Porter Workflow

English | 简体中文

Porter Workflow is an automated video localization pipeline: give it a video URL, get back a standardised asset bundle plus hard-subbed release videos with bilingual and Chinese-only subtitles.

It ships as one engine with three frontends:

All three call the same porter library. The engine contains no argument parsing and never writes to stdout — a requirement, because in an MCP stdio server stdout is the JSON-RPC channel.

Status: v0.2.x is the current development line on main. The v0.2 rewrite — one engine, three frontends — has landed. The v0.1 implementation is preserved in git history at commit b5fd577.

Installation

Requires Python ≥ 3.11.

# CLI, minimal install (pure-Python; translation works without a key, but
# transcription needs the [asr-local] extra or an API key -- see below)
uvx porter-workflow "<URL>"

# CLI with every optional backend
uvx --from "porter-workflow[all]" porter "<URL>"

# Persistent install
pipx install "porter-workflow[all]"

System dependencies

These are not Python packages and must be on PATH:

Check everything at once:

porter doctor

Transcription is key-free and offline, or an API key

Local Whisper runs on your own machine: no key, no network after the first run. Measured 2026-09-24 on real speech -- 18 s of audio became 3 cues in 7.2 s on CPU with the default small model:

whisper-local is tried first, so it is what a run uses whenever it is installed. It sits behind an extra because faster-whisper and ctranslate2 are heavy, and CTranslate2-backed rather than PyTorch:

# Key-free and offline. This is the recommended install.
uvx --from "porter-workflow[asr-local]" porter "<URL>"

# Or everything at once -- [all] includes [asr-local]
uvx --from "porter-workflow[all]" porter "<URL>"

The first run downloads the model from Hugging Face (464 MB for the default small); after that it is fully offline. Device and precision are chosen for you -- CUDA if it loads, CPU otherwise -- and can be pinned:

porter config set asr.whisper_local_model=medium
porter config set asr.whisper_local_device=cuda
porter config set asr.whisper_local_compute_type=float16

A bare uvx porter-workflow "<URL>" still cannot transcribe, because no ASR backend is in a minimal install: it downloads and standardises the video and then fails with every speech-to-text backend failed. The two key-free endpoints are the ones that are dead -- both are reverse-engineered, and Google Web's response is malformed at the HTTP level, so the read fails outright rather than merely returning nothing. Bcut's failure is the more ambiguous of the two -- the host answers, so it could be a quota or a changed field rather than a dead endpoint -- which is why it keeps an "unverified" label in the source.

So to get subtitles, do one of:

# Option A (recommended): local, key-free, offline
uvx --from "porter-workflow[asr-local]" porter "<URL>" --burn skip

# Option B: an LLM key (also improves translation quality a lot)
export OPENAI_API_KEY=sk-...
porter "<URL>" --burn skip

# Option C: install VideoCaptioner and use its engines
pip install videocaptioner
porter "<URL>" --asr-engine bijian --burn skip

Translation is unaffected. Bing, Google and MyMemory all still work without a key; each was probed against its live endpoint and returned real Chinese. Only transcription ever needed credentials.

porter doctor reports which route a job will actually take before you start it.

Optional extras

videocaptioner is deliberately not a dependency. It is GPL-3.0 (this project is MIT) and pins python<3.13. Porter detects it at runtime and, if present, uses it as an extra ASR/translation backend across a process boundary. Install it yourself if you want those engines:

pip install videocaptioner

Slow or blocked? Mirrors for China

Both of porter's downloads default to endpoints that are painful from mainland China: the packages come from PyPI, and the Whisper weights from Hugging Face. On a machine that looks Chinese -- by timezone, by a UTC+8 offset, or by a zh locale -- porter switches to domestic mirrors on its own:

The faster-whisper-* repositories on ModelScope hold the official CTranslate2 weights rather than a re-conversion: their config.json is byte-identical to Systran/faster-whisper-small's. Each model size is downloaded from there once.

To force a direction, or to use a mirror of your own choosing:

PORTER_MIRROR=cn     # always prefer the mirrors
PORTER_MIRROR=off    # never; always use the official sources

An index you set yourself always wins and is never overwritten. Setting UV_DEFAULT_INDEX (or the older UV_INDEX_URL) is enough on its own. PIP_INDEX_URL is honoured as well, and is copied across to uv -- because uv ignores PIP_INDEX_URL by itself, so pip config set global.index-url <mirror> alone would silently do nothing at all.

Usage

# Inspect a link without downloading anything
porter inspect "<URL>"

# Full pipeline: download → transcribe → translate → burn
porter "<URL>" -o ./porter_output

# Subtitles only, no video encoding
porter "<URL>" --burn skip

# Diagnose the environment
porter doctor

# Show resolved configuration (secrets masked)
porter config list

Output layout

<output_dir>/<video_id>_<safe_title>/
├── raw/
│   ├── video.mp4                 H.264 + AAC + faststart master
│   ├── audio.wav                 16 kHz mono reference track
│   ├── audio_enhanced.wav        denoised track used only for ASR
│   ├── transcript.json/.txt      reconstructed sentence script book
│   ├── subtitle.srt              platform-provided source subtitles, if any
│   ├── cover.jpg
│   └── metadata.json
└── cooked/
    ├── subtitle_bilingual.{srt,ass}
    ├── subtitle_zh.{srt,ass}
    ├── video_bilingual.mp4
    └── video_zh.mp4

Configuration

Resolved in this order, highest priority first:

1. --config <path>
2. $PORTER_CONFIG
3. Environment variables (OPENAI_API_KEY, PORTER_ASR_ENGINE, …)
4. Project-level ./porter.json
5. User-level config directory (platformdirs)
6. Built-in defaults

MCP server

{
  "mcpServers": {
    "porter": {
      "command": "uvx",
      "args": ["--from", "porter-workflow[mcp]", "porter-mcp"]
    }
  }
}

Long jobs are exposed as a start/status/result/cancel API rather than one blocking call, because encoding a 1080p video can take tens of minutes.

Development

uv sync --extra all --extra dev      # installs exactly what uv.lock pins

ruff check src          # includes T20: the engine must not print()
mypy src
lint-imports            # enforces the engine/frontend boundary
pytest

CONTRIBUTING.md documents the full gate, the architecture rules those tools enforce, and the licence rules for new dependencies.

Repository layout:

src/porter/        engine (library) — all business logic lives here
src/porter_cli/    CLI frontend
src/porter_mcp/    MCP frontend
skills/porter-skill/   Agent Skill assets (SKILL.md, scripts, references)
tests/{unit,regression,integration}/

License

MIT. See LICENSE.

Acknowledgements

Porter is built on other people's work, and the debt is worth stating plainly:

• yt-dlp — downloading from five platforms is a solved problem rather than five scrapers, because of this.
• VideoCaptioner (GPL-3.0) — its sentence segmentation, alignment and subtitle-processing design shaped the approach in porter.subtitles and the ASR chain. Ideas only: no code was copied, and it is never imported — only run as an external process, which is why porter can stay MIT.
• FFmpeg and libass — the entire media layer: standardisation, denoising and professional ASS rendering.

THIRD_PARTY_NOTICES.md records every component, its licence, and whether porter depends on it, optionally installs it, or only spawns it as a subprocess.

Contributing

Contributions are welcome. CONTRIBUTING.md covers the development setup and the gate every change must pass; SECURITY.md explains how to report a vulnerability privately; release history lives in CHANGELOG.md.

This project follows the Contributor Covenant.