ArcReel: Self-Hosted AI Video Workspace on Docker

Most AI video tools hand you a finished file and stop there. ArcReel is built the other way around: it keeps every intermediate piece — characters, locations, props, storyboard, individual shots — as an asset you can inspect, version and regenerate on its own. All of it runs on hardware you control.

How a story becomes a video

The project accepts three kinds of input: a novel, a completed script, or product material. From there the work moves through a fixed chain, driven by AI agents that either run straight through or pause for a human decision.

• analyze the source material and plan the project
• generate reusable character, scene and prop assets
• break the story into episodes and write the lines
• build the storyboard
• render video segments and voiceover
• composite the result or export it as a draft

Keeping a character recognizable

Consistency across shots is where generative video usually falls apart. ArcReel's answer is reference imagery that carries across shots, backed by a version history you can roll back to. Because assets are separate objects, a single bad prop or scene can be regenerated without rerunning the whole pipeline — which matters when every rerun costs money.

Vendors and the cost meter

Text, image, video and text-to-speech work is handed off to outside providers configured in one place. The project promises cost visibility both before and after a generation, aggregated across those providers. The repository landing page does not name the supported vendors; it defers that to the documentation, so the practical question of which models you can plug in cannot be answered from the README alone.

Deployment and licensing

Installation is a Docker Compose stack; the interface then sits on port 1241, with PostgreSQL or SQLite behind it. The license is GNU AGPL v3.0 with additional terms, and the project lists a contact address for commercial use. One caveat is stated outright in the repository: do not put the service on the open internet as-is.

Who it fits

The export path reveals the intended audience. Drafts are produced for JianYing, the editing app widely used in mainland China, so an editor can keep working on the timeline instead of receiving a flat render. For CapCut, the international sibling, the repository states compatibility has not been verified — worth testing before building a workflow on it.

What we could not verify

Only the project's own repository was available as a source for this report. No independent newsroom had reproduced the pipeline by press time, so render times, output quality and the actual provider cost per minute of finished video remain unconfirmed.

A version number was also unavailable: the landing page carries a release badge, but the number itself was not present in the retrieved text. The star and fork figures are a snapshot from the day of retrieval and move continuously.