BrightBean: TikTok Intelligence for AI agents

I'm Jan, the maker of BrightBean, and I built this out of pure frustration.

I was building AI agents that generate TikTok content, and every single time, the output felt… off.

The hooks were generic. The pacing was wrong. The scripts read like blog posts, not videos.

LLMs are incredible writers, but they have zero intuition for what actually works on TikTok.

So I started studying what the best creators do differently. I indexed tens of thousands of high-performing organic TikTok videos, reverse-engineering the hook types, video structures, pacing rules, and proof patterns that drive real retention. Then I turned all of that into a structured API that AI agents can actually consume.

BrightBean gives your AI agent three things:

/hook - Viral hook patterns grounded by your niche and goal. Not catchy phrases, actual opening archetypes with first-2-second requirements and proof asset suggestions.

/recipe - Shot-by-shot video format. Timing, beats, visual cues, CTA placement. Think of it as a director's brief your agent can follow.

/score - Preflight draft scoring. Before you waste compute rendering a bad video, score it against viral benchmarks and get concrete fixes.

It works via API and MCP, so it plugs into whatever agent framework you're using, OpenClaw, OpenAI Agents, LangGraph, OpenHands, or anything that can call an HTTP endpoint.

I'd love to hear from anyone building agentic content workflows.

What's the biggest bottleneck you're running into?

What would make this more useful for your stack?

Thanks for checking it out, excited to hear your feedback!