Prisma is an experimental application designed to explore the reasoning limits of Google Gemini 3.0, functioning as a "visual thinking laboratory." It targets engineers, researchers, and power users seeking to understand and leverage advanced AI reasoning capabilities. The primary benefit is a transparent, visualized workflow for complex problem-solving, allowing users to observe and analyze the AI's internal decision-making process.

How It Works

Prisma employs a "Deep Multi-Agent Reasoning" architecture. A Manager agent dynamically decomposes user queries into sub-tasks, dispatching them to specialized, parallel "Expert" agents. A recursive refinement loop enables the Manager to review expert outputs, identify flaws, and initiate up to three correction rounds. Finally, a Synthesis agent consolidates results into a coherent response. This approach offers enhanced robustness and transparency over monolithic models by visualizing the AI's complex problem-solving process.

Quick Start & Requirements

• Install/Run: Clone the repo, run npm install, create .env.local with your GEMINI_API_KEY, then npm run dev. Access http://localhost:3000.
• Prerequisites: Node.js environment, Google Gemini API Key for local execution.
• Online Demos: Google AI Studio (https://aistudio.google.com/app/) offers a free, no-key-required experience. A web demo is available but requires a user-provided API key.
• Tech Stack: React 19, Vite 6, @google/genai (v1.34+), Tailwind CSS.

Highlighted Details

• Deep Multi-Agent Reasoning: Dynamic planning, parallel expert execution, and recursive self-correction (up to 3 rounds) for complex problem-solving.
• Full-Process Visualization: Real-time node graphs display reasoning states (Analyzing, Working, Reviewing, Synthesizing), with access to expert 'Internal Monologue' and 'Draft Output'.
• Fine-grained Control: Configurable 'Thinking Budget' per stage (planning, execution, synthesis) and model switching between Gemini 3 Flash (speed) and Pro (depth).
• Modern Engineering: Supports custom API gateways, Markdown/LaTeX rendering, and local session storage for privacy.

Maintenance & Community

The project welcomes Pull Requests for improvements and bug fixes. Sponsorship is accepted via Afdian (https://afdian.com/a/gemini-nexus). No explicit community channels like Discord or Slack are listed.

Licensing & Compatibility

The project is released under the MIT License, which generally permits commercial use and integration into closed-source projects.

Limitations & Caveats

This is an experimental application exploring Gemini 3.0's reasoning capabilities. The web demo requires a user-provided API key, unlike the recommended Google AI Studio version. Recursive refinement is limited to a maximum of three iterations.