Cavalry Bridge for Claude

Let Claude see, edit and preview your Cavalry scene: read layers and connections, change attributes with JavaScript, hook nodes together, switch Duplicator distributions, and render a frame to check the result.

English | Русский

Install (2 steps, no terminal)

1. Download cavalry-claude-bridge.mcpb from the latest release and double-click it. Claude Desktop opens an install dialog: click Install. (Or: Claude Desktop → Settings → Extensions → drag the file in.)

2. In Cavalry run Scripts → Stallion. The window should say Listening on 127.0.0.1:8080.

Now ask Claude: "Show me my Cavalry scene." Claude asks for permission before each tool call.

You do not need to install Node.js or edit any config file: Claude Desktop ships its own Node.js, and the .mcpb contains the whole server in one file.

Related MCP server: cavalry-mcp

Requirements

What Claude can do

Try: "Put a Fibonacci duplicator on the circle and add a Random to the clone positions", "Which layers are connected to the Duplicator?", "Render the frame and tell me what looks off".

Spend fewer tokens

• Name layers by id (duplicator#1) when you can, so Claude does not need to search.

• Claude reads the API notes with cavalry_help on demand instead of you pasting them into the chat.

• Renders default to 50% size; ask for a bigger one only when you need detail.

Troubleshooting

Safety

Claude can run any JavaScript inside Cavalry, including deleting layers. This isn't limited to the scene: Cavalry's own scripting API can make outbound network requests, open arbitrary URLs or local file:// paths, and run its own HTTP server — all from your machine. Save your scene before letting it make big changes, and read the permission prompts. Stallion itself listens on 127.0.0.1 only; run it when you need it.

Use with other MCP clients (Claude Code, Cursor, ...)

The .mcpb is a zip file: unzip it and use server/index.js, or build from source (below).

{
  "mcpServers": {
    "cavalry": {
      "command": "node",
      "args": ["/absolute/path/to/server/index.js"],
      "env": { "CAVALRY_PORT": "8080" }
    }
  }
}

How it works

Claude Desktop ──stdio (MCP)──> server/index.js ──HTTP POST──> Stallion (in Cavalry, 127.0.0.1:8080)
                                       ^                              │ runs the script, then
                                       └──────── temp JSON file <──── api.writeToFile(result)

Stallion accepts scripts but does not return their result, so every script is wrapped: it runs inside try/catch, captures console.log, and writes a JSON result file that the server reads back. Cavalry's API throws plain strings, which the wrapper turns into readable errors.

Build from source

git clone <this repo> && cd cavalry-claude-bridge
npm install
npm run build      # type-check + bundle to server/index.js
npm test           # end-to-end test against a fake Cavalry
npm run pack       # dist/cavalry-claude-bridge.mcpb

Releasing: bump version in manifest.json and package.json, then push a tag vX.Y.Z. The workflow builds, tests and attaches the .mcpb to a GitHub release.

The Cavalry API notes live in knowledge/api-notes.md (split into topics by ## headings and served by cavalry_help). Corrections are welcome: every line should be something you verified on a real scene.

Optional: extras/skill is a Claude Skill with the same notes, for plans that support Skills. The extension does not need it.

Credits

• Stallion by scenery.io: the bridge inside Cavalry.

• kacperchlebowicz/Cavalry-mcp: an earlier Cavalry MCP server (MIT) that this project started from.

Not affiliated with Scene Group (Cavalry) or Anthropic.

License

MIT, see LICENSE.