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?
dartos 50 days ago [-]
MCP is analogous to REST.
It doesn’t define what behavior should be where in a given app, just how to communicate what that behavior is and how to invoke it.
breckenedge 50 days ago [-]
RPC, not REST
paulddraper 49 days ago [-]
Like neither one.
RPC and REST are architectural patterns/philosophies, not protocols.
SOAP and HTTP are protocols, like MCP.
dartos 49 days ago [-]
If we’re being technically correct, yes, but I was just trying to give an analogy to someone who I figured was pretty new to web tech.
paulddraper 49 days ago [-]
We have time, but there is a lot to do.
fragmede 50 days ago [-]
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.
breckenedge 50 days ago [-]
Curious, I’ve yet to see Claude effectively use Unity.
2 days old :D where are people seeing these releases announced? I swear we need an MCP weekly email digest
emersonmacro 50 days ago [-]
Pulse MCP has a weekly email digest
prats226 49 days ago [-]
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?
breckenedge 48 days ago [-]
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.
tmitchel2 49 days ago [-]
It feels nuts to me that there is a push away from strict APIs to conversational interfaces for products and then the actual technology itself under the hood is translating that into a strict set of API calls in order to understand something. Would it not be better to seek interoperability with fairly well scripted natural language handshake. I feel like MCP is built for understanding language and Syntax to a greater degree but not random tools and APIs.
_joel 50 days ago [-]
Maybe tell it it's a champion Minesweeper player and that loss is not an option :)
lgas 50 days ago [-]
Based on some of the recent leaked prompts I imagine something like "The mines are connected to actual bombs that will blow up your family if you make a mistake" might work best.
bredren 50 days ago [-]
Realize this must be somewhat tic, but curious about a link to example leaked related prompt?
By just looking at the README from the repo (Would look more deeply into this later) you're replying with an image of the current status?
If you expect Claude to interpret the image corretly may be you're asking for too much.
Besides the image (Gotta say I didn't know you could fed Claude images in MCP that's incredible cool) I'd rather / also return some json payload that informs Claude which positions has "cleared" neighbor positions, and their value. E.g.:
(Might not be valid json, just wrote that by hand on the fly)
I would report only on the positions that has cleared neighbors, and hope for the best. Good luck!
(Impressive work BTW, I think we haven't even started to see the possibilities of MCP and I love people being this imaginative)
ericol 50 days ago [-]
An interesting exercise here would be to make the MCP server show the actual state in a window and return to Claude just the json payload with the status.
It doesn’t define what behavior should be where in a given app, just how to communicate what that behavior is and how to invoke it.
RPC and REST are architectural patterns/philosophies, not protocols.
SOAP and HTTP are protocols, like MCP.
Did Mario not do it for you? https://youtu.be/dCC7QoV5a6E
I would report only on the positions that has cleared neighbors, and hope for the best. Good luck!
(Impressive work BTW, I think we haven't even started to see the possibilities of MCP and I love people being this imaginative)
use this format for board representation
``` { "game_state": { "board_size": { "width": 9, "height": 9 }, "mines_total": 10, "mines_flagged": 2, "game_status": "in_progress", // "in_progress", "won", "lost" "time_elapsed": 45, "difficulty": "beginner" // "beginner", "intermediate", "expert", "custom" }, "board": [ ["1", "?", "?", "2", "1", "1", "1", "1", "0"], ["1", "2", "?", "2", "?", "1", "1", "?", "0"], ["0", "1", "1", "2", "1", "1", "1", "1", "0"], ["0", "0", "0", "0", "0", "0", "0", "0", "0"], ["1", "1", "0", "0", "0", "0", "0", "0", "0"], ["?", "1", "0", "0", "0", "1", "1", "1", "0"], ["1", "1", "0", "0", "0", "1", "F", "1", "0"], ["0", "0", "0", "0", "0", "1", "1", "1", "0"], ["0", "0", "0", "0", "0", "0", "0", "0", "0"] ], "last_action": { "action_type": "reveal", "x": 3, "y": 2, "result": "revealed_number", "timestamp": 1710931245 } } ```
and this format for llm response generation
``` { "action": { "action_type": "reveal", // "reveal", "flag", "unflag", "chord" "x": 5, "y": 3, "confidence": 0.95, "reasoning": "This cell is surrounded by revealed cells with low numbers, making it a safe choice." }, "game_analysis": { "identified_safe_cells": [[5, 3], [2, 5]], "identified_mine_cells": [[6, 1], [8, 2]], "uncertain_cells": [[1, 1], [2, 2]], "strategy": "Targeting isolated revealed areas first to gain more information." } } ```
It should fix all your issues plus also make it cheaper to play
* What is the data format it gets? Does it unambiguously correspond to output (i.e. without mistaking rows for cols, or indexes starting at 0 or 1)?
* What is the prompt?
* Is the model allowed to think? (If it is just JSON response, I expect it to suck, as tokens are units of thinking.)
MCP = Model Context Protocol
https://modelcontextprotocol.io/
"MCP is an open protocol that standardizes how applications provide context to LLMs."