@rckflr/modelar-live

Live host for Modelar — a 3D modeling studio operated by AI agents. One process, three roles:

1. MCP server — the published modeling tools (create_model, update_model, delete_model, duplicate_model, arrange_models, get_state) run in a QuickJS-WASM sandbox; the skill is fetched from the live origin and SHA-256-verified before every load. The host holds the project state, so tool calls carry no state document.
2. Studio mirror — every state change is pushed over WebSocket to any open studio tab, so you watch the LLM build in real time in the viewport.
3. Static server — serves that studio itself at http://127.0.0.1:12416/ (loopback, auto-connects to the mirror).

Use

npx -y @rckflr/modelar-live
# then open http://127.0.0.1:12416/

Wire an MCP client to the same process (shared state, same mirror):

[mcp_servers.modelar-live]
url = "http://127.0.0.1:12416/mcp"
startup_timeout_sec = 30

(Codex also accepts stdio: command = "npx", args = ["-y", "@rckflr/modelar-live"] — that session gets its own host, without a shared mirror.)

Controls (in the studio's Headless live card)

• Pausar / Reanudar — hold the LLM's next tool call
• Deshacer última — revert the LLM's last applied tool call
• Enviar al headless — push the studio's current project into the session

Environment

• MODELAR_LIVE_PORT (default 12416) — HTTP + WS port
• MODELAR_ORIGIN (default https://modelar.rckflr.workers.dev) — origin whose llms.txt skill is loaded and whose studio is proxied when no local build exists

Rendering stays in the browser (WebGL): headless builds are real, and the live mirror is where you see them.