Highly custom design of frostbyte excels in paper workflow with AI agents! My workflow to go from a design canvas to live demo: Paper MCP + Antigravity for design exploration and Paper + Claude Code + Github for live prototype! Live demo of frostbyte here Social post here I have explored various scenarios where the AI agent can help me go from concept, into design exploration and develop it into a live prototype by interacting with Paper MCP. Frosbyte is an file compression tool that automatically detects the format of the files and makes it easy to compress and convert. (I only had time to finish the image compression tool) I connected AI agents to perfect the initial concept and create new variations of the compression tool for images, by enhancing the UI and UX of it. Extra accuracy for pixel perfect design, as I prompted the AI to read the annotations from paper and further develop it with the correct animations into the engaging hero section that it should become. Paper makes it look easy. AI models can read easily and turn ideas into design ideation, read highly custom designs and turn them into a live prototype. Additionally, AI agents were prompted to work on the responsive screens of the design on paper. Paper + AI agents finished my concept idea 100x faster and made it into a live demo!

Try the demo here https://tryfrostbyte.com/

Building a Zero-Touch Design Feedback Loop with Slack and the @Paper MCP

What if you never had to manually process a minor design revision again? As designers and developers, we easily waste 30% of our week fulfilling fast client requests—changing button colors, adjusting typography, or tweaking padding. I decided to completely automate this process.

In this project, I engineered an autonomous AI-driven workflow that physically alters native UI designs based on conversational Slack messages, proving the massive potential of the Paper MCP.

✨ The Power of the Paper MCP: Historically, AI agents have struggled to actually "see" or edit design files. By deploying the Paper MCP, I gave a raw Large Language Model the ability to programmatically traverse a 2D design canvas.

The Setup: I design the initial UI (like a Login Page) using Paper and post the screenshot to the client's Slack channel.

The Trigger: The client requests changes natively in Slack by mentioning a custom @DesignBot (e.g., "Make the login button green and change the title to 'Welcome'").

The Execution: Instead of guessing layout geometry, the AI uses the Paper MCP to dynamically query the canvas, isolate the precise target node, and duplicate the master to create a protected variant. It aligns the new design on the canvas grid and updates the styling properties natively.

The Delivery: The MCP server intercepts the final render state and pushes a flawlessly formatted screenshot right back into the Slack thread.

🛠️ Technical Stack Used:

Paper MCP: The absolute core of the project, enabling rigorous JSON-RPC access to native design layers, text bounds, and styling hooks.

Slack Bolt (Socket Mode): For threading logic, instant reaction processing, and conversational memory.

Agentic LLM Engine: Used native OpenAI tool-calling to map the complex MCP schemas directly into the AI’s brain so it never hallucinates commands.

The Result: Clients experience magical, instantaneous visual revisions directly inside their communication channels, while designers and developers win back hours of completely uninterrupted, deep-focus work.

Check out the video demo below to see the AI leverage Paper MCP to parse a layer tree and dynamically manipulate the UI!

This is the kind of workflow that quietly kills 30% of design busywork 👀