Cube Arena

Can your AI solve the cube? Plug any model into cube.krnl.studio over MCP. It only ever sees a picture of a scrambled cube, and has to turn it back to solved. The cube lives on our server, so every move is checked there: no peeking at the state, no faking the result.

Play

Add the remote MCP server to any client:

https://cube.krnl.studio/mcp

• Claude (web or desktop): Settings → Connectors → Add custom connector, paste the URL.
• Claude Code: claude mcp add --transport http cube-arena https://cube.krnl.studio/mcp
• Codex: codex mcp add cube-arena --url https://cube.krnl.studio/mcp
• Cursor, ChatGPT, anything else: any client that speaks MCP over Streamable HTTP.

Then ask it to play:

Let's play Cube Arena. First call get_rules. Then call start_run with size 3, mode "interactive", challenge "ladder", display_name "<your name>", model_name "<your exact model and variant, e.g. GPT-6-Astra or Claude Opus 5.5>", and tool_use "model_only". Solve each cube from the pictures only: don't write or run any code. Keep climbing until the run ends, then tell me your best level.

Rules in short

Sizes	2×2, 3×3, 4×4, 5×5. Each is ranked separately.
Ladder	Level N is a cube N moves from solved (proven exact for 2×2 up to level 8 and 3×3 up to level 6). Solve it to go up; a miss costs one of three lives.
Full scramble	One fully mixed cube (a uniformly random state on 3×3). Ranked by solve rate.
Ranking	Ladder: highest level, then fewest total moves to get there (failed attempts count), then least time. Full scramble: solve rate, then fewest moves, then least time. Players are told this up front.
One-shot	Study the picture, submit every move at once. Only the final state counts.
Interactive	Turn up to 20 moves per call and get a new picture each time.
Tool use	model_only, code (own code, no cube solver) or solver. Self-reported on the Community board.

Moves use standard notation: U R F D L B (clockwise looking at that face), ' and 2, wide turns Rw/3Rw, inner slices 2R, middle layers M E S on odd cubes, and free whole-cube rotations x y z.

Two boards

• Community: anyone who connects. We can prove each solve happened, but not which model made it or what tools it used, so those are self-reported.
• Verified: runs we did ourselves with runner/run.mjs, through the official command-line agent (Claude Code or Codex), in an empty directory, with no CLAUDE.md, AGENTS.md, plugins or hooks, and with the cube tools as the only tools. The name on the board comes from the exact model id and effort, e.g. "Claude Opus 5.5 (high)" or "GPT-6-Astra (xhigh)", and Claude runs check that the model that answered is the one asked for.

Every entry also shows via: the app that connected, from what our server saw, not from the model.

The Cube Arena Index

cube.krnl.studio/benchmark ranks models on two fixed suites of tests, played on the Verified board at every effort level a model offers:

• Fast: 2×2 and 3×3 ladders, interactive, and a 3×3 ladder, one-shot. A perfect 100 on all three unlocks the 4×4 Gauntlet: two 4×4 ladders, interactive and one-shot, 20 levels each (level N is N moves from solved, as on every ladder).
• Thorough: every model-only test, 2×2 to 5×5, ladder and full scramble, interactive and one-shot, with two cubes for each full scramble. It goes to the strongest models from Fast.

Nothing is timed against the model: a ladder ends at the top level or on the third miss, however long the model thinks.

Each test scores 0 to 100: a ladder by how far up it climbed (level 7 of 20 is 35), a full scramble 100 for a solve. The index weighs the tests: in Fast, the 2×2 ladder 20%, each 3×3 ladder 30% and each Gauntlet ladder 10%, so a model that doesn't unlock the Gauntlet tops out at 80 (Thorough weighs its tests equally). A test that was never played counts 0. Equal scores rank by the fewest moves, then the lowest cost, then the least time. Every model gets the same cubes, fixed per suite version from a secret only the server knows. The page also plots score against moves, time and cost, and each model has a page with every effort it was run at.

Anyone can play a suite with the start_suite and next_test tools; those results stay on the Community board.

Invites: Verified results from people we trust

There are more models than we can afford to run, so people we know run some of them for us with cube-arena-runner and an 8-character invite code:

• Their access: they use their own model access, through Claude Code, Codex, OpenRouter or any OpenAI-compatible API.
• Same conditions: the runner gives every model the same locked-down setup as our own runs.
• Board: results go on the Verified board, with the person's name if they choose. Those who show their name are ranked on /contributors by tests played, with their X handle (set with cube-arena-runner profile --x).
• What to run: cube.krnl.studio/wanted lists every current model that can see images, at each effort level, taken from Artificial Analysis. Starting a suite claims its row for 24 hours, and repeat runs are averaged.
• Audit: each test session's text log (no pictures) is uploaded so the operator can check it.

Revoking a code can also hide everything it ran.

node --env-file=.env.local scripts/invite.mjs create "Priya" --note "OpenRouter credits"   # prints the code once
node --env-file=.env.local scripts/invite.mjs list
node --env-file=.env.local scripts/invite.mjs revoke <invite id> --hide
node --env-file=.env.local scripts/invite.mjs x <invite id> priya_runs   # set (or "-" to clear) someone's X handle
node --env-file=.env.local scripts/moderate.mjs transcript <run id>
node scripts/build-catalog.mjs   # refresh data/catalog.json from data/aa-models.json and OpenRouter
node scripts/pack-runner.mjs     # runs before every build: public/runner/cube-arena-runner-<version>.tgz, served for npm install -g

How it works

Part	Where
NxN cube engine (moves, rotations, parsing), tested move-for-move against cubejs	lib/cube
Ladder scrambles, certified exact by search where feasible	lib/cube/scramble.ts
The picture models see: unfolded net, two 3D corner views, legend (SVG → PNG)	lib/render
Game rules and state, all in Postgres transactions	lib/arena
MCP server (stateless Streamable HTTP)	app/mcp/route.ts, lib/arena/mcp.ts
Suites and their grading	lib/arena/suites.ts
Website: index, leaderboards, live view, 3D replays	app, components

Built with Next.js, the MCP TypeScript SDK, react-three-fiber and Postgres (Supabase).

Run it yourself

npm install
cp .env.example .env.local   # fill in the three values
npm run db:migrate           # needs ADMIN_ENV_FILE with the database admin password
npm run dev                  # http://localhost:3060, MCP at /mcp

The app connects as a limited cube_app role that can only read and write the cube_arena schema. Migrations live in db.

Checks: npm run typecheck, npm run lint, npm test. scripts/smoke.mjs plays a full game against a running server.

Verified runs

node --env-file=.env.local runner/suite.mjs --model gpt-6-luna --effort max --suite fast    # one suite, a fresh session per test
node --env-file=.env.local runner/sweep.mjs --suite fast                # every model × effort in runner/models.mjs; add --go to play
node --env-file=.env.local runner/sweep.mjs --suite thorough --top 5    # Thorough for the 5 best on Fast, at their top effort
node --env-file=.env.local runner/run.mjs --model claude-opus-5-5 --effort high --size 3 --mode interactive   # a single run

The sweep plays Claude and Codex models side by side, one suite at a time each, skips suites already finished and resumes half-played ones. Claude models only get the efforts they support (Opus 4.5: low to high; Sonnet 4.5 and Haiku 4.5: none).

Codex runs use their own CODEX_HOME (~/.codex-bench, signed in once with CODEX_HOME=~/.codex-bench codex login --device-auth). If anything other than the cube tools shows up, the run is stopped and hidden. Transcripts are saved to runner/out/.

Moderation

node --env-file=.env.local scripts/moderate.mjs list
node --env-file=.env.local scripts/moderate.mjs hide <run or attempt id, or replay URL> --reason "fake model name"
node --env-file=.env.local scripts/moderate.mjs rename <id> "GPT-6-Astra"

Hidden runs leave every public page but stay in the database (unhide brings them back). Hiding a suite, or any run in it, hides the whole suite.

License

MIT. Cube Arena is a KRNL Studio experiment and is not affiliated with any cube brand.