Cyclotron by videogame.ai

A free, open-source game editor by videogame.ai, built from Blender 5.2 compiled to WebAssembly, Code-OSS, three.js, TypeScript and React.

On Windows, open PowerShell and paste:

irm https://cyclotron.videogame.ai/install.ps1 | iex

On a Mac with Apple Silicon, or Linux on x64, open Terminal and paste:

curl -fsSL https://cyclotron.videogame.ai/install.sh | sh

• Nothing else to install. The line uses your Node.js 24 or later if you have it, or downloads Node.js 24 from nodejs.org (checked against its published SHA-256) into ~/.volter/node. It makes the starter racing game at ~/Cyclotron/my-race and opens it in your browser; paste the same line again to reopen it. With Node.js 24 already, the same is npx @volter/cyclotron create my-race --template playable (on Windows, in Command Prompt: PowerShell's default policy refuses npm's npx.ps1). To change the editor itself, see Building from source.

• To try it without installing, open cyclotron-web.videogame.ai: the editor with Blender and Play in your browser, without Chat.

• Opens Canyon Comet, a kart race through a desert canyon modelled in the editor. Press Play in the Game panel below the viewport: the camera moves from your editing view to the chase camera behind your kart. Press RACE! (or Enter, or an arrow key) and, after a three-second countdown, race five rivals over three laps: arrow keys or W/S to drive, Space to drift, Shift to boost with collected coins. AUTO in the HUD, or Autoplay in the Game panel, lets a bot drive your kart. Escape moves the camera back and returns the untouched model. A lap takes about 12 seconds.

• src/models/canyon.blend is the scene and src/models/canyon.py, with the step scripts beside it, the bpy that built it; src/models/canyon.play.ts is the script Play runs, and src/ui/ is the React interface drawn over it.

• In the Chat pane, one click opens Sign in with ChatGPT, or Chat runs the coding agent you already have (Codex or Claude Code; others from its agent picker). The editor adds no account and bills nothing.

• Three first things to ask the agent, one per file:

  Write src/models/palms.py to plant five palm trees on the sand outside the start straight, clear of the road. Open Canyon, run the script through the project's Blender MCP, and save canyon.blend.
  

  Edit src/models/canyon.play.ts so a coin boost lasts 2 seconds instead of 1.4 and raises top speed to 38 metres per second instead of 33. Keep drift boosts unchanged.
  

  Edit src/ui/race-hud.tsx to show your speed in km/h at the bottom centre, above the controls line. Keep the circuit map on the right.
  

• npx @volter/cyclotron create my-models makes a plain modelling project with a cube.

For modelling, run npx @volter/cyclotron with no arguments. Inside an existing project it opens that project. Elsewhere it creates a ready cube project at ~/Documents/Volter Models/Untitled Model and opens it with Chat alongside the viewport. Later launches reopen that starter with your edits intact; an occupied unrelated folder is preserved and a numbered folder is used instead.

Chat and your agent

Chat uses the agent already signed in on your machine: it starts Codex or Claude Code by itself when it finds one installed and signed in, resuming the project's last conversation with the agent that held it. With none, Chat shows one button, Sign in with ChatGPT. It installs OpenAI's Codex from its official npm package (@openai/codex) when Codex is missing, then opens Codex's ChatGPT sign-in in your browser, and the chat opens once you're signed in (docs/CHAT-WELCOME.md). Until a coming Chat release removes it, the welcome also shows Sign in with Claude, which does the same for Claude Code (@anthropic-ai/claude-code). Chat checks readiness and reconnects without a reload. Other agents Volter Harness supports (Gemini, Grok and more) can be chosen from Chat's agent picker.

For native Codex or Claude Code Chat in a trusted workspace, set volter.chat.runtimePolicy to yolo in this project's Workspace Settings (the project's .vscode/settings.json). On the next start or resume, the native workbench asks you to enable YOLO for that project. Automatic workspace trust does not answer this prompt. Your confirmation stays in your user profile, bound to the folder and settings file; changing that file asks again. Declining, missing confirmation or an older workbench keeps ordinary approvals. It applies on the next start or resume, including configured model and inference launches. YOLO runs without approval prompts and removes Codex's sandbox; the Chat permission picker shows the policy carried by its opened runtime. Set the workspace value to default to use ordinary Ask/Auto approve choices on the next start or resume. A user-wide value does not enable project YOLO. An older runtime that cannot advertise YOLO enforcement is refused by name. This source change requires the paired Supercode change; publication and live build-agent qualification remain open.

If Enter is pressed before Chat has an agent and available model, Chat explains why nothing was sent and keeps the draft. Choose the agent, then send again (source and qualification).

The permission picker beneath the composer shows its selected mode throughout the conversation, including when the toolbar compacts. It stays ahead of optional provider controls on that row. This changes the native control's presentation; the harness and conversation still determine its permission choices and defaults (source and qualification).

Other extensions install from the Extensions view; Claude Code's official extension is @id:Anthropic.claude-code. No agent extension is bundled.

Every Cyclotron project declares the Blender MCP server in both .mcp.json and .codex/config.toml, with the same command. Codex loads it once you trust the folder in Codex. Creating a project never writes to your global ~/.codex.

Native Codex Chat keeps Codex's network permissions. Reaching the editor's local server requires networking allowed by the person's Codex configuration; the editor adds no network override. See the networking policy.

New projects include AGENTS.md with the model workflow and CLAUDE.md importing it. Both agents receive the same guidance for inspecting the live scene, updating scripts, using Blender MCP and verifying the saved model. The defaults also require an in-scene supplied or generated reference, regular visual comparison, live gameplay with an autoplay/demo controller, and React game UI shown in the editor's UI canvas and in Play. The agent derives appearance and requirements from supplied images, videos, specifications and existing assets. For a new game, it first matches a static scene and React UI screenshot to the reference, including framing and display colors, before implementing gameplay. These workflows come from the project defaults rather than a detailed user prompt.

Another agent can drive the conversation already visible in Chat, from the project directory:

npx --no-install cyclotron chat status
npx --no-install cyclotron chat send "Make the cube blue with softly rounded edges."
npx --no-install cyclotron chat stop

These commands use the native Chat and its Supercode runtime, retain the person's draft, and keep prompts and replies in the same transcript. send requests a native Chat submission and returns submissionRequested: true with submissionConfirmed: false. This action returns before submission settles, so its receipt does not establish that a turn started. status reports the native session ID, busy state and pending requests. Finish approvals in Chat; sending another prompt while a turn or request is pending is refused. Choose the agent/model or complete sign-in in the editor before sending when Chat is not ready.

Working in a project

cd my-race
npm run dev

From the project directory, npx --no-install cyclotron provides (the unscoped npm package cyclotron is not ours; --no-install runs the project's own copy):

• edit opens the project;
• status reports the session;
• console reports unresolved diagnostics;
• eval drives the editor's automation API;
• blender-mcp serves the Blender MCP interface over stdio;
• close stops the session.

edit prints the editor's address, http://editor-<id>.localhost:<port>/?project=<name>: each project has its own host (a stable hash of its folder), and http://127.0.0.1:<port>/ redirects there. A tool that drives the editor in a browser binds that printed host and reuses its tab; binding 127.0.0.1:<port> opens a fresh tab each time, the editor yields the previous one, and tabs pile up. Running edit again prints the same address and focuses the open tab. See CHANGELOG.md (0.5.190).

MCP initialization does not start Blender; the first scene request attaches to or opens the project's editor. Append --existing-session after blender-mcp to require an editor that is already open.

Modelling edits save to the project's .blend file. Undo and Redo go through the editor's history, and Blender keeps the snapshots; redo never reruns a script. History lasts for the session and resets when another .blend opens. An arbitrary Python execution counts as an edit, even one that fails part-way, so use the inspection tools for read-only queries.

To move a project to a new release, from whichever version it is on, run this in its folder:

npx @volter/cyclotron@latest upgrade

It moves the project's @volter packages and volter.project.json's engine.version together, and prints what to run next (npm install, then reopen the editor). The product then downloads the matching workbench. Name @latest: in a project, a plain npx @volter/cyclotron runs the project's own installed version, and a project made before 0.5.203 needs the newest release's upgrade to move it to the renamed packages and the editor/ folder.

Package map

The npm scope is @volter. Install one product; its supporting packages come with it.

Blender and Code-OSS are separate forked repositories, pinned by this repository's build configuration. ARCHITECTURE.md describes how the pieces fit.

Building from source

With Node.js 24 and npm:

After changing the Zod settings document, run npm run generate-schema, then node scripts/workbench/generate-settings.mjs --write, to derive both settings artifacts.

npm ci
npm run build
npm run typecheck
npm test
npm run check:release
npm run check:packed-imports

To rebuild after a change, name the packages it touched: node scripts/build-release-packages.mjs release/modeling.json @volter/editor-core @volter/cyclotron. scripts/workbench/build-release.mjs builds the Code-OSS workbench. The current source pin includes the shared Chat draft cache correction; that page records its source and remaining Windows qualification.

Only the packages listed in release/modeling.json and release/game.json are published, from the publish branch after a main commit is promoted to it.

License

Licenses vary by component; see LICENSE.md and each package's license and notice files.