ArcReel: Open-Source Multi-Agent Pipeline That Turns Stories Into Videos

By Prahlad Menon Published 2026-06-26 3 min read

Most AI video tools skip the hard part. They’ll generate a clip from a prompt, but ask them to maintain a consistent character across multiple shots and they fall apart. The protagonist in scene one looks nothing like scene three.

ArcReel takes a different approach: build character consistency into the pipeline architecture from the start.

The Pipeline

The system breaks video creation into distinct stages, each handled by a specialized agent:

Novel/Script → Character/Scene/Prop Design → Screenplay → Storyboards → Video Clips → Final Video

Before any video generation happens, the system extracts characters and “clues” (key props, locations, recurring visual elements) and generates reference designs. Every subsequent storyboard and video shot references these designs, maintaining visual coherence across the entire production.

Multi-Agent Architecture

Built on Claude Agent SDK, ArcReel uses an orchestrating Skill that detects project state and dispatches focused sub-agents:

analyze-characters-clues — Global character and prop extraction
split-narration-segments — Segment splitting for narration mode
normalize-drama-script — Drama/episode structure normalization
create-episode-script — JSON screenplay generation
Asset generation sub-agents — Character portraits, storyboard frames, video clips

Each sub-agent completes one task and returns a summary. The heavy context (full novel text, detailed descriptions) stays inside the sub-agent — only distilled results bubble up to the main agent. This keeps the orchestration context clean while handling long-form content.

Provider Flexibility

Video generation supports multiple backends:

Provider	Models	Capabilities
Gemini	Veo 3.1, Veo 3.1 Fast	Text-to-video, image-to-video, extension
火山方舟	Seedance 2.0, 1.5 Pro	Text/image-to-video, audio, seed control
Grok	Grok Imagine Video	Text/image-to-video (1-15s)
OpenAI	Sora 2, Sora 2 Pro	Text/image-to-video (4/8/12s)
Vidu	Q3 Turbo/Pro, 2.0	Reference-based video, audio
Kling	v2.5, v3, v3 Omni	Text/image/reference-to-video

Image generation has similar multi-provider support (Gemini Nano Banana, GPT Image 2, Grok, etc.). You can switch providers at global or project level.

Three Content Modes

Narration — Splits by reading rhythm, optimized for storytelling videos
Drama — Organizes by scene/dialogue structure for episodic content
Ad/Short — Generates product shots with target duration, single video per episode

The Architecture That Matters

What makes this interesting isn’t the individual features — it’s the sequencing:

Extract characters and visual elements first
Generate reference designs before any scene work
Propagate those references through storyboards and video generation

This solves the consistency problem by making it impossible to skip. You can’t generate a video clip without first having character reference images that anchor the visual identity.

Quick Start

git clone https://github.com/ArcReel/ArcReel.git
cd ArcReel/deploy
cp .env.example .env
docker compose up -d
# Visit http://localhost:1241

Configure at least one provider API key (Gemini, Grok, OpenAI, etc.) in the settings page, and you’re ready to upload a story.

Cost Reality Check

API costs add up. The README includes detailed pricing tables for each provider. A multi-episode production with high-quality video will cost real money. But the architecture is designed to let you preview and iterate at each stage before committing to expensive video generation.

Why This Matters

The “story to video” space is heating up. What distinguishes ArcReel is the explicit multi-agent design with character/prop consistency baked into the workflow — not bolted on as an afterthought.

If you’ve tried generating a coherent short film with existing tools and hit the “everyone looks different in every shot” wall, this is the architectural approach that addresses it.