17 native MCP tools (5 read-only + 12 mutation) plus universal api_call/api_search passthrough to ~9,500 REST endpoints. Production-deployed since March 2026 — now with a versioned public tool contract, a 5-minute read-only connect path, and deny-by-default, audit-logged permissions.
Your AI assistant speaks MCP. RakuAI is the runtime on the other end of that connection, driving real spatial output to glasses, headsets, and the browser.
The full native MCP surface: 5 read-only + 12 mutation tools. Schema-validated, audit-logged, permission-gated. Plus universal api_call/api_search passthrough to ~9,500 raku-api REST endpoints (see spatial-engine.html for the full inventory). The 17-tool surface is now pinned as a versioned public contract (tools.v1.json, schema_version 1) with an anti-drift test that fails CI if the published surface diverges from the live registry.
| Tool | What it does | Permission |
|---|---|---|
get_scene_state(include_physics, include_transforms) |
Read-only scene-graph snapshot | Read-only, all modes |
capture_scene(...) |
Capture a render-time observation (color, depth, segmentation) | Read-only, all modes |
get_metrics() |
Performance telemetry (frame time, draw calls, memory, latency) | Read-only, all modes |
xr_get_session() |
Inspect headset / glasses session state | Read-only, all modes |
xr_list_anchors() |
List spatial anchors in the session | Read-only, all modes |
add_scene_node(...) |
Insert an entity into the scene graph | Mutation, sandbox/dev |
update_scene_node(...) |
Change properties on an existing node | Mutation, sandbox/dev |
set_node_transform(...) |
Position / rotation / scale | Mutation, sandbox/dev |
remove_scene_node(...) |
Delete a node | Mutation, sandbox/dev |
load_world_model(adapter_name, config) |
Bind a world-model adapter (Runway, Veo, Marble, Genie, content-pack) | Mutation, sandbox/dev |
ingest_frame(adapter_name, frame_data, frame_index, timestamp) |
Feed a world-model output frame into the runtime | Mutation, sandbox/dev |
set_render_target(target_type, config) |
Configure output (WebGL/VR/native/offscreen) | Mutation, sandbox/dev |
start_simulation(tick_rate, max_duration, realtime) |
Step physics + world model under sandbox mode | Mutation, sandbox/dev |
xr_set_head_pose(...) |
Drive head pose (sim, recording, or remote) | Mutation, sandbox/dev |
xr_simulate_input(...) |
Inject controller / hand-tracking events | Mutation, sandbox/dev |
xr_create_anchor(...) |
Anchor a node to world coordinates | Mutation, sandbox/dev |
xr_remove_anchor(...) |
Release an anchor | Mutation, sandbox/dev |
17 native tools, grouped by what they're allowed to do. Swipe or use the arrows.
If your model speaks MCP, it speaks RakuAI. Point an existing MCP client at the runtime and get a working read-only session — the 5 read-only tools — on a free sandbox. No model lock-in, no engine SDK, no mutation risk.
raku-mcp in your MCP client config (Claude Desktop, or any MCP-speaking assistant).
get_scene_state and get_metrics to see the runtime respond. That delivers immediate value.
Step-by-step: the connect-read-only quickstart, the LLM integration guide, and a drive-a-scene walkthrough. On-site: the Claude Desktop guide.
The contract is identical whether the caller is Claude, ChatGPT, Gemini, or Copilot — that is the wedge. We publish a parity matrix that tracks each vendor against all 17 native tools plus passthrough. Parity holds at the protocol level; client-specific schema handling is tracked openly as a living artifact.
| Vendor | 17 native tools | Passthrough | Status |
|---|---|---|---|
| Claude | Protocol-level | Protocol-level | Published |
| ChatGPT | Protocol-level | Protocol-level | Published |
| Gemini | Protocol-level | Protocol-level | Published |
| Copilot | Protocol-level | Protocol-level | Published |
Honest note: the matrix is published, not yet independently verified. No vendor has been run through a full 17-tool + passthrough conformance sweep on the record — per-client end-to-end verification is open Phase 1 work, tracked in the matrix itself.
What lets a security owner approve an autonomous agent acting in space? Mutation is gated by design, every call is logged, and the relay enforces tiered limits.
/api/raku/* surface, redaction-safe.
Honest note: the audit export is relay-scoped and account-wide today, not yet per-connector; quota enforcement is active where the relay's counter store is configured. Per-connector granularity and signed audit logs are roadmap.
The 17 tools above are one versioned contract. Same names, same schemas, same permission semantics whether the caller is Claude, ChatGPT, Gemini, or Copilot. The spatial web is standardizing in the open, and this contract is where RakuAI plugs any assistant into it.
We share the full contract with teams evaluating an integration: every tool schema, the permission and audit semantics behind each call, error behavior, and the versioning rules we hold ourselves to. Tell us what you are building and we will send it over.
Unity MCP (just released): scene manipulation, asset placement, build automation — Claude Desktop generates static scenes at design time.
RakuAI MCP (production since March 2026): world-model orchestration — any LLM streams intent into the simulation step every frame.
Unity is the tool your AI picks up. RakuAI is the body your AI inhabits.
The contract is the 17 native tools. The adapters are how partners plug in.
The adapter contract is public. Today the proven adapters are first-party / reference; a broad third-party adapter ecosystem is a roadmap goal, not the current state.
tools.v1.json with an anti-drift test, so what you integrate against won't silently change.We label what has shipped and what hasn't — honest framing is a feature, not a footnote.
tools.v1.json) + anti-drift CI testapi_call/api_search passthrough to ~9,500 endpointsPick a path.