AKITA - the Goofiest Website Builder

Akita is a generative website builder that replaces the traditional chat window with a co-creative interface. Instead of prompting in text, users manipulate visual components in a drag-and-drop canvas while Claude reasons in real time.

This prototype achieves stability through a clear architectural separation of concerns:

• A React-based frontend manages deterministic UI state and drag-and-drop interactions.
• A Node.js + Express backend orchestrates Claude via a controlled CLI environment.
• WebSockets provide low-latency, real-time synchronization between AI reasoning and UI rendering.
• By using structured tool calls (e.g., layout retrieval and preview rendering) rather than raw text streaming alone, the system constrains generative output into predictable UI updates. This prevents hallucinated structure from breaking the interface and ensures consistent rendering in the live preview panel.

Our interface itself is the innovation because:

• Traditional AI products center around a chat box. Akita eliminates that constraint.
• Users build by touching the interface, not by describing it.
• Claude reasons over layout state rather than plain conversation history.
• The preview is not an output artifact — it is the active workspace.
• The system turns AI from a “respondent” into a design engine.
• This playful inversion of the chat paradigm transforms AI from a reactive assistant into a collaborative UI runtime. The result feels less like prompting a model and more like sculpting with intelligence.

Technologies & Frameworks we used include:

• Frontend:
  React – Component-based architecture for dynamic canvas rendering
  Vite – High-performance development and bundling
  HTML5 + CSS3 – Responsive layout rendering
• Backend:
  Node.js + Express – API orchestration and request handling
  WebSockets – Real-time bidirectional communication
  node-pty – Sandboxed Claude CLI execution
• AI Integration:
  Claude (via CLI + structured tool calls) – Layout reasoning and generative HTML synthesis

No prior work at all.