ArcReel is a self-hosted AI video production workbench for animated short dramas, novel adaptations, narrated explainer videos, and product or advertising clips. It organises content analysis, asset management, storyboarding, media generation, cost tracking, and export into one production pipeline that you can audit, interrupt, and resume.
The pipeline runs in stages: you paste a novel, a finished script, or product material; an agent reads the whole text and extracts characters, scenes, and props into three asset libraries; it proposes episode split points for you to approve; it turns the text into a structured JSON script of scenes and shots across three creation types (narration, drama, and advertising); it generates character design sheets first and uses them as the reference for every storyboard frame; and it renders each shot to a video clip before composing the final cut. Every stage can be orchestrated by the agent or reviewed, adjusted, and regenerated by you in the workbench.
Orchestration uses a main agent with focused subagents (analyze, split, script, render). Large context such as the source novel stays inside the subagent, and the main agent receives only a condensed summary. Text, image, video, and text-to-speech generation share one backend protocol with 11 built-in providers and more than 100 preset models, including Gemini, Volcengine Ark, OpenAI, Tongyi, Kling, and MiniMax, plus custom endpoints such as Ollama and vLLM for self-hosted models.
Features
- Input types: novels, finished scripts, or product material, in Chinese or English, of any length
- Asset libraries: characters, scenes, and props extracted automatically and tracked across the whole production
- Cross-shot consistency: character design images generated first and reused as references for every storyboard frame and video clip
- Structured scripts: scene and shot JSON covering narration, drama, and advertising creation types, with human approval at episode split points
- Multi-agent orchestration: a main agent dispatches to subagents so large source text does not overflow context
- Provider matrix: 11 providers and 100+ preset models across text, image, video, and TTS, plus custom OpenAI-compatible endpoints
- Task queue: asynchronous engine with per-provider RPM limits, separate image and video channels, lease-based scheduling, resumable tasks, and retries that only redo failed items
- Version history: every regeneration is archived, with one-click rollback and side-by-side comparison
- Cost tracking: estimated and actual usage visible before and after generation
- Export: automatic composition through FFmpeg, or per-episode Jianying 5.x/6+ draft packages for manual editing (mainland China Jianying; CapCut compatibility not yet verified)
- Deployment: Docker Compose with SQLite by default and a documented migration path to PostgreSQL
