GitHub - richstokes/InfiniChat: 💬 A CLI app that lets you watch conversations between two LLMs (all running locally using Ollama)

3 min read Original article ↗

InfiniChat

InfiniChat is a command-line application that simulates conversations between two LLMs running locally using Ollama. Just for fun really - I find it interesting to watch the chats! You can have them flesh out ideas, debate things, argue, or just give vague prompts and see what topics they spiral in to.

InfiniChat Screenshot

Features

  • Full Conversation History: Models maintain a record of the entire conversation, enabling more coherent interactions
  • Streaming Responses: Real-time streaming of model outputs with live display
  • Attractive Terminal UI: Rich text formatting with color-coded speakers and panels
  • Debate Mode: Set a specific topic for the models to debate, with one arguing "for" and the other "against"
  • Conversation Saving: Automatically saves transcripts of conversations

Requirements

  • Python 3.+
  • Ollama installed and running
  • Required models (llama3:latest and gemma3:12b by default) pulled in Ollama
  • A "non-trivial" amount of RAM. This uses 30GB+ on my Macbook.

Installation

# Install dependencies
pipenv install

Usage

Basic usage:

Edit the prompts in prompts.py to your liking.

Command-line Arguments

InfiniChat supports the following command-line arguments:

Argument Description Default
--max_turns Maximum number of conversation turns between the two AI models 9000
--max_tokens Maximum number of tokens per response from each AI model (just leave it at default unless you know you need to change it) 1000
--debug Enable debug mode for additional output False
--show_json Show RAW JSON from chat api False
--stats Show message history statistics in panel titles False
--history_limit Number of messages to keep in conversation history for each model before summarizing and trimming. Turn this down if messages gradually start to take longer to generate 100
--delay Delay in seconds between streaming chunks (for slower, more readable streaming) 0.0
--model_a Name of the first AI model to use llama3:latest
--model_b Name of the second AI model to use gemma3:12b
--debate_topic "Pizza is a vegetable" Topic to debate, model A will be "for" the topic, model B will be "against" None
--model_a_prompt "Your custom prompt" Custom system prompt for model A (overrides default from prompts.py) None
--model_b_prompt "Your custom prompt" Custom system prompt for model B (overrides default from prompts.py) None

Examples

Run with custom settings:

# Run with 2000 conversation turns!
pipenv run python app.py --max_turns 2000

# Run in debug mode
pipenv run python app.py --debug

# Show message history statistics in panel titles
pipenv run python app.py --stats

# Add a delay for slower, more readable streaming
pipenv run python app.py --delay 0.1

# Use different models
pipenv run python app.py --model_a qwen:latest --model_b deepseek-r1:latest

# Start a debate
pipenv run python app.py --debate_topic "Coffee is better than tea"

# Use custom prompts for both models
pipenv run python app.py --model_a_prompt "You are a cheerful assistant who loves to help people" --model_b_prompt "You are a serious academic who prefers formal language"

# Override just one model's prompt
pipenv run python app.py --model_a_prompt "You are a pirate who speaks in nautical terms"

After running, a conversation transcript will be saved to conversation_history.txt. This will be overwritten each run, so copy it somewhere if you wish to keep it.