Gardn | Devpost

About the project

Gardn came from a pretty simple frustration: too many AI-generated websites all feel like they were made in the same factory. Clean, polished, technically fine, and somehow still completely dead inside.

When people design manually, they usually pull inspiration from real websites they love. They save references, compare patterns, notice details, and build taste over time. But when people use AI tools, that process often gets skipped. You give a vague prompt, the model guesses, and suddenly you have another generic landing page with the same soft gradients, same pill buttons, same fake personality.

I wanted to build a better workflow for taste-driven design.

Gardn lets users create collections of websites they love called gardens. Instead of bookmarking random links or trying to describe a visual style in words, users can curate examples that actually reflect the design language they want. From there, the goal is to let users connect AI agents to those gardens through MCP (Model Context Protocol), so the agent can use a curated set of references to generate website designs that are much more aligned with the user’s real preferences.

In other words, instead of prompting an AI with “make it modern and minimal,” you can give it a garden containing Website A, Website B, and Website C, and let it learn from actual taste instead of vague adjectives. The idea is to make AI-assisted design feel more intentional and less like rolling the dice with autocomplete.

How we built it

We built Gardn around the idea of structured inspiration. The core experience is centered on creating and managing gardens, where each garden acts as a curated collection of design references. Users can save websites they like, organize them into collections, and use those collections as a source of truth for future design generation.

The bigger technical idea behind the project is connecting those gardens to AI agents through MCP, so that an agent can access a user’s saved references and generate designs based on them. That turns a garden from just a bookmark collection into a usable design context layer for AI tools.

A big part of the build was thinking through how inspiration should be stored and passed to an AI system in a way that preserves intent. The challenge was not just saving links, but designing a product concept where curated references become meaningful input for generation.

About the project

Gardn came from a pretty simple frustration: too many AI-generated websites feel like they were made in the same factory. Clean, polished, technically fine, and somehow still completely dead inside.

I wanted to build a better workflow for taste-driven design.

Gardn lets users create collections of websites they love called gardens. Instead of bookmarking random links or trying to describe a visual style in words, users can curate examples that actually reflect the design language they want. From there, the goal is to let users connect AI agents to those gardens through MCP (Model Context Protocol), so the agent can use a curated set of references to generate website designs that are much more aligned with the user’s real preferences.

How we built it

The bigger technical idea behind the project is connecting those gardens to AI agents through MCP, so an agent can access a user’s saved references and generate designs based on them. That turns a garden from just a bookmark collection into a usable design context layer for AI tools.

A big part of the build was figuring out how inspiration should be stored and passed to an AI system in a way that preserves intent. The challenge was not just saving links, but designing a workflow where curated references become meaningful input for generation.

Challenges we ran into

One of the biggest challenges was figuring out how to turn a website into usable data for an LLM. After that, the question became: how would this scale to 3 sites? 10? 50? 100?

We solved this using Featherless.ai and Browserbase. With Featherless as our model provider, we were able to give AI models access to a real web browser so they could visit sites the user likes directly. This let the agent capture screenshots and inspect the site’s HTML, CSS, and JavaScript to construct what we call a Soul Design Document, or DESIGN.md.

Each processed site generates its own Soul Design Document and stores screenshots for future reference. This gives our MCP server a structured way to provide an AI agent with the context and tools it needs to build websites that actually reflect the user’s taste.

What we learned

We learned that curation is a missing part of a lot of AI workflows. People do have taste, references, and preferences, but most tools do a terrible job of letting them express that in a structured way.

We also learned that better AI output is not always about better prompting. Sometimes it is about better context. If you give an agent a strong set of references, you can shift the interaction from “guess what I mean” to “work from what I already know I like.”

Building AI products responsibly sometimes means slowing the model down with better human input instead of asking it to do everything from scratch.

Why this matters

Gardn is an attempt to reduce AI “slop” by making inspiration explicit. Instead of replacing taste, it helps users capture and apply it. We think that if AI is going to be part of the design process, it should work from real creative direction, not generic trends and prompt lottery nonsense.

The goal is simple: help people grow a garden of references, then let AI build from that garden in a way that feels intentional, personal, and actually good.