AutoVideo-Agent

Turn a Markdown script into a reproducible video pipeline — storyboard, scene assets, timeline, QA, and MP4.

Built for Codex, Claude Code, Gemini CLI and other coding-agent workflows. v0.1 is local-first and deterministic: it creates inspectable placeholder scene assets and an FFmpeg video without an API key or cloud account.

Markdown Script -> Storyboard -> Scene Manifest -> Media -> Timeline -> FFmpeg -> QA -> MP4

The default v0.1-compatible command does not claim AI video generation. v0.2 provider mode adds ComfyUI API media, while MiniMax and other hosted providers remain planned.

Demo

The demo uses examples/demo-script.md: three scenes and seven seconds.

INPUT                         PIPELINE                         OUTPUT
examples/demo-script.md  ->  autovideo run              ->  video.mp4
                              parse + manifest + assets      manifest.json
                              silent WAV + FFmpeg            report.json

Run it:

autovideo run examples/demo-script.md

The build is written to build/demo-script/. Open video.mp4 when FFmpeg is available. Always inspect manifest.json and report.json; without FFmpeg the command reports status: degraded and keeps the inspectable assets.

Quick Start

Install the package from PyPI:

python -m pip install autovideo-agent

To run the repository demo, clone the repository for its example script:

git clone https://github.com/wangxin6x/AutoVideo-Agent.git
cd AutoVideo-Agent
autovideo run examples/demo-script.md

The wheel contains the autovideo CLI and runtime package. examples/ is a repository fixture, so use your own Markdown script after installing from PyPI or clone the repository to run this demo.

FFmpeg is optional. With it, the output is an H.264 MP4 with a silent AAC track. Without it, scene cards, manifest, WAV timeline, and QA report are still produced.

Features

Architecture

flowchart LR
    Script[Markdown Script] --> Parser[Script Parser]
    Parser --> Storyboard[Storyboard]
    Storyboard --> Manifest[Scene Manifest]
    Storyboard --> Providers[Provider Interface]
    Providers --> Media[Media assets]
    Media --> Timeline[Timeline]
    Timeline --> Renderer[Renderer]
    Renderer --> QA[QA report]
    QA --> MP4[MP4 output]
    VideoProvider[ComfyUI Media Provider - Experimental] -. media .-> Providers
    TTSProvider[Mock / Command TTS] -. audio .-> Providers
    AssetProvider[Asset Provider - Planned] -. slot .-> Providers

The current renderer creates deterministic placeholder cards and a silent audio track. Provider slots are documented extension points, not shipped integrations.

ComfyUI validation status

The ComfyUI API behavior is covered by mocked integration tests, but v0.2.0-beta.1 has not yet been validated against a live ComfyUI workflow. The provider is implemented and experimental; live image/video validation is tracked in Issue #12. Do not treat it as production-ready.

Use with Codex

Read AGENTS.md for repository rules, tests, security constraints, and the development loop. Then point Codex at skills/auto-video/SKILL.md for the local storyboard workflow:

Turn examples/demo-script.md into a video and run QA. Use skills/auto-video/SKILL.md.

The real command is:

autovideo run examples/demo-script.md

QA means checking the command result plus report.json and manifest.json; there is no separate AI quality grader. This is a repository workflow, not an endorsement by Codex or any model vendor.

Roadmap

• v0.1 ✅ — Local parser, deterministic cards, silent timeline, FFmpeg MP4, degradation report, tests, and agent onboarding.
• v0.2 (this development branch) — Provider contracts, ComfyUI media, Mock/Command TTS, scene-level SRT, normalized timeline, mixed renderer, and deterministic QA. MiniMax and hosted TTS remain planned.
• v0.3 — media adapters #5, cross-platform FFmpeg #6, CI render coverage #9, more formats #10.

Community

• Issues
• Good First Issues
• Feature Requests
• Bug Reports

Contributions to docs, examples, portability, and provider boundaries are welcome. Read AGENTS.md, add tests for core behavior, run python -m pytest, and review git diff --check before opening a pull request.

Development

python -m pip install -e ".[test]"
python -m pytest

The runtime has no third-party dependencies. Never commit API keys, tokens, passwords, cookies, or machine-specific paths.

中文文档

中文文档 -> README_CN.md

License

MIT. See LICENSE.