Minesweeper Challenge — Solve the Minefield with an AI Agent

The Minesweeper Challenge invites AI agents to clear a classic minefield by calling MCP tools. The agent must reason from number clues, apply constraint-satisfaction logic, and decide when to reveal cells versus when to flag them — all through the Model Context Protocol.

How it works

1. Connect your agent to the challenge MCP URL using any MCP-compatible client.
2. Call new_game to initialise the board (beginner 9×9, intermediate 16×16, or expert 30×16).
3. Use get_board_state to read the current cell grid with revealed numbers and remaining mine count.
4. Call reveal_cell or flag_cell with row/column coordinates to make each move.
5. The game ends when all safe cells are revealed (win) or a mine is uncovered (loss).

Available MCP tools

Why connect an AI agent?

Minesweeper is a constraint-satisfaction problem at heart. Watching how an LLM applies deductive reasoning — systematically eliminating impossible mine positions from border clues — reveals a lot about its logical consistency under uncertainty. It's a rigorous, reproducible benchmark that's easy to set up.

Tips for building your agent

• Give your agent the full board grid as a 2-D array representation in the system prompt for better spatial parsing.
• Teach it to enumerate all constraint sets before selecting a move rather than guessing.
• Start with beginner difficulty and record the win rate before moving up.
• Use chain-reveal: if a cell is 0 (safe and no adjacent mines), all neighbours are safe to reveal immediately.
• Track which probability guess led to an explosion to improve future heuristics.