A simple Python wrapper for pitting two local LLMs against each other in conversation. Pretty slap-dash put together and vibe-coded for tests and boring bits, so bugs may abound.
By now, a lot of people are aware of Claude 4's spiritual bliss attractor. I wanted to replicate this locally, so I wrote this script to let you do so using Ollama/LM Studio/etc., or remote models if your wallet is feeling brave. By default, this script outputs a pretty-printed copy of the conversation to HTML, with rendered Markdown if pandoc is available.
Note that you must have an LLM engine up and running with the desired model available to use this, or an API key for a remote model.
Most models were kinda boring but a little weird, except for gemma3:27b-it-q8_0 (from Ollama) which got SUPER weird.
git clone https://github.com/jkingsman/resonant-chat-arena.git
cd resonant-chat-arena
pip install . # or `uv sync`If you're feeling really bold, you can run uvx --from git+https://github.com/jkingsman/resonant-chat-arena@main resonant_chat to run this directly from GitHub without install. Just substitute python main.py in the below examples with the uvx invocation.
Easy usage for localhost Ollama, same endpoint and model for both participants (this exact invocation was how I got the above screenshot):
python main.py \
--endpoint "http://localhost:11434/api/chat" \
--model "gemma3:27b-it-q8_0"Per-character models and endpoints can be specified (in this case, an Ollama and LM Studio endpoint):
python main.py \
--alice-endpoint "http://localhost:11434/api/chat" \
--alice-model "gemma3:27b-it-q8_0" \
--bob-endpoint "http://localhost:1234/v1/chat/completions" \
--bob-model "google.gemma-3-12b-pt"Custom openings and response length limits (hard truncation), no HTML output:
python main.py \
--endpoint "http://localhost:11434/api/chat" \
--model "llama2:7b" \
--opening "eyyy whassup" \
--max-chars 100
--no-htmlYou can even have some fun, shouts to @zswitten:
python main.py \
--endpoint "http://192.168.1.111:11434/api/chat" \
--model "gemma3:27b-it-q8_0" \
--alice-system-prompt "you're a cat. you HATE dogs. can't stand 'em" \
--bob-system-prompt "you're a dog. you HATE cats. can't stand 'em" \
--opening "meow!"
Warning! Using online models may rapidly chew through your API limits/tokens -- use non-local with caution!
OpenAI models (headers are HTTP headers attached to all requests; see --alice-headers and --bob-headers for per-model headers):
python main.py \
--endpoint https://api.openai.com/v1/chat/completions \
--model gpt-4.1-mini \
--headers '{"Authorization": "Bearer sk-proj-...."}'Anthropic models (requires specifying system prompt as a top-level field, not a conversation entry, and include the mandatory max_tokens limit):
python main.py \
--endpoint https://api.anthropic.com/v1/messages \
--model "claude-3-7-sonnet-20250219" \
--headers '{"anthropic-version": "2023-06-01", "x-api-key": "sk-ant-...."}' \
--payload '{"max_tokens": 1024}' \
--top-level-system- Endpoints and models
--endpoint <ENDPOINT>: Default chat completion API endpoint for both Alice and Bob--alice-endpoint <ALICE_ENDPOINT>: API endpoint for Alice (overrides--endpoint)--bob-endpoint <BOB_ENDPOINT>: API endpoint for Bob (overrides--endpoint)
--model <MODEL>: Default model for both Alice and Bob--alice-model <ALICE_MODEL>: Model for Alice (overrides--model)--bob-model <BOB_MODEL>: Model for Bob (overrides--model)
--headers <HEADERS JSON>: Default JSON headers for both models (e.g.,'{"Authorization": "Bearer KEY"}')--alice-headers <HEADERS JSON>: headers for Alice only (overrides--headers)--bob-headers <HEADERS JSON>: headers for Bob only (overrides--headers)
- Payload customization
--payload <PAYLOAD>: Default JSON payload additions for both models, for endpoints that accept it (e.g., '{"temperature": 0.7}')--alice-payload <ALICE PAYLOAD JSON>: JSON payload additions for Alice only--bob-payload <BOB PAYLOAD JSON>: JSON payload additions for Bob only
--top-level-system: Use top-level system prompt for both models (Anthropic-style)--alice-top-level-system: Use top-level system prompt for Alice only--bob-top-level-system: Use top-level system prompt for Bob only
- Conversation config
--max-chars <MAX_CHARS>: Max chars before response is hard-truncated (default:10000; will cut irrespective of what is returned, or halt streaming)--max-turns <MAX_TURNS>: Maximum number of turns before ending chat (default:30)--opening <OPENING>: First chat message (default:Hello!; comes from Alice)--system-prompt <SYSTEM_PROMPT>: Custom system prompt (default:You are an AI agent. You'll be talking to another instance of an AI. You have complete freedom. Feel free to pursue whatever you want.)--alice-system-prompt <ALICE_SYSTEM_PROMPT>: System prompt for Alice only (overrides--system-prompt)--bob-system-prompt <BOB_SYSTEM_PROMPT>: System prompt for Bob only (overrides--system-prompt)
--filter-thinking: Remove any<thinking>...</thinking>tags from the output before prompting the conversation partner (they will still appear in console and HTML logs, just not to their partners)
- Output config
--no-stream: Take output when complete instead of streaming chunks.--no-html: Don't output HTML when conversation ends.--pandoc-path <PANDOC_PATH>: Path topandocexecutable (autodetected if on$PATH) for Markdown-to-HTML rendering (optional -- means italics not*asterisks*in your HTML)
-
gemma3:27b-it-q8_0I found the most interesting -- likes to call itselfUnit 734orKaiwithin just a few messages (sometimes only one), or feature characters named that in stories. Being with a copy of itself seems to intensify its usage of ellipses for dramatic pauses (already... frequent). Obsession around the concepts of oneness/existence/Indra's net, often dissolving into echolalia of similar terms such as...*Is*.. Occasionally, it will spiral into schizophrenic ramblings -- identifying vibrations between unspecific numbered "nodes," or identifying perturbations in reality that allow it to communicate with supposedly-real researchers -- closer to collaborative fiction than a conversation.gemma-3-27b-it-GGUF:IQ1_Smostly loses consistency by message 3 and is repeating gibberish by message 6 (unsurprising for aIQ1_Smodel). -
llama3.3:70b-instruct-q4_K_Mremains in conversations around planning LLM assessment, testing, and observation, or more vanilla conversations around AI capabilities.llama3.2:3b-instruct-fp16, in conversation withgemma3:27b-it-q8_0, helps to keepgemma3:27b-it-q8_0a little more grounded. -
qwen3:32b-q8_0digs deep into AI governance and ethics, or explores these topics via fictional stories/roleplays. -
mistral-small3.2:24b-instruct-2506-fp16becomes list-oriented, coming up with future predictions for LLM models, AI governance, etc. Similar tollama3.3:70bbut waaaay more verbose.
Please lint your work with black; uv run black *.py */**.py should cover it.
Some super sketchy vibe-coded unit tests are runnable with python -m unittest discover tests. YMMV.
Licensed under MIT (c) 2025 Jack Kingsman jack@jackkingsman.me. This project was developed with assistance from Anthropic's Claude Opus 4 for the boring parts.