HELIX Docs
CLI & AI tools

MCP server for AI agents

The HELIX Model Context Protocol server gives AI coding agents typed tools to discover, validate, build and publish on HELIX.

The MCP server (@hypersoniclabs/helix-mcp) gives an agent a structured, typed interface to the platform instead of scraping CLI output. It is a thin layer over the CLI: tool schemas route arguments into the same code, so a tool call and a terminal command behave the same.

Generated reference

The server registers more than a hundred tools. The MCP tool reference is generated by starting the published server and listing its tools, so names, descriptions and argument schemas match exactly what an agent sees.

Connect it

For a complete Claude Desktop walkthrough, including Windows setup and troubleshooting, see Connect Claude Desktop.

Add the server to your agent host's MCP configuration (Claude Desktop, Cursor, Codex and others):

{
  "mcpServers": {
    "helix": {
      "command": "npx",
      "args": ["-y", "@hypersoniclabs/helix-mcp@latest"]
    }
  }
}

Discovery tools work signed out. Signing in stays human: the agent calls start_login, the human finishes the browser sign-in, and the agent calls check_login. The token is stored locally and never enters the conversation.

Headless and remote hosts need token login first

start_login uses a callback on the machine that runs MCP. On CI, an SSH-only server or a container, create a client token in Account → API keys, run helix login --token "$HELIX_CREATOR_TOKEN" on that host, then start MCP. The CLI and the MCP server share ~/.helix/credentials.json.

Source upload is opt-in

publish_world uploads the playable build by default. Set uploadSource: true only when the human also wants HELIX to keep the editable project for the website editor. Otherwise the project must live in version control: an optimised published bundle is not a source backup.

Start with get_started

An agent should begin with get_started. It returns the project recipe for what is being built (kind is character, multiplayer or scene for worlds, or one of the item and asset kinds) and points to the right docs and skills. From there:

  • read_doc and read_template serve the bundled platform documentation and the worked multiplayer templates.
  • read_skill serves bundled skills for avatars, vehicles, assets, modelling, multiplayer and QA.
  • validate_world, inspect_world and capture_world_screenshot check a built world before publish_world.
  • publish_continuum_package and publish_owned_item publish Packages and move existing Items onto them.

The server carries its own versioned documentation, so for agents it is the best source for the exact version they run. This site exists for people, for search, and for the concepts that do not belong inside a tool description.

For language models reading this site

  • /llms.txt is a structured index of every page.
  • /llms-full.txt is the entire corpus in one file.
  • Every page is available as clean Markdown (the Copy Markdown button, and the content routes).

Publish an item from a GLB

Seal a .glb as an immutable Continuum Package, turn it into an Item, and optionally open a Marketplace offer. Also where wearables and other character add-ons fit.

Connect Claude Desktop

Add HELIX as a local MCP server in Claude Desktop so an AI agent can validate and publish worlds.

On this page

Connect itStart with get_startedFor language models reading this site