Yea Claude sucks at minesweeper (and many spatial reasoning tasks), but isn’t an idea of MCPs is that Claude should be able to ask an MCP what the next best move is rather than figuring it out itself? Like offload hard thinking/reasoning to purpose-built solvers because they are deterministic? Though I guess you’d expect a reasoning model to be able to come up with its own solvers on the fly, especially for well-known problems. Maybe having access to an MCP itself is confusing it?
MCP generically connects Claude to an environment, so it can be used to connect Claude to minesweeper, and then also to connect it to a CSP solver. Or a calculator and a dictionary. Or your GitHub and a devbox. Or Unity and a 3d printer.
This seems like intended usage? The server actually executes the moves and interacts with the environment, the core orchestration or reasoning is offloaded to claude?
Right, reasoning was offloaded to Claude. Claude is obviously terrible at Minesweeper. I’d like to see Claude orchestrate both playing the game as well as using another MCP to help it pick the next best move. Otherwise it’s just wasteful getting an LLM to reason about an already solved problem, it just chews up API requests. I followed the Manifold market for a while getting ChatGPT to play Sudoku —- each puzzle cost ~$20 to complete.
