I Tested MoneyPrinterTurbo, the 138k-Star Tool That Turns One Sentence Into a Short Video

You give it one sentence. It writes the script, finds footage, records the voiceover, burns in subtitles, adds music, and hands you a finished vertical video.

That's the promise of MoneyPrinterTurbo — 138,685 stars, 22,139 forks, MIT licensed, created in March 2024 and still pushed to this week (latest release v1.3.9, October 21). Promises are cheap, so I installed it on a headless Linux cloud VM and ran it end to end. Here is what actually happened — including the part that failed.

What it is

MoneyPrinterTurbo is a pipeline, not an editor. One topic goes in one end, a finished short video comes out the other:

1. Script — an LLM writes the voiceover text from your topic
2. Footage — stock clips matched by keyword, or your own local files
3. Voice — text-to-speech reads the script
4. Subtitles — speech recognition turns the audio back into timed captions
5. Assembly — footage, voice, subtitles and background music are composed into a 9:16 HD video

It has four entry points: WebUI, API, CLI, and an AI-agent interface. I used the CLI.

The design decision that matters most: every stage is replaceable. Don't like the auto-generated script? Paste your own. Don't want stock footage? Point it at local files. Don't want the robot voice? Supply your own audio. This test leans on that hard, as you'll see.

The test setup

• Headless Linux cloud VM, Python 3.12.3, ffmpeg already installed
• Official pip route, in a fresh venv
• No paid API keys anywhere; data directory redirected with MPT_HOME so nothing touched global config

What worked

Install — one command, after one gotcha.

pip install mpt   # installed mpt 1.3.9, matching the latest release

The gotcha: this machine's /tmp is a 512 MB tmpfs. pip unpacks some large wheels (the mpt package itself is ~97 MB) and died with No space left on device — while the home directory had 137 GB free. Worse, the first background install reported success while having installed nothing. Setting TMPDIR to a directory on the big disk fixed it. If you run this on a small cloud box, you will probably hit the same wall.

The full pipeline run — zero keys.

For the successful run I supplied all three inputs myself: my own script, three local images as footage, and my own audio file:

mpt cli \
  --video-script "..." \
  --video-source local \
  --video-materials materials/mat1.jpg,materials/mat2.jpg,materials/mat3.jpg \
  --video-aspect 9:16 \
  --custom-audio-file narration.mp3 \
  --bgm-type none

Result: task status SUCCESS, and a finished final-1.mp4. I verified the output with ffprobe rather than trusting the success message:

• 1080×1920 (9:16 vertical)
• 15.07 seconds
• Audio track present
• The still images were automatically given a slow zoom (Ken Burns) effect and stitched in sequence

So the core claim holds: the pipeline really does assemble a watchable vertical video with no human editing — at least when you feed it your own materials.

What failed

The built-in free voiceover (Edge TTS) never connected.

My first run used the default free TTS voice. The script and footage stages passed; the job then died at the audio stage. Edge TTS could not reach Microsoft's speech endpoint from this VM (speech.platform.bing.com:443, SSL error), retried three times, and gave up.

I want to be precise about what this means: it's a network problem of this cloud environment, not proof the feature is broken — on a normal home connection it usually works. But in my test it did not work, so I won't write "free voiceover: tested, works." I routed around it with my own audio file, which is exactly what the replaceable-stage design is for.

What I did not test

Honesty section, because a pipeline is only as good as its untested links:

• Auto scriptwriting — needs an LLM API key from one of the supported providers. Skipped; I supplied the script.
• Online stock footage (Pexels/Pixabay) — free, but needs a registered API key. Skipped; I used local files.
• Whisper subtitles — runs locally but downloads a ~3 GB model. Not downloaded this time, so the test video has no subtitles.
• WebUI and one-click cross-platform posting — not started. (On posting: automated publishing runs into each platform's rules and risk controls; my own practice is to publish manually.)
• AI-generated footage — paid sources, and the CLI requires an explicit charge confirmation flag. Untouched.

In other words: the fully automatic chain — topic in, video out, nothing supplied by you — depends on an LLM key + a footage key + TTS connectivity, and I did not verify that whole chain in this environment. What I verified is the self-supplied mode, end to end.

The cost ledger

The path I tested cost exactly $0: open-source software (MIT, commercial use allowed), my own script, local images, an existing audio file, no accounts registered.

The costs appear only where you ask for more automation: an LLM key for auto-scripting (some providers have free tiers), a free-but-registered key for stock footage, ~3 GB of disk for the subtitle model, and real money only if you opt into AI-generated footage.

Verdict

Worth it if you produce informational or list-style short videos at volume and your bottleneck is assembling materials — scripting, sourcing, voicing, captioning — rather than editing craft. Start in full self-supplied mode, get the pipeline running, then automate stage by stage. Each stage you automate is a stage that can break; this way you choose which risks to take.

Not for you if you need edited storytelling, transitions and pacing (the output is standardized assembly), or if you refuse to touch a command line (there is a WebUI, but I didn't test it, so I won't vouch for it).

The formula I'd take away:

Short-video throughput = templated scripts + local footage library + a pipeline where every stage can be swapped out.

The day the voiceover stage failed and I still shipped a finished video is the whole argument for that last term.

Tested 2026-11-04 on mpt 1.3.9. Star/fork counts read from the GitHub repo page the same day.