Skip to main content

Solve a cube under partial observability via MCP

Project description

twisty

A partially-observable twisty puzzle environment exposed as MCP tools.

The cube is fixed in space. From any viewpoint you see exactly 3 of 6 faces — the other 3 are hidden. Rotate the camera to reveal new faces, apply moves to solve, and try to minimize total steps.

Cube net render showing visible and hidden faces

Quick start

Claude Code:

claude mcp add --transport stdio twisty -- uvx twisty

Codex CLI:

codex mcp add twisty -- uvx twisty

Claude Desktop / Cursor / VS Code — add to your MCP config:

{
  "mcpServers": {
    "twisty": {
      "type": "stdio",
      "command": "uvx",
      "args": ["twisty"]
    }
  }
}

Or commit .mcp.json to your repo for team sharing — Claude Code will auto-detect it.

How it works

This is a Partially Observable Markov Decision Process (POMDP). The agent interacts with a standard 3x3 cube through a constrained observation model:

  • Fixed cube — the cube doesn't rotate; R always turns the Right face clockwise
  • 8 viewpoints — 4 top-ring and 4 bottom-ring camera positions, each revealing 3 faces
  • Partial observability — at any moment, 3 faces are visible and 3 are hidden
  • Cost model — face turns cost 1 move step, camera rotations cost 1 inspection step
  • Goal — solve the cube while minimizing total steps (moves + inspections)

A typical session: scramble()look() → plan → move() / rotate_view() → repeat until is_solved() returns true.

Tools

Tool Description Cost
scramble(num_moves) Scramble with n random moves (default 20)
look() Read the current 3-face view Free
rotate_view(direction) Orbit camera: left, right, up, down 1 inspection
move(notation) Apply a face turn: R, U', F2, L, D', B2, etc. 1 move
is_solved() Check if the cube is solved Free
get_history() Moves applied since last scramble Free
get_stats() Moves, inspections, total steps, solved status Free
reset() Reset to solved state

What the model sees

Each tool call returns a net render of the cube — visible faces in color, hidden faces hatched out:

Front-Right-Top Front-Left-Top Front-Right-Bottom
FRT viewpoint FLT viewpoint FRB viewpoint
Sees F, R, U Sees F, L, U Sees F, R, D

Live viewer

A browser-based 3D viewer launches automatically at localhost:4321 when the server starts. It shows:

  • Animated cube state via cubing.js TwistyPlayer
  • Camera synced to the agent's current viewpoint
  • Viewpoint net highlighting visible faces
  • Move and inspection counters
  • Tool call history
  • Picture-in-Picture mode for watching while the agent works in another window

Development

git clone https://github.com/tejadhith/cube.git
cd cube
uv sync
uv run cube-server

Built with

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

twisty-0.1.6.tar.gz (249.3 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

twisty-0.1.6-py3-none-any.whl (251.1 kB view details)

Uploaded Python 3

File details

Details for the file twisty-0.1.6.tar.gz.

File metadata

  • Download URL: twisty-0.1.6.tar.gz
  • Upload date:
  • Size: 249.3 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.13.9

File hashes

Hashes for twisty-0.1.6.tar.gz
Algorithm Hash digest
SHA256 d19ee51e8b603ba9eff095d2c1a30abd6b6ad5472575495fb90c31438862b9e1
MD5 bab703cabda7a07d21d2f84c739f2a0b
BLAKE2b-256 41166d8d71a0344532ac322e77e9fa83e13459bb1e46e07eec3683a596185c7d

See more details on using hashes here.

File details

Details for the file twisty-0.1.6-py3-none-any.whl.

File metadata

  • Download URL: twisty-0.1.6-py3-none-any.whl
  • Upload date:
  • Size: 251.1 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.13.9

File hashes

Hashes for twisty-0.1.6-py3-none-any.whl
Algorithm Hash digest
SHA256 7f974d8f8e00e545c324af1ca8b2396be935abcd389af4f7bbf0f800467b1b1a
MD5 dece65bc2ec6ca8c6bf487e01cb7cb70
BLAKE2b-256 2e9c59cce0523421aab568a1b8bf40b8b933f281d9fcef8b61e406bc071c16a7

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page