Desktop app turns novels into short dramas via agent triage

What it does

Toonflow is an Electron-based workbench that takes a story or script and pushes it through a structured pipeline: chapter extraction, script adaptation, storyboarding, character generation, and final video assembly. It does not generate media itself; it orchestrates external LLM, image, and video APIs—such as Claude, GPT Image, and Seedance—while managing context and continuity for you. The interface is an infinite canvas where nodes for scripts, characters, and shots can be arranged non-linearly.

The interesting bit

Instead of a simple prompt chain, the workflow is handled by a three-layer agent system—decision, execution, and supervision—that retains memory across sessions through a local ONNX vector store. The app also externalizes agent “skills” as editable Markdown files, so you can tune prompts without touching source code, and it lets you inject custom TypeScript adapter logic for new model endpoints at runtime.

Key highlights

• Infinite canvas UI for non-linear arrangement of scripts, characters, storyboards, and video nodes.
• Three-layer agent collaboration (decision, execution, supervision) with persistent cross-session memory via local ONNX vector retrieval.
• Chapter event graph automatically extracts and structures plot points from source novels to preserve long-context coherence during adaptation.
• Supplier system accepts custom TypeScript logic for new AI model endpoints without recompiling or restarting the app.
• Core agent prompts are exposed as Markdown “Skill” files for live editing.

Caveats

• The app is essentially an orchestration layer: you must bring your own API keys for LLMs, image generation (Nano Banana Pro), and video generation (Sora or Doubao). The published demo cost about ¥130 in API fees.
• Docker online deployment is marked “待完善” (to be perfected); only local builds are currently supported.
• macOS users may hit certificate warnings on first launch.

Verdict

Worth a look if you are a writer or indie creator experimenting with AI-driven short-form video and want a structured canvas instead of a folder full of ChatGPT transcripts. Skip it if you are hoping for a fully offline, self-contained video generator—this is a conductor, not an orchestra.