↑Swipe up to play full screen
  • A Craftax day is 300 steps.
  • Hunger and thirst count steps, not seconds.


W A S Dwalk
Shift + W A S Dturn without stepping
← → ↑ ↓look around
Spaceact on the tile ahead
1–0, −pick a hotbar slot
Einventory and crafting
Space or Qclimb a ladder
Zsleep (energy not full)
Rrest (health not full)
Xplace block ahead
C V B2D, 1P/3P, bird's-eye
Mcolor, phosphor, laser
Shift + Mthe same, backward
F5First or third person
Tthe next theme
Shift + Tthe same, backward
Pagent plays / take over
Ksound on or off
Ggraphics quality
Hthese controls
Escpause

The agents you can watch: pick one from the bot window in the game; each runs in your browser.

Rekursiv priml 19486% completion

rekursiv.ai's priml recipe (reproduced) · 15M parameters · 194 / 226 in priml's own evaluation

A convolutional board encoder and 4 MinGRU layers of 1,024 that also read their last action, fine-tuned to fight the final boss: beats the necromancer in 43% of games.

PufferLib 16272% completion

PufferAI · 14M parameters · 162 / 226 in PufferLib's own evaluation

PPO with 4 minGRU layers of 1,024, trained by PufferLib 5.0; reaches the sixth floor. Converted unchanged to ONNX to play here.

TransformerXL PPO19% completion

Gautier Hamon · 4.7M parameters · 41.9 / 226 training return at 1e9 steps

PPO with a gated Transformer-XL remembering its last 128 steps; of the ten published Craftax runs, the one that goes deepest.

PPO-RNN (Craftax paper)15% completion

Matthews et al., Craftax paper (reproduced) · 6.9M parameters · 35 / 226 (15.3%, reported in the Craftax paper)

The Craftax paper's PPO-RNN baseline, retrained to its reported score: one GRU of 512.

NeuralCrystal 2712% completion

NeuralCrystal · 21M parameters · 27 / 226 over 580 games

An optical neural network: a simulated crystal with 80 etched surfaces sees the game board, lights up its next move, and changes its recurrent memory.

Try your agent
  1. Export it to ONNX: obs [1, 843] in (the engine's packed observation), logits [1, 43] out; any memory as inputs fed back from outputs named X_out.
  2. Check it on the engine: python -m craftax_agents.check_model your.onnx
  3. Play it on agentcraft3d.com: Test My Agent below (or Settings, Agent), with a link to the file (Hugging Face works). It stays in your browser.
Publish your agent
  1. Open a pull request to craftax-agents adding its entry to models.json: its name, a line about it, and the link to its file.
  2. Once it's merged, your agent is in everyone's list on agentcraft3d.com.
Loading…