A generic MCTS framework driven by an external AI agent, supporting game simulations (Chess, Shogi, Ligand design) and CartPole control policy evaluation.
This MCTS simulator server has moderate definition quality with significant gaps. All 6 tools have names following verb_noun conventions (reinitialize, run, get patterns), but descriptions are inconsistent in depth. Input schemas are present for all tools with proper JSON Schema structure, but parameter descriptions vary widely. Several tools lack detailed descriptions of their behavior and return values. The server includes a well-designed prompt (mcts_autonomous_search) that compensates somewhat by providing workflow context, but individual tool definitions could be clearer for LLM reasoning.
Executes Python control policy code in a Gymnasium CartPole-v1 environment. Returns normalized score (0.0-1.0) and evaluation logs.
Tool descriptions lack detail on return values and expected output structure. 'Returns the best move found so far' does not specify format, type, or how to use it next.
Missing output schema documentation in tool definitions. The fastmcp_server.py source does not include explicit return type schemas or field documentation for any tool response.
run_mcts_round parameter 'actions_to_expand' has a generic description ('Optional list of action strings') but lacks guidance on when to use it, how to obtain valid values, or what happens if invalid actions are provided.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 13 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 45 | - | v1 |
reinitialize_mcts requires state_module and state_class as strings, but the descriptions do not clarify error handling if these do not exist, or what to do if the class does not inherit from GameStateBase.
No error recovery guidance. If reinitialize_mcts fails to load a module or class, the tool provides no actionable error message or recovery suggestion.