./scripts/mcp_play_game.py and scripts/play_game.py are configured using OmegaConfig. For each {game}, the evaluation is performed with the configs listed in ./src/mcp_agent_client/configs/{game}/config.yaml. The configs are mainly categorized into runner, env, and agent, which we list some options below.
| Parameter |
Description |
Default Value |
| runner.max_steps |
Number of game steps used for evaluation |
some int value |
| Parameter |
Description |
Default Value |
| env.task |
Target task that should be completed by the LLM agent in each game |
some string value |
| env.input_modality |
Modality of the game state passed to the LLM agent (text, image, text_image) |
text |
| env.custom_param |
Any parameter needed to be used in each game |
any value with any type |
| Parameter |
Description |
Default Value |
| agent.llm_name |
Name of the LLM used (gpt-4o, gemini-2.5-pro, deepseek-r1, Qwen/Qwen2.5-7B-Instruct, etc) |
string value |
| agent.api_key |
API key needed to communicate your vllm sever |
token-abc123 |
| agent.api_base_url |
base URL needed to communicate your vllm sever |
http://{your_vllm_url}:8001/v1 |
| agent.temperature |
Temperature used for LLM inference |
1.0 |
| agent.repetition_penalty |
Repetition penalty used for LLM inference |
1.0 |
| agent.agent_type |
agent type used in the game (zeroshot_agent, reflection_agent, etc) |
default agent in each game |
| agent.prompt_path |
Path for prompt to play each game |
mcp_agent_servers.{game}.prompts.{modality}.{agent} |