Designing Zeno through an MCP connector
A remote MCP server (OAuth 2.1 + PKCE) that lets Claude design against Zeno’s own component and principle catalogs, page state, and sample records, with handoff in a chat thread.
Here’s a behind-the-scenes one. A lot of the friction in building software isn’t in writing the code. It’s in keeping the design honest. A page mocked up in a vacuum drifts from the components, tokens, and conventions the rest of the app already uses, and someone reconciles it later. We built a connector that lets an AI design against Zeno’s own design system.
A remote MCP server for design sessions
We stood up a remote MCP (Model Context Protocol) server that exposes Zeno’s design system as a set of tools. Connect Claude to it and the model can browse the module catalog, read the component and principle libraries, inspect a page’s current state, and pull sample records to see how the data behaves in a layout: long names, big numbers, missing optional fields. It designs with the building blocks our developers ship with.
OAuth 2.1, and module-scoped slices
The connector authenticates over OAuth 2.1 with PKCE, so it slots into the native claude.ai connector flow without us handing out long-lived secrets. The design system is served in module-scoped slices rather than one enormous catalog dump, so a session pulls only the part of Zeno it’s working on.
Handoff happens in a chat thread
When a design is ready, the connector opens a handoff as a chat thread with the mockup, the context, and the reasoning attached. A developer reads it, replies, and marks it landed or rejected in the same place the rest of the team talks. Nothing gets emailed into a void.
What grounding the model changes
We use AI heavily in how we build Zeno, and one lesson keeps repeating. A model with access to the system produces mockups that translate to code on the first try. A model working from a description of the system produces mockups someone has to reconcile. That’s the same bet behind the AI features inside Zeno: narrow scope, grounded in your own records, and a person approves the output.
This one’s mostly internal plumbing, but we think the approach is worth sharing. Curious how we use AI to build? [email protected].