Supercharge your AI in the real world.
Starter projects for the RakuAI runtime over the 17-tool native MCP surface. Wire your AI assistant, then iterate against a real engine. Multi-vendor by design — Claude, ChatGPT, Gemini, Copilot.
Each pattern below shows one way an external agent can drive the runtime over MCP — the same 17 native tools (get_scene_state, add_scene_node, load_world_model, … — 5 read-only + 12 mutation, see mcp.html) cover everything from read-only inspection to full simulation control. Read-only patterns work in production; mutating patterns are sandbox/dev only, deny-by-default.
Use get_metrics and get_scene_state from any client — Claude Desktop, Cursor, custom GPTs, in-house agents. Safe in every environment, including production.
PRODUCTION-SAFE
Register a world-model backend (video predictor, neural radiance field, physics foundation model, content pack) with load_world_model. The adapter speaks the standard contract; the engine retains authority.
SANDBOX / DEV
Push generative-model frames into the scene graph with ingest_frame. Visuals come from the adapter; gameplay (physics, collision, scoring) stays with the engine. Neither side has to know about the other except through the MCP boundary.
SANDBOX / DEV
set_render_target configures where the world renders: WebGL, VR headset, native window, offscreen. The same scene, the same authority, different surfaces.
SANDBOX / DEV
start_simulation begins the simulation loop with a configurable tick rate. Refused in production by design — external agents can ask, not tell. Sandbox and dev environments accept it.
SANDBOX / DEV
Compose the 17 native tools however you need from your own agent. The MCP server enforces deny-by-default permissions per tool per caller and writes a full audit log. Build on the boundary, not through it.
PRODUCTION-SAFE (read-only)
Any MCP-capable client can reach the runtime. Today: Claude Desktop (wiring guide live). In flight: Cursor, GitHub Copilot in agent mode, custom GPTs over the MCP HTTP transport. The MCP surface is model-agnostic by design — every frame, every NPC, every interaction routed through the AI of your choice.
The adapter contract on GitHub documents how to plug a world-model backend into the runtime over the MCP boundary. Full server, adapter, and scene-state schemas are available with SDK access under NDA.