axjs/multi-agent-chat-lab

A browser-based interface for experimenting with multi-agent conversations

MultiAgent ChatLab

A browser-based interface for experimenting with multi-agent conversations using various LLM APIs. Have conversations between different AI models or observe how they interact with each other - all directly in your browser. For more information check out this piece

Features

• Pure client-side implementation - no server required
• Support for multiple LLM providers:
  • Google (Gemini models)
  • Anthropic (Claude models)
  • OpenAI models
• Real-time conversation history including Markdown rendering support (via Showdown) and color-coded messages for easy agent tracking
• Editable message history to facilitate experimentation
• Simple tabbed interface supports up to 6 agents (each can have own system prompts and models)

Quick Start

1. Clone this repository or just download multiagent-chatlab.html
2. Open multiagent-chatlab.html in your browser
3. Configure the agents (with your own LLM API keys, names, and system prompts)
4. Start chatting!

Security Note

This tool runs entirely in your browser. Your API keys are never sent anywhere except directly to the respective LLM providers' APIs. However, always use API keys responsibly.

How It Works

The ChatLab allows you to:

1. Customize Agents (you can pick model, name, and system prompt)
2. Send messages as a user or pick Agent to respond next in a conversation
3. Edit or delete any message in the conversation history

Each message in the conversation is color-coded by sender for easy tracking of who said what (for up to 6 total participants). The interface automatically handles the appropriate API formatting for each provider's chat history requirements and does simple prefix adding / removing behind-the-scenes.

Requirements

• A modern web browser
• API key(s) for the LLM provider(s) you want to use:
  • Google AI API key for Gemini models
  • Anthropic API key for Claude models
  • OpenAI API key for OpenAI models

Development

This project is built using:

• Preact for UI components
• htm for JSX-like syntax without build steps
• Showdown for in-browser Markdown rendering
• Pure browser JavaScript - no bundler or build process required

Contributing

Some areas that could use improvement:

• Additional LLM provider support
• Improved error handling
• More customization options for system prompts
• Conversation templates/presets

License

MIT License

Credits

Created by Benjamin Tseng