MiniVAgent is a node-based workflow application powered by Nano Banana Pro for building AI-driven image and text pipelines. Create flexible generation flows by connecting nodes on a visual canvas, leveraging Google Gemini models and other modern AI tools.
- Visual Node Editor: Drag-and-drop interface for building AI workflows
- Multiple Node Types:
- Text Generation (Gemini Flash 2.5, Gemini Flash 3, GPT-5.4)
- Image Generation (Nano Banana 1, Nano Banana Pro, Nano Banana 2, Imagen 4, GPT Image 1 / 1.5)
- Vision/Image-to-Text Analysis
- Image Source (URL or Upload)
- Note/Documentation Nodes
- Prompt Enhancement: Optional AI-powered prompt optimization
- Flow Execution: Run individual nodes or entire workflows
- AI Flow Assistant: Natural language graph builder
- Sample Workflows: Pre-built examples to get started
- CLI Support: Execute graphs from command line for automation
- Node.js (v18 or higher)
- A Gemini API key (required) or OpenAI API key (optional)
-
Clone the repository and install dependencies:
git clone https://github.com/bmustata/minivagent.git cd minivagent npm install npm run dev -
Create a
.env.localfile in the root directory:GEMINI_API_KEY=your_gemini_api_key_here # Required for base functionality OPENAI_API_KEY=your_openai_api_key_here # OptionalGet your Gemini API key from https://aistudio.google.com/api-keys.
Get your OpenAI API key from https://platform.openai.com/api-keys.
⚠️ Security Warning: Never commit.env.localor any file containing your API key to version control. It is already included in.gitignore. Keep your key private — anyone with access to it can make requests billed to your account.
Development Mode (Hot Reload):
npm run devThen open http://localhost:3202 in your browser.
- UI (Vite Dev Server): http://localhost:3202
- API Server: http://localhost:3201/api/
Compiled Mode (Production Build + Watch):
npm run dev:compiledThen open http://localhost:3201 in your browser.
- UI (Static Build): http://localhost:3201
- API Server: http://localhost:3201/api/
Note: In development mode, Vite runs the client on port 3202 with hot-module replacement. In compiled mode, the client is built to
/distand served by Express on port 3201.
- Build Workflows: Drag nodes from the toolbar onto the canvas
- Connect Nodes: Click and drag from output handles (right) to input handles (left)
- Configure: Click nodes to edit prompts and settings
- Execute: Click the play button on individual nodes or run the full flow
- View Results: Generated images appear in a gallery modal
- AI Assistant: Use the Flow Assistant to describe a workflow in natural language and auto-build the graph
- API Endpoints Reference
- Node Types Reference
- Supported Models
- Graph ID Conventions
- Project Structure
| Type | Purpose |
|---|---|
| Text Generator | Generate or transform text using AI; supports prompt enhancement (Gemini, OpenAI) |
| Image Generator | Generate images from text prompts and optional reference images; supports aspect ratio, output format (JPEG/PNG) (Gemini, OpenAI) |
| Vision / Image-to-Text | Analyze or describe images using AI (Gemini, OpenAI) |
| Image Source | Load or upload a reference image (URL or file upload) |
| Note | Add documentation, prompts, or plain text that can feed into other nodes |
| Compare | Side-by-side comparison of two images with passthrough outputs |
| Split Text | Split text into parts by a separator — ideal for creating multiple variations of scenes, products, descriptions, and more |
This project is intentionally minimal and designed for local use only. Please review the following constraints before deploying:
- Local-Only Usage: Built to run on a local machine or within a private network. Not intended to be exposed to the public internet.
- No Authentication or Authorization: The application does not implement any authentication or access control mechanisms.
- Single-User Design: No concept of multiple users, roles, or tenants. The system assumes a single-user environment.
MiniVAgent is released under the Apache 2.0 License and encourages all types of contributions. No contribution is too small, and we want to thank all our community contributors.

