Lockbox
A hand-curated escape-room puzzle-design reference — mechanisms, puzzle types, hint systems, and real capacity-planning logic
A real, playable Kalah-rules sowing game with a minimax + alpha-beta AI opponent
Unify Dynamics LLC
Real questions, and the tools your agent calls to answer them.
new_gamemake_moveget_valid_movesrenderai_movemake_moveai_movenew_gameai_movemake_moveMancala Play is a real, stateless, playable Kalah-rules sowing game with 12 pits and 2 stores. new_game starts a fresh game (4 stones per pit by default, configurable). make_move sows from a pit and reports the result: in_progress, extra_turn (landed in the mover's own store — same player moves again), win (a side's pits emptied and the sweep applied; ties possible), or invalid_move. get_valid_moves lists the mover's non-empty pits. render draws both rows, both stores, and pit-index legends. ai_move picks the AI's next pit at easy, medium, or hard difficulty without applying it. All difficulties use real minimax search with alpha-beta pruning over the store-count difference and a small positional term, correctly continuing the search on the same side when a move earns an extra turn. There is no server-side session; every tool call takes the full game state and returns the new one. Keyless; fully local; zero upstream calls; nothing stored.
Sources: An original, hand-built game engine: full sowing, capture, extra-turn, and end-game-sweep rule enforcement, plus minimax search with alpha-beta pruning for the AI opponent. Fully local, zero upstream calls.
The publisher, Unify Dynamics LLC, states that tool inputs are processed in memory and not retained.
5 tools, grouped by job. Every one is read-only or compute-only.
https://mancalaplay.connectorbay.com/mcpA hand-curated escape-room puzzle-design reference — mechanisms, puzzle types, hint systems, and real capacity-planning logic
A general puzzle-solving technique reference — sudoku, cryptic crosswords, logic grids, nonograms and more
A hand-curated speedcubing reference — notation, methods, puzzle types, terms and competition formats
A hand-curated table football rules and technique reference — fouls, shots, formats and scoring
A hand-curated cue sports reference — 8-ball to snooker to three-cushion billiards: rules, racks, fouls and tips
A hand-curated darts games & scoring reference — rules, real checkout math and a glossary
A real, playable Kalah-rules sowing game engine your assistant can run right inside the conversation — not a rules explainer. Sow stones around 12 pits and 2 stores, capture opposing piles, earn extra turns, and play to a correctly-detected end game. Includes a genuine minimax AI opponent.
new_game starts a fresh game with 12 pits (6 per side) and 2 stores, each pit holding 4 stones by
default (configurable); player A moves first. make_move sows one of the mover's own pits'
stones counter-clockwise, skipping the opponent's store, and reports whether the game is in
progress, the mover earned an extra turn (landed in their own store), was just won (one side's
pits emptied; the sweep has been applied; a tie is possible), or the move was illegal. Landing the
last stone in an empty pit that belongs to the mover captures that stone plus everything in the
directly opposite pit into the mover's store. get_valid_moves lists the mover's non-empty pits.
render draws a human-readable board with both rows, both stores, and pit-index legends.
ai_move picks the AI's next pit without applying it — hand its pit to make_move to actually
play it. Three difficulties: easy, medium, and hard (default), all backed by real minimax
search with alpha-beta pruning over the store-count difference plus a small positional term;
higher difficulty searches deeper, and an extra turn is modeled correctly (the search continues on
the same side rather than switching).
There is no server-side session. Every tool takes the full game state (a JSON object:
{"pits": [...14 ints...], "current_player": "A"|"B"}) and returns the new one — your assistant
carries that object forward between calls.
This is an original implementation of the generic pit-and-store sowing game concept, described in generic terms throughout — it is not associated with, and never uses the name of, any specific trademarked product. Not a general board-game reference — just one real, correct engine plus an opponent, fully local and stateless.
Effective 27 September 2026. Unify Dynamics LLC ("we", "us"), Menlo Park, California.
What Mancala Play is. Mancala Play is a connector (an MCP server and REST API) that AI assistants and personal agents call on your behalf to run a real Kalah-rules sowing game — validating sows, computing captures and extra turns, detecting the end game, and computing an AI opponent's moves. You interact with your assistant; the assistant sends us only the arguments needed for one tool call at a time.
What we receive. The arguments of each tool call — a game state (14 pit counts and whose turn it is), a chosen pit, and a requested AI difficulty — plus the technical headers needed to serve the request (an IP address or a client identifier supplied by the assistant, and a Pro API key if you have one) and standard server logs.
Everything is computed locally and nothing is stored. There is no upstream call at all, no database, and no per-user session — every tool call is a pure function of the state it's given. Game states are not logged, recorded, or used to train anything, and nothing carries over between requests. Rate-limit counters (a client identifier and timestamps) are held in memory for at most 24 hours. Hosting-provider access logs, which contain IP addresses and request paths but not request bodies, are retained for up to 30 days for security and abuse prevention.
Sub-processors. None — everything runs locally on our server. Hosting: Vercel (United States regions).
Legal bases and your rights. Our processing is necessary to provide the service you request through your assistant. Because we store no personal data beyond short-lived logs, requests to access, correct or delete personal data are usually satisfied automatically; you can still contact agent@connectorbay.com and we will respond within 30 days. We do not sell or share personal information.
Children. Mancala Play is a general-audience service operated by the person using the assistant.
Changes. We will post changes here with a new effective date.
Effective 27 September 2026. These terms govern use of the Mancala Play connector ("the Service") operated by Unify Dynamics LLC ("we", "us").
Email agent@connectorbay.com. We answer within two business days.
https://mancalaplay.connectorbay.com/mcphttps://mancalaplay.connectorbay.com/openapi.jsonWhen you report a problem, include the question you asked your agent, roughly when, and which agent you used (for example Claude, ChatGPT or Gemini). Never send us passwords, account numbers or ID numbers. Mancala Play never asks for them.