v0.16.0 · Commercial beta · Now accepting applications

From a novel
to a finished reel,
agent-driven.

ArcReel is an open-source AI short-drama & manga workspace. Drop in a novel, and an agent orchestrates scriptwriting, character design, storyboarding and final video synthesis — keeping characters, scenes and props visually consistent across every shot.

Apply for commercial beta Star on GitHub · 2.5k

AGPL-3.0

License

Docker · WSL2 · macOS

Runs anywhere

5 providers · 40+ models

Gemini · Volcengine · Grok · OpenAI · Vidu

Multi-agent orchestration

Subagent architecture

pipeline.reel · v0.16.0

06 stagesagent-orchestrated

Upload novel

Drop in the source text. Chinese or English, any length.

Build asset library

Agents scan the full work and index every character, scene, and prop.

Plan & split episodes

Progressive human-in-the-loop episode breakdown with AI-suggested cut points.

Generate script JSON

Normalize prose into structured scene/shot JSON — narration or drama mode.

Character & storyboard frames

Reference sheets first; then every shot with cross-scene consistency.

Synthesize video

Image-to-video per shot, then FFmpeg-composed final reel or Jianying export.

Built for consistency across every shot.

Characters keep their faces. Props keep their shapes. Scenes keep their mood. ArcReel's agent graph treats continuity as a first-class artifact — not an afterthought.

Agent workflow

Orchestration Skill + focused Subagents

A state-aware orchestrator detects your project phase and dispatches subagents for each task. Large context (novel text, references) stays inside subagents — only distilled summaries reach the main thread.

Character DNA

Cross-shot consistency

Reference sheets are generated first; every downstream storyboard and video clip is conditioned on them. Characters, scenes, and props are tracked as persistent assets across every cut.

Queue

Async task engine

RPM-rate-limited, lease-based scheduling with independent image / video channels. Resumable.

Versioning

Every regen is history

One-click rollback. Compare variants side by side. Nothing is ever lost.

Export

Jianying-ready drafts

Ship per-episode ZIPs into Jianying 5.x / 6+ for human finishing. FFmpeg pipeline for automated cuts.

Bring your own model stack.

A unified backend protocol across image / video / text. Five preset providers and 40+ models out of the box — or plug in any OpenAI-compatible or Google-compatible endpoint, including self-hosted Ollama and vLLM.

Modality

Gemini

Volcengine

Grok (xAI)

OpenAI

Vidu

Image

Nano Banana 2 / Pro

Seedream 5.0 / Lite / 4.5 / 4.0

Grok Imagine Image / Pro

GPT Image 2 / 1.5 / 1 Mini

Vidu Q2 / Q1 Image

Video

Veo 3.1 · Fast · Lite

Seedance 2.0 · Fast · 1.5 Pro

Grok Imagine Video

Sora 2 · Sora 2 Pro

Vidu Q3 Turbo · Pro · 2.0

Text

Gemini 3.1 Pro / 3 Flash / 3.1 Flash Lite

Doubao Seed 2.0 / 1.8 series

Grok 4.20 / 4.1 Fast

GPT-5.5 / 5.4 / Mini / Nano

—

Preset providers

40+

Models supported

∞

Custom endpoints

Content modes · narration / drama

Apply for the commercial beta.

The open-source build is free forever. The commercial tier adds hosted infrastructure, pooled provider credits, priority queues, SSO, white-label branding, and an SLA. Limited spots in this cohort.

Join the community.

Swap tips, show reels, troubleshoot, and shape the roadmap with other creators and devs building on ArcReel.

Get early access to the reel.

We post release notes, prompt recipes, and sneak peeks of unreleased agents in the group. Join before the next cohort closes.

◆Weekly model & feature digests
◆Direct access to core maintainers
◆Showcase channel for community reels
◆First look at commercial features before GA

Feishu · Chinese-first

Scan with Feishu / Lark

Frequently asked.

The open-source build is AGPL-3.0 and free forever — self-host with Docker Compose on Linux, macOS or WSL2. You only pay for the upstream AI provider usage (Gemini, Volcengine, Grok, OpenAI, or your own). The commercial tier is an optional managed offering for teams that want hosted infra, pooled credits, and SLAs.

Before any storyboard is rendered, a subagent scans the novel and builds a library of characters, scenes, and props. Reference sheets are generated first and every downstream image/video generation is conditioned on them. Continuity is designed in, not a prompt hack.

Five preset providers ship out of the box: Gemini (Nano Banana 2 / Veo 3.1), Volcengine (Seedream / Seedance), Grok, OpenAI (GPT Image / Sora 2), and Vidu (Vidu Q3 video / Vidu Q2 image) — 40+ preset models across image, video and text. You can also add any OpenAI-compatible or Google-compatible endpoint — including self-hosted Ollama and vLLM — and ArcReel will auto-discover models.

Yes. Per-episode drafts export as a ZIP compatible with Jianying desktop 5.x and 6+. Import, fine-tune the cut, add music, done.

A POSIX-like environment: Linux, macOS, or Windows with WSL2 / Docker Desktop. A few low-level dependencies are POSIX-only, so native Windows isn't supported yet.

Same core pipeline, but hosted — no setup, pooled provider credits at better rates, priority GPU queues, team workspaces with SSO, white-label branding for agency use, and a support SLA.

All of those are closed-source SaaS — your scripts, character assets and generated content live in their cloud. ArcReel is AGPL-3.0 open source and self-hostable: data stays on your machine, you hold the model API keys, and full novels never leave your server. ArcReel is also model-neutral: Sora 2, Veo 3.1, Seedance 2.0, Nano Banana 2, GPT Image 2 and Vidu Q3 are all wired up out of the box, plus Ollama / vLLM for local LLMs. Commercial platforms win on zero-ops and ready templates; ArcReel wins on control, hackability, and zero subscription fees.

All three ship preset in ArcReel. Pick by use case: Veo 3.1 produces the most natural long takes and camera motion, with Fast / Lite variants for speed-cost tradeoffs. Sora 2 / Sora 2 Pro give the strongest realism and acting detail. Seedance 2.0 has the richest multi-modal reference (mix up to 9 images, 3 videos, 3 audio clips per shot), strong Chinese-scene understanding, and token-based pricing that stays predictable on long projects. Vidu Q3 is a fourth option — reference-to-video plus audio generation, credit-based pricing, and solid Chinese-scene handling. Switch globally or per-project — no need to commit upfront.

Nano Banana 2 (Gemini 3.1 Flash Image) supports multi-reference input. ArcReel's flow: a Subagent first scans the full novel to extract a cast list, then generates a reference sheet for each character; every downstream storyboard frame conditions on that reference. ArcReel also tracks 'clues' — key props and scene elements — the same way. Consistency is baked into the pipeline, not prompt magic.

No. All image / video / text generation goes to provider APIs (Gemini, Seedance, Sora, Grok, etc.) in the cloud. Your machine only runs orchestration logic, FFmpeg compositing, and Jianying draft packaging. A normal laptop with macOS / Linux / WSL2 + Docker Compose is enough. If you wire in self-hosted Ollama / vLLM for local LLMs, that machine needs a GPU — but with the five preset providers, no GPU required.

It's the Skill + Subagent pattern from the Claude Agent SDK. The main Agent only reads project state and decides the next step; concrete tasks (character extraction, script normalization, storyboard, video) are dispatched to focused Subagents. Each Subagent handles its large context (novel text, reference images) internally and returns only a distilled summary. The main Agent's context never gets blown up, so the pipeline stays stable across long works — and resumes from any stage.

Progressive human-in-the-loop. ArcReel first peeks at the structure, a Subagent proposes cut points (by plot arc, character entry, POV shift), you confirm or adjust in the web UI, and only then does it physically split. Cut everything at once, or step through episode by episode — each episode generates its own script JSON, reference sheets, storyboard, and video. Two content modes: narration (split by reading rhythm) or drama (split by scene / dialogue).

Yes. ArcReel uses unified ImageBackend / VideoBackend / TextBackend protocols — anything OpenAI-compatible or Google-compatible plugs in. Add a custom provider in settings with Base URL + API Key; ArcReel hits /v1/models to discover what's available and infers media type by name. Custom providers get the same treatment as the five presets: project-level switching, cost tracking, version history. Third-party API gateways work too.

AGPL-3.0 allows commercial use but adds a copyleft and network-service clause: if you modify ArcReel and serve the modified version over a network, you must release your modifications under AGPL-3.0 to end users. Internal use only, or building one-off projects for clients (no network service), doesn't trigger the clause. If you'd rather not open source your modifications, ask us about commercial licensing — dedicated instance + a different license.

AI manga drama leans anime / animation style with snappy pacing and strong visual styling; AI short drama leans live-action realism with melodrama / vertical-video pacing; narration video is text-to-speech + images + simple motion, focused on storytelling rather than performance. ArcReel exposes two content modes: narration (split by reading rhythm) and drama (organized by scene / dialogue). Layer on style references and art-direction settings to switch between manga and short-drama looks.

From a novel to a finished reel,agent-driven.

Open-source AI short-drama & manga workspace · novel to reel · multi-agent orchestrated