What it does Toonflow is an Electron/Vue3 desktop app that tries to automate the full pipeline from text to animated video: ingest a novel or script, extract chapter events into a structured graph, generate a screenplay, storyboard it on an infinite canvas, and export stitched video. It bundles three “Agent” layers—decision, execution, and supervision—and keeps local persistent memory via ONNX vector search so multi-session projects don’t lose context.

The interesting bit The “programmable supplier system” lets you write TypeScript adapter logic in the settings UI to plug in your own LLM, image, and video endpoints without touching source code or restarting. That’s a pragmatic nod to the reality that most users will hit rate limits, pricing tiers, or regional blocks with any single provider.

Key highlights

• Infinite canvas UI for non-linear arrangement of scenes, characters, and assets
• Chapter event graph extraction to reduce long-context loss when adapting novels
• Skill files: core prompts for the script and production agents are external Markdown you can edit live
• Docker and native builds for Windows, Linux, macOS (with the usual unsigned-binary warnings on Mac)
• Demo claims ~2 hours and ¥130 ($18) to produce a 2-minute short using Claude Opus, GPT Image 2, and Seedance 2.0

Caveats

• You must bring your own API keys for language, image, and video models; the README lists specific services (Sora, Doubao, Nano Banana Pro) with no clarity on OpenAI-compatible fallbacks
• Docker “online deployment” is marked “待完善” (pending); only local build works now
• Default login is admin/admin123, which suggests this is not hardened for multi-user or exposed deployments

Verdict Worth a look if you’re experimenting with AI-driven content farms or need a structured GUI over scattered generative tools. Skip it if you were hoping for a batteries-included, offline-capable creative suite—the model bills are still yours to pay.