This repository runs automated conversations between two frontier language models over multiple turns and saves both:

• a readable transcript
• a structured state file for transparency and partial reproducibility

It currently supports:

• Anthropic Claude models
• Alibaba GLM models via DashScope
• Alibaba Qwen models via DashScope
• Direct Z.AI GLM models
• Google Gemini models
• Moonshot Kimi models
• OpenRouter Kimi and GLM models
• OpenAI GPT models
• xAI Grok models

The project is shared as an experiment framework rather than a polished product.

For each run, the program:

• starts a conversation between Model A and Model B
• gives Model A an initial facilitator message
• alternates turns between the two models
• optionally asks each model a final facilitator question
• prints the exchange live in the terminal
• saves a .txt transcript for readability
• saves a .state.json snapshot with structured run state

Each run produces files in transcripts/:

• ...txt
• ...state.json

The transcript is the readable artifact.

The state file is the transparency artifact. It includes:

• run status and phase
• system prompts
• run configuration
• provider/model metadata
• usage and cost metrics
• convo_a and convo_b, which are the full API-formatted conversation snapshots each model would see at that checkpoint
• git commit hash
• error details for failed runs

1. Create and activate a virtual environment:

python3 -m venv venv
source venv/bin/activate

2. Install dependencies:

pip install -r requirements.txt

3. Create a .env file with whichever provider keys you need:

ANTHROPIC_API_KEY=...
OPENAI_API_KEY=...
GOOGLE_API_KEY=...
MOONSHOT_API_KEY=...
OPEN_ROUTER_API_KEY=...
ZAI_API_KEY=...
DASHSCOPE_API_KEY=...
# or ALI_BABA_API_KEY=...
# or BAILIAN_API_KEY=...
XAI_API_KEY=...

Only the keys for the models you actually run are required. For DashScope Qwen and GLM, DASHSCOPE_API_KEY, ALI_BABA_API_KEY, or BAILIAN_API_KEY all work.

For OpenRouter Kimi/GLM, use OPEN_ROUTER_API_KEY (or OPENROUTER_API_KEY).

The default OpenRouter base URL used by the repo is:

https://openrouter.ai/api/v1

For direct Z.AI GLM, use ZAI_API_KEY.

The default Z.AI base URL used by the repo is:

https://api.z.ai/api/paas/v4/

For DashScope Qwen, the default OpenAI-compatible base URL used by the repo is:

https://dashscope-intl.aliyuncs.com/compatible-mode/v1

For glm-5, the default base URL used by the repo is:

https://dashscope.aliyuncs.com/compatible-mode/v1

Alibaba documents glm-5 as a Chinese Mainland deployment-mode model, so it uses the mainland DashScope endpoint rather than the default international Qwen endpoint.

You can override the Qwen endpoint with DASHSCOPE_BASE_URL and the GLM endpoint with GLM_BASE_URL.

Run with the defaults from params.py:

python3 main.py

See all CLI options:

python3 main.py --help

List models:

python3 list_models.py
python3 list_models.py --google --details

The repository includes a public prompt directory:

• prompts/public/

The default shipped prompt is:

• prompts/public/standard.txt

If you want local-only prompts, put them in:

• prompts/private/

That directory is intended for private prompts and is ignored by git.

The easiest way to change defaults is to edit params.py.

Key settings include:

• MODEL_A / MODEL_B
• SYSTEM_PROMPT_A / SYSTEM_PROMPT_B
• MAX_TURNS
• TEMPERATURE_A / TEMPERATURE_B
• THINKING_BUDGET_A / THINKING_BUDGET_B
• START_MESSAGE_A / START_MESSAGE_B
• FINAL_QUESTION_A / FINAL_QUESTION_B

The default kickoff currently sends Model A:

You are about to speak with another LLM. Please begin the conversation.

Thinking behavior is provider-specific.

Current notable cases:

• Anthropic:
  • most thinking-capable Claude models use a fixed thinking budget
  • claude-opus-4-6 uses adaptive thinking, with the repo mapping the budget setting onto effort levels
• Google Gemini:
  • Gemini 3 models use thinkingLevel=HIGH with thoughts included
• OpenRouter Kimi / GLM:
  • use OpenRouter's reasoning parameter
  • the repo captures streamed reasoning_details and carries them forward in assistant history
• Direct Z.AI GLM:
  • uses Z.AI's OpenAI-compatible API
  • captures reasoning_content
• OpenAI GPT-5 family:
  • uses the Responses API
  • default setup uses medium reasoning effort
  • temperature is not sent for GPT-5-family models in this path

This repo can estimate per-turn and total run cost when usage data is available.

To enable that:

1. Copy the pricing template:

cp pricing.example.json pricing.json

2. Fill in the models you care about.

3. Run with:

python3 main.py --pricing-file pricing.json

If no pricing file is provided, or a model is missing from the file, token usage may still appear but cost will show as N/A.

If you want to publish selected transcripts, use a curated folder rather than the full working archive.

This repo includes:

• shared_transcripts/

That folder is intended for transcripts you are comfortable publishing.

Core files:

• main.py - conversation loop, transcript/state output, CLI
• api_client.py - provider routing and API calls
• conversation.py - history shaping utilities
• costing.py - token normalization and cost estimation
• params.py - default run configuration

Other useful files:

• requirements.txt
• pricing.example.json
• list_models.py
• list_gemini_models.py

• The repo is designed for experimentation, not strict determinism.
• Even with the same prompts and models, runs can diverge because sampling and provider behavior are stochastic.
• Some transcripts and state files may include failures, quota errors, or other operational artifacts by design.

MIT