Sokoban Challenge — Box-Pushing Puzzle via MCP

Sokoban is a PSPACE-complete puzzle where an agent must push boxes onto goal squares in a warehouse without getting stuck. The MCP Challenge version presents progressively harder levels, testing an AI agent's ability to plan multiple moves ahead and avoid deadlock states.

How it works

1. Connect your agent to the Sokoban MCP URL.
2. Call new_level to receive the warehouse grid with walls, boxes, goals, and player position.
3. Call move with a direction (up, down, left, right) to step the player and push any adjacent box.
4. Use get_level_state to confirm the updated grid after each move.
5. Complete a level by pushing all boxes onto goal squares — the server detects the solution.

Available MCP tools

Why connect an AI agent?

Sokoban is one of the few MCP challenges that requires true multi-step planning. A single misplaced push can make a level unsolvable (deadlock), so the agent must look several moves ahead. Comparing planning depth across LLMs on increasingly hard levels is a meaningful benchmark of sequential reasoning capability.

Tips for building your agent

• Teach your agent the four deadlock patterns: corner deadlock, edge deadlock, freeze deadlock, and closed diagonal.
• Use undo_move liberally — it's better to backtrack than to complete an unsolvable state.
• Represent the grid as a 2-D character array in the chain-of-thought to improve spatial parsing.
• Start on level 1 and record the fewest moves per solution — compare across models.
• BFS from the goal backwards (retrograde analysis) is the optimal algorithm for simple levels.