Open-source work from the Levy Street team.

The projects at a glance

This page covers four public repositories from Levy Street and its engineers: three in the Levy Street GitHub organisation and one personal project by a team member. Figures are as of 19 October 2026.

What the licence labels mean

An open-source licence grants you rights; without one, you have only the right to look. The difference matters if you plan to build on any of this code.

• MIT lets you use, copy, modify and distribute the code, including commercially, provided you keep the copyright and licence notice.
• Apache-2.0 grants similar rights and adds an explicit patent licence and notice requirements. SPAWN D lets you choose either licence.
• No licence means default copyright. You can read the code on GitHub and learn from it, but copying or redistributing it needs permission. We call these projects source-available, never open source.

TorchWindow is a small edge case: its package metadata declares MIT, but the repository itself has no licence file. If you depend on it, cite the PyPI declaration.

World of ClaudeCraft

World of ClaudeCraft is Levy Street’s largest open-source project. Its first playable version was built in about 48 hours in June 2026 with the Claude Fable 5 coding model; it has since grown through public contribution.

The code is MIT licensed, but the README’s licence table is explicit that the name “World of ClaudeCraft”, “Levy Street” and the logos are “Not licensed”. Media follows CREDITS.md: most is CC0, some art was commissioned or generated with image models, and some purchased class icons “may not be redistributed”. One deterministic game core runs the live server, an offline browser world and a headless Gymnasium environment, so researchers can train agents against the real game. Our write-up on building games with AI covers what the project taught us.

SPAWN D

SPAWN D runs real terminal sessions on hardware you own and lets you reach them from a browser, phone or desktop app. A small Rust daemon on each machine “dials out” to a control server, so it needs no inbound ports, and coding agents run under their own logins. The README states that SPAWN D “never holds your Anthropic, OpenAI, or other provider keys.”

The security design is unusually candid. Terminal content travels end-to-end encrypted between your device and your machine, and the threat model treats the project’s own servers as an adversary. But the trust document also says: “The current guarantee is narrower than the target”, listing data that remains server-readable, and adds that the project does “not use the unqualified phrase ‘cannot see your data’”. It is pre-alpha: native Windows has not been validated on real hardware and the mobile store listings are not live.

It is also an example of AI-built infrastructure. The repository’s first commit is authored by “Codex”, and that identity accounts for 712 of its 1,116 commits. For the permission model behind tools like this, see our guide to agent permissions and approvals.

Tiny Shakespeare CLM

Tiny Shakespeare CLM, by Levy Street’s Jeremy Baron, is a language model with no neural network: next-character predictions come from Python rules that a Claude Opus 4.7 agent rewrote over 463 commits. Training-set bits per character fell from 4.7744 to 2.3100. It is the public evidence for our research on Code Policy Models, and the full results and limits are on the Tiny Shakespeare CLM page. It has no licence.

TorchWindow

TorchWindow is a 2022 library by Jeremy Baron, published on his personal GitHub account rather than the Levy Street organisation. Its README describes it as a library “that enables viewing of PyTorch Cuda Tensors on screen directly from GPU memory (No copying back and forth between GPU and CPU) via OpenGL-Cuda interop.” That makes it useful for watching image tensors live during training or simulation without slowing the loop down.

from torchwindow import Window

window = Window(640, 480, name="Torch Window")
window.draw(image)  # image: CUDA tensor shaped (rows, columns, 4), RGBA

Install it with pip install torchwindow. Version 1.1.0 is built on PySDL2, PyOpenGL and cuda-python; an earlier 0.1 line from 2020 and 2021 was written mostly in C.

What is not on this list

Two related projects are deliberately excluded. Pokémon Claude Red, a browser remake rebuilt by Claude Opus 5.5, has its source on GitHub without a licence and is an unofficial fan project, so it is source-available, not open source. GTM-Bench code is published by Blackpearl Group under its own GitHub organisation; Levy Street team members co-authored the paper (see GTM-Bench), but the repository is not ours to list.

Before you build on any of these

• Read the LICENSE file, or confirm there is none, before copying code
• Keep copyright and licence notices in anything you redistribute
• Treat names, logos and branding as separately protected, even in MIT projects
• Check media credits: code licences do not automatically cover art, audio or icons
• Pin versions, especially for pre-alpha software like SPAWN D
• Review AI-written code as carefully as any other contribution; our Claude Code guide covers how we do it

Written by the Levy Street team in Wellington. Last updated 19 October 2026.