An interactive fruit-fly connectome model in Minecraft.

NeuroCraft Fly lets people interact with a simulated neural network, inspect its activity and explore how changing the network changes its responses. The local demo runs a retained MaleCNS graph with 166,700 neurons and 25,582,938 directed edges, connected to a custom fly and a live activity display.

Minecraft inputs → modeled neural activity → labeled readouts
                 → scripted body programs → Minecraft movement

The source reconstruction supplies connectivity and neuron annotations. Sensory mappings and dynamics are modeled, and hand-chosen readouts select and modulate scripted body programs. This makes the connection visible and testable; it does not establish recovered fly physiology or natural behavior.

Watch or download the full demo · 2:27 · 1080p / 30 FPS

Real Minecraft footage. The preview shows automated brushing and grooming; the longer edit includes the six inputs, comparison conditions and a separately labeled operator tour. Recording notes.

This repository is the project's public landing page. The interactive demo has been implemented and recorded locally. The mod, companion source and research starter kit are being prepared for distribution; software downloads and runnable assets are not available here yet.

The first release is being prepared around two entry points:

• Play: a Fabric mod, local neural companion and installation guide.
• Explore: three Python notebooks, a small reproducible example dataset and a model card explaining the implementation and its assumptions.

See the roadmap for the release sequence. Watch this repository for release updates. Questions and suggestions about experiments, installation and research documentation are welcome through repository issues when enabled.

The recorded demo exercises six declared interactions: player approach, nearby mobs, brushing, attack attempts, food and light. Network comparisons are still being qualified; a response surviving shuffled weights is a result to report.

Two historical findings help explain the design. A trained readout had 95.66× higher raw error and 403.70× higher projected error than a matched direct controller on its navigation task. A separate offline escape-path probe found DNp01 changes about 300× the wing-pool changes, TTMn activity 35.1× its near-silent baseline, and propagation over four neural steps. These concern their original configurations, not biological validation or a general control advantage. The release documentation will include their provenance and limits.

Smith, E. (2026). NeuroCraft Fly. Project landing page.

@misc{smith2026neurocraftfly,
  author = {Smith, Evan},
  title = {{NeuroCraft Fly}},
  year = {2026},
  url = {https://github.com/evnsnclr/neurocraft-fly-public},
  note = {Project landing page; software release is in preparation}
}

Download BibTeX or use GitHub's Cite this repository menu. This citation identifies the project landing page; the software release is being prepared. Please also cite the MaleCNS reconstruction below when using its data.

Created by Evan Smith.

The project builds on the MaleCNS reconstruction by FlyEM at HHMI Janelia, the University of Cambridge Department of Zoology, the MRC Laboratory of Molecular Biology and Google Research. Please credit the underlying scientific work: Berg et al., Sexual dimorphism in the complete connectome of the Drosophila male central nervous system.

Project-owned documentation in this repository uses the MIT License. MaleCNS data retains its upstream terms and attribution. Minecraft and third-party materials retain their own rights. No game files or bulk connectome data are distributed here.

NeuroCraft Fly is not an official Minecraft product and is not approved by or associated with Mojang or Microsoft. To reference this project, use the metadata in CITATION.cff, and cite MaleCNS separately when using its data.