scene, sitting alongside the agent tooling that floods the same charts.
Installing it takes about two minutes, and you can run it from a browser, from the command line, from an API or from an AI agent.
Installation, four ways
The official package on PyPI is named mpt, maintained by harry0703. The MoneyPrinterTurbo name on PyPI belongs to a third party, which is the kind of detail worth checking before you pip install something.
```shell python3 -m venv .venv source .venv/bin/activate python -m pip install mpt mpt webui ```
Python 3.11 or newer, with 3.11 recommended. The WebUI opens in your browser automatically. To pin a port and skip the browser, use mpt webui --port 8503 --headless. Before generating anything from the CLI, mpt cli --help lists the options.
Other routes exist for different needs. Windows users get a one-click bundle for trying it fast. Source installs and secondary development use uv for dependency management. Docker is the path when you want an isolated environment. Google Colab has a notebook for people who would rather not configure anything locally.
The fourth route is the interesting one. The repository ships an agent skill document, so an agent that can read a skill file and run shell commands will install the project, configure the API keys it needs and generate the video on its own. The README's example prompt is basically the whole workflow.
What the pipeline does
The project exposes four interfaces: AI Agent, WebUI, API and CLI, so it works both as a quick demo and as a component in an automated flow. It supports batch generation of multiple finished videos, keeps a task history, and can import, export and restore both generation settings and API keys.
Footage comes from your own local images and video, or from the free libraries at Pexels, Pixabay and Coverr. Subtitles are generated, and so is the background music.
The model layer is where it gets crowded in a useful way. Kimi, OpenAI, Anthropic Claude, Google Gemini, DeepSeek, Alibaba Tongyi, Azure OpenAI, Volcano Ark, xAI Grok, MiniMax and Xiaomi MiMo are all supported for script and text work. For visuals, Volcano Ark Seedance is wired in natively to generate video shots from script segments, and the sponsor integrations cover GPT Image 2.5, Seedream for cover images, and Wan, Seedance and MiniMax H3 for video.
It also speaks to the aggregator layer: OpenRouter, LiteLLM, OneAPI, Cloudflare AI Gateway, ModelScope, Ollama, Groq, AIHubMix, EvoLink, Pollinations and a Claude Code subscription can all stand in for a direct provider key.
Configuration lives in a config.toml created on first launch, with API keys entered in the WebUI. Data goes to ~/Library/Application Support/MoneyPrinterTurbo on macOS, %LOCALAPPDATA%\MoneyPrinterTurbo on Windows and ~/.local/share/MoneyPrinterTurbo on Linux. Set MPT_HOME to move it. Finished videos land in storage/tasks.
The business model is visible in the README
Nearly every sponsor block sells API access. APIMart at $0.006 per GPT Image 2 image, Infistar with image generation from 0.06 yuan per picture, MiniMax H3 video at 0.09 yuan per second for 768p, RunningHub reselling Seedance, Ofox bundling Seedance and Wan behind one key. Kimi offers new users a 10 percent API credit rebate through a tracked link.
The tool is free and MIT licensed, and the project is a well-built front end for other people's models. Its maintainer is being paid, or at least paid in referrals, by making it easy to spend money on tokens. Assume that when you read the pricing notes.
What it costs to run
The economics are worth spelling out, because the reason this project spread is the price of the output.

Stock footage from Pexels, Pixabay and Coverr costs nothing. A script generated by a good text model costs a fraction of a cent. Subtitles and a music bed are local operations. What remains is video generation, and that is the only line item that moves the bill: Seedance, Wan and MiniMax H3 are all wired in, and the sponsor integrations advertise rates like 0.21 yuan per second, 0.09 yuan per second for 768p and 0.15 yuan per second for 2K.
A ten-second clip at those rates lands under a cent. A two-minute video with a dozen shots lands somewhere between a few cents and a couple of dollars, depending entirely on which model renders the footage. The pipeline itself adds nothing to the cost, because it is local code on your machine.
Nobody improved the writing. The marginal cost of another finished video fell to roughly the price of the API calls, and that is what changed mass short-form production.
The workflow details that matter in practice
A few features separate a demo from something you would actually use. Batch generation matters, because the useful case is twenty variants of the same script, not one video. The task history matters, because when something renders wrong you need to find which run produced it. Export and restore of settings and API keys matters more than it sounds, because a configuration that cannot be moved between machines is a configuration you cannot share with a teammate.
The model coverage matters too, and it is broader than the marketing suggests. Seedance through Volcano Ark generates video straight from script segments, so the pipeline can produce its own footage rather than only assembling library clips. Script writing can run on Kimi, Claude, Gemini or DeepSeek depending on which key you have and what the script needs. Aggregators like OpenRouter and LiteLLM sit in the middle, so one key can reach most of the model list.
For a team, that flexibility is the point. The video model changes every quarter. The script model changes every quarter. The pipeline that holds them together does not.
The part nobody solves
The pipeline solves throughput, not taste. A stock-footage montage with machine-written narration and auto subtitles is exactly the content people mean when they complain about AI sludge. You can produce ten of these in the time it takes to watch one and make well.
The workflow therefore needs a human somewhere in it, and the README is honest that the human part is yours. The useful version of this tool is the boring one: a draft generator you edit, where the script, the stock choice and the subtitle wording get your judgment before anyone sees the result.
Used that way, it is a genuinely good time saver for anyone who has ever spent an evening matching clips to a paragraph. Used as an unattended content farm, it is a faster way to produce sludge, and it has 140,000 stars because both use cases exist.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.