| name | moneyprinterturbo-expert |
| description | Expert knowledge of the MoneyPrinterTurbo codebase: the full AI short-video generation pipeline (LLM script -> search terms -> TTS audio -> subtitles -> stock-video sourcing -> clipping/concat -> final compositing), its FastAPI service, Streamlit WebUI, configuration system, and deployment. Use when working on, debugging, extending, or explaining anything in this repository: adding LLM/TTS/video-source providers, fixing subtitle timing, changing video composition, configuring config.toml, task/state management, or Docker setup. |
MoneyPrinterTurbo Expert
MoneyPrinterTurbo generates complete short videos (9:16, 16:9 or 1:1) from a
single subject string: an LLM writes the script, TTS narrates it, subtitles are
aligned to the speech, stock footage is fetched and cut to match the audio, and
everything is composited with MoviePy/FFmpeg into final-{n}.mp4.
Mental model (read this first)
One orchestrator drives everything: app/services/task.py::start(task_id, params, stop_at="video").
It runs the stages below in order, persisting progress into a state store after
each one. stop_at checkpoints (script, terms, audio, subtitle,
materials, video) let callers run a partial pipeline (this is how the
/audio and /subtitle API endpoints reuse the same code path).
| # | Stage | Service / key function | Output |
|---|
| 1 | Script | llm.py::generate_script() | plain-text script |
| 2 | Search terms | llm.py::generate_terms() | JSON list of English terms |
| 3 | TTS audio | voice.py::tts() | audio.mp3 + SubMaker timings |
| 4 | Subtitles | voice.py::create_subtitle() or subtitle.py::create() (whisper) | subtitle.srt |
| 5 | Materials | material.py::download_videos() | cached stock clips |
| 6 | Combine | video.py::combine_videos() | combined-{n}.mp4 |
| 7 | Composite | video.py::generate_video() | final-{n}.mp4 |
All task artifacts live in storage/tasks/{task_id}/. Stock-video cache lives
in storage/cache_videos/ (configurable via material_directory).
The single source of truth for request parameters is
app/models/schema.py::VideoParams (Pydantic). Enums: VideoAspect
(16:9→1920x1080, 9:16→1080x1920, 1:1), VideoConcatMode
(random/sequential), VideoTransitionMode (Shuffle, FadeIn, FadeOut,
SlideIn, SlideOut). Task states in app/models/const.py: -1 failed,
4 processing, 1 complete.
Entry points
- WebUI:
streamlit run webui/Main.py (port 8501). i18n JSON in
webui/i18n/ (de, en, pt, ru, tr, vi, zh).
- API:
python main.py → uvicorn/FastAPI on port 8080 (app/asgi.py,
routes in app/router.py + app/controllers/v1/video.py).
- CLI:
cli.py.
Reference files (load on demand)
- pipeline.md — deep dive into every stage: LLM prompt
engineering, all TTS providers and SubMaker subtitle alignment, whisper
fallback + Levenshtein correction, material sourcing, clip
selection/letterboxing/transitions, FFmpeg concat, final compositing and
subtitle rendering.
- configuration.md — full
config.toml reference:
20+ LLM providers, TTS/voice naming conventions, video sources and API keys,
whisper, redis, proxy, codec settings.
- api-and-webui.md — REST endpoints, task queue and
concurrency limits, memory vs redis task managers, state store, Streamlit UI
structure and i18n.
- patterns.md — reusable engineering techniques this repo
teaches: thread-safe API key rotation, hardware-codec fallback with runtime
detection, TTS streaming timeout via daemon thread + queue, duration
estimation for no-voice mode, subtitle/script alignment, rounded subtitle
backgrounds with PIL.
- recipes.md — step-by-step guides: run locally / via
Docker, add a new LLM provider, add a new TTS provider, add a new stock
video source, add a transition effect, debug common failures.
- examples.md — copy-paste-ready practical examples: CLI
invocations, full API request/poll/download flows, driving the pipeline
from Python, config.toml presets (free / local / production), subtitle
styling and prompt-customization payloads, quota-free smoke tests.
- local-setup-guide.md — dispatch guide (in
Portuguese) for an agent installing and running everything on the user's
PC: Docker vs native Python paths per OS, free API keys, config, WebUI
launch and visual verification, troubleshooting. Use when the task is
"set this up on my machine".
Conventions and gotchas (always relevant)
- Config is read once from
config.toml at repo root by
app/config/config.py; copy config.example.toml to start. Never commit
real API keys.
PYTHONPATH must include the repo root; FFmpeg binary resolution order is
IMAGEIO_FFMPEG_EXE env var → system PATH → bundled imageio_ffmpeg.- Subtitle fonts come from
resource/fonts/ (default STHeitiMedium.ttc);
BGM from resource/songs/ (bgm_type = "random" picks one at random).
- Docker images patch the ImageMagick
policy.xml — required for MoviePy text
rendering; replicate this in any new image.
- File uploads/downloads must go through
app/utils/file_security.py::resolve_path_within_directory() to prevent
path traversal. Keep this invariant when adding endpoints.
- Hardware encoders (
h264_nvenc, h264_amf, h264_qsv, ...) are optional;
every write path must keep the automatic libx264 fallback
(video.py::_write_videofile_with_codec_fallback).
- Tests live in
test/services/; run with pytest from the repo root.
