ArcReel: The Multi-Agent Pipeline That Turns a Story Into a Video

By Prahlad Menon 4 min read

The viral post that introduced most people to ArcReel called it MIT-licensed. It’s not — it’s AGPL-3.0. If you’re planning to build something on top of it, that distinction matters. More on that below.

The license correction aside, ArcReel is worth understanding on its own terms. It’s an open-source AI video workspace that takes a novel or story as input and produces a finished short video as output — screenplay, character designs, storyboards, video clips, final assembly. The whole pipeline. It hit 1,321 stars and 271 forks since its February 2026 debut, and as of today it’s still actively updated.

The Problem It’s Actually Solving

Character consistency is the persistent, underaddressed failure mode of AI video. Current tools either ignore it entirely or attempt to solve it at generation time through prompt engineering — referencing character descriptions in every frame request, hoping the model stays coherent. It doesn’t, reliably.

ArcReel takes a different approach: generate locked character reference images first, before any scene work begins. Every subsequent image and video generation references those locked designs. Same treatment for key props and locations, which the project calls “clues.” The pipeline won’t move to storyboards until the character and clue assets exist.

This is the architecturally correct answer. It’s not a new idea — film production has always locked character designs before production begins — but it’s the first open-source AI pipeline that enforces this as a structural constraint rather than a suggestion.

How the Pipeline Works

The flow is sequential and state-aware:

  1. Upload your novel or story text
  2. Script generation — two modes: narration (audiobook-style voiceover) or drama (scene-based dialogue and action)
  3. Character design — AI generates reference images for each character; these are locked
  4. Clue design — key props and locations get their own reference images
  5. Storyboard — per-scene images generated referencing the locked character and clue designs
  6. Video clips — one clip per scene
  7. Assembly — FFmpeg stitches the final video, or you export a CapCut/Jianying draft

The agent knows where a project is in this sequence and can resume from any point. If clip generation fails halfway through, you don’t restart from scratch.

The Multi-Agent Architecture

ArcReel is built on Anthropic’s Claude Agent SDK. The orchestration layer uses a Skill called manga-workflow that detects project state and dispatches focused Subagents — each one handles a single task and returns a summary. The named Subagents include analyze-characters-clues, split-narration-segments, normalize-drama-script, create-episode-script, and various asset generation agents.

This is good agent design. Context isolation per task keeps the main orchestration agent from accumulating noise across a long pipeline. Each Subagent does one thing cleanly. The pattern maps well to the actual structure of video production.

For text generation it leans on Claude for orchestration, with Gemini 3.1, Grok 4, and GPT-5.4 available. Image generation options include Seedream 5.0, Grok Imagine, and GPT Image 1.5. Video generation supports Veo 3.1, Seedance 2.0, Grok Video, and Sora 2. Custom OpenAI-compatible providers are also configurable.

Cost and License Reality

The built-in cost estimator is a thoughtful addition — it shows projected spend per scene, episode, and project before you commit to generation. It handles multi-currency (USD and CNY, since several ByteDance-affiliated providers bill in yuan). Video generation with Veo 3.1 or Sora 2 on per-second billing adds up quickly for anything longer than a short clip. Budget accordingly.

On the license: AGPL-3.0 has copyleft implications. If you build a service or product on top of ArcReel and distribute it, you may be required to open-source your modifications and the surrounding code. The MIT framing in circulation is incorrect — read the actual license before making product decisions.

ArcReel is a Chinese-origin project (the community is on Feishu) with a bilingual CN/EN README. It’s two months old. Deploy via docker compose up -d on localhost:1241.

Where It Sits

ArcReel doesn’t solve AI video generation — it uses the same underlying models everyone else uses. What it contributes is a pipeline architecture that enforces the right sequencing: lock character designs first, then generate everything else against them. For narrative video specifically, that’s the right abstraction.

Early, real costs, AGPL constraints. Worth watching if you’re working in this space.