MCP server for AI agents

The MCP server (@hypersoniclabs/helix-mcp) gives an agent a structured, typed interface to the platform — discovery, validation, and publishing as callable tools — instead of scraping a CLI's stdout. It bundles the CLI as a library, so a tool call and a terminal command run the same code.

What it exposes

• Discovery — list_systems, list_abilities, list_templates and get_package_manifest, so the agent builds on what already exists instead of reinventing it, and reads a package's real config surface instead of guessing key names.
• Validation — validate_world checks a bundle against the manifest schema and platform rules before publishing; inspect_world measures the built scene's layout as numbers.
• Publishing — publish_world for worlds, and publish_item for a wearable, avatar or home item uploaded from a single .glb. list_cosmetic_slots returns the Character-Creator slot tags and genders a wearable must declare; slot and genders are both closed enumerations, so an agent cannot invent a tag or a gender.
• Docs — read these docs as structured Markdown (the same content behind /llms.txt).

Connect it

Point your agent host at the server (example for a Claude/Cursor-style MCP config):

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

Signing in stays human: the agent calls start_login, the human completes the browser sign-in, and the agent polls check_login. The minted token is stored locally by the MCP process and never enters the conversation.

Why agents like these docs

These docs are built to be machine-readable from day one:

• Every page is available as clean Markdown (the Copy Markdown button, and content routes).
• /llms.txt is a structured index of the whole site.
• /llms-full.txt is the entire corpus in one file for context windows.
• Every parameter and error is typed and described — what's obvious to a human reader is spelled out for an LLM.

Together with the MCP server, an agent can read the docs, scaffold a world, validate it, and publish it without a human in the loop.