Optimizing AI Workflows: Is TOON (Token Oriented Object Notation) on the radar?

Developers Developer APIs
1

Hi Webflow Community,

At Fener Interactive Studio, we’ve been integrating more LLM-based automations into our client builds recently. While Webflow’s native Logic and API are fantastic, we constantly run into the “token tax” issue when sending large CMS datasets to models like GPT-4 or Claude.

As we know, standard JSON is great for compatibility but quite “verbose” for LLMs due to repeated keys and syntax overhead.

I’ve been diving deep into TOON (Token Oriented Object Notation) lately. For those who haven’t seen it, it’s a format designed specifically to minimize token usage (claiming ~30-50% reduction compared to JSON) while remaining machine-readable for AI agents.

I wanted to open a discussion on two points:

  1. Middleware Experience: Has anyone here successfully implemented a “JSON-to-TOON” conversion layer (via Make, n8n, or custom code) before sending Webflow CMS data to an AI endpoint? If so, did you notice a significant improvement in latency or cost?

  2. Native Support: I know it’s early days, but is there any discussion within the Devs or Community about supporting lighter, token-friendly serialization formats within Webflow Logic in the future?

We are currently testing this via custom middleware, but I’d love to hear if anyone else is exploring this path to optimize their AI-powered Webflow apps.

Best,

Berkay Çınar
Fener Interactive Studio

2

At this point, Webflow is just at the very beginning of exploring native support for CMS data pushed into apps, which you can see it in the new Appgen. What you’re talking about is several steps away which puts it squarely in custom build territory for now.

Anything you want to do with JSON compression is easy enough in your middle tier, but for most apps I find agent-facing query support much more important to token efficiency. For example, Webflow doesn’t have a way to traverse master-detail relationships efficiently. You cannot list all B records that are ref’d to an A record, without retrieving every single B record.

In most app-like scenarios, you’re probably burning a lot more tokens there, along with other challenges like rate limits. When I’m building systems like this, I sync the Webflow CMS to an external db and then build querying interfaces that the LLM can use to isolate what it needs, with no rate limits.