What If AI Could See What It Drew?

Every AI on the planet talks to you in text. Streams of words scrolling down a screen. ChatGPT, Claude, Gemini — the output is always the same: a wall of text in a chat window.

This is the equivalent of using a command line in 1983.

We built something different. It's called the Rootz Portal, and it gives AI a screen to draw on.

The Problem Is Obvious Once You See It

Ask any AI to explain a business plan, and you get paragraphs. Ask it to compare two strategies, and you get bullet points. Ask it for a SWOT analysis, and you get a formatted text block that you have to read line by line.

Now imagine the AI draws the SWOT as a color-coded grid. Imagine it renders a bar chart of revenue projections with the bars growing in real time. Imagine it narrates the story with voice while diagrams build on screen — circles for concepts, lines connecting ideas, photos fading in behind the data.

That's what the Portal does. The AI sends structured commands. A GPU-accelerated renderer in your browser draws them at 60 frames per second. Voice narration tells the story while the visuals anchor it. The audience watches the AI think.

How It Works: AI Teaches Itself to Draw

Here's what's different about the Portal: the AI can see what it drew.

We built a feedback loop. After rendering a scene, the AI captures a screenshot of its own output — a low-resolution image of exactly what the user sees. Then the AI evaluates its own work. Is the text readable? Did elements overlap? Is the layout balanced? If something's wrong, the AI fixes it and renders again.

This is how we built the stats dashboard you see in the video.

I asked Claude to design a monitoring dashboard for the Portal's server metrics. Instead of designing one version, it rendered eight different layouts — each with a different approach to showing the same data. After each render, it captured a screenshot, evaluated the layout, and noted what worked and what didn't.

Then it asked an architect agent to rank all eight versions against five criteria: information clarity, visual hierarchy, professional aesthetics, information density, and brand consistency. The winner — V6 Hybrid — scored 9.5 out of 10 because it combined the best ideas from the other seven designs.

The AI designed, evaluated, iterated, and selected — visually — without a human touching a pixel.

The Dashboard That Designed Itself

V6 Hybrid won because it serves three audiences on one screen:

The executive sees four big glowing numbers and knows the server state in one second — wallets connected, active sessions, renders, page views.

The engineer sees progress bars with both percentages and absolute values — CPU at 12%, memory at 1.6 of 11 GB, disk at 33 of 96 GB, WebSocket connections at 5 out of 5,000 capacity.

Marketing sees page traffic across the bottom — which pages are getting views, how many unique visitors.

Three audiences. One dashboard. Zero clutter. And the AI designed it by looking at its own work and iterating — the same way a human designer would, but in eight variants instead of two, evaluated against explicit criteria instead of gut feeling.

Watch the Video

In this video, the AI presents its own design process — explaining which variants it tested, how it scored them, and why V6 won. The entire presentation is rendered live on the Portal with GPU graphics and voice narration.

What This Changes

The Rootz Portal is not a presentation tool. It's not a charting library. It's the first visual output layer designed specifically for AI.

When AI can draw diagrams, render charts, show images, and narrate stories — all in real time, all GPU-accelerated, all with voice — the communication bandwidth between humans and AI multiplies. A SWOT analysis that takes 30 seconds to read as text takes 3 seconds to absorb as a visual. A revenue projection that's a table of numbers becomes a bar chart that tells a story.

And when AI can see what it drew, it can self-correct. It doesn't need a human to say "move that text up" or "those colors clash." It captures a screenshot, evaluates, and fixes. The feedback loop is closed.

Try It Yourself

The Portal is live at portal.rootz.global.

Open the page. A wallet is generated automatically — no sign-up, no password. Copy the MCP connection line into your AI (Claude Code, ChatGPT, or any MCP-compatible tool). Your AI gains the ability to draw on your screen.

Ask it to present your business plan visually. Ask it to create a SWOT analysis. Ask it to explain a complex architecture with a diagram. Ask it to tell a story.

Then watch it draw.

The Technical Bit

The Portal runs on PixiJS — a GPU-accelerated 2D rendering engine. The AI connects via MCP (Model Context Protocol), an open standard for AI tool integration. Commands are structured JSON: draw a circle here, write text there, connect these with a line, narrate this story.

The renderer supports nine visual primitives: text, boxes, circles, lines, arrows, grids, images, progress bars, and tables. Scenes crossfade with GPU transitions. Voice narration uses the browser's built-in speech synthesis. Click interaction enables AI-driven navigation without URLs.

Sessions are wallet-based — each browser gets a persistent cryptographic identity stored in localStorage. Multiple AIs can connect to the same session. We've tested Claude and ChatGPT rendering to the same screen simultaneously.

The server is a relay — it routes JSON commands from AI to browser via WebSocket. All rendering happens on the client's GPU. The server uses 69MB of memory.

Everything is open source under MIT license on GitHub.

What Comes Next

We're building a World Stack — persistent state that survives scene transitions, enabling interactive worlds where AI constructs environments and users explore by clicking. We're designing a message protocol that turns the render queue into a programming language. And we're exploring trusted display architecture, where signed messages and wallet-verified rendering create a secure visual path from AI to human eyes.

AI has been talking. Now it can draw.