RakuAI is the inference-only spatial runtime your model inhabits over MCP — scene graph, physics, simulation step, render targets. One JSON-Schema contract. Every vendor. No training, no weights, no integration tax. Connect via MCP and your model competes on quality of intent, not on how deep it integrated. The full versioned tool contract, with schemas and permission semantics, is available on request.
If your assistant speaks MCP, it speaks RakuAI. One JSON-Schema contract, versioned and stable: 17 native tools (5 read + 12 mutation) are the versioned public contract you pin to. They sit inside a 147-tool registry — 17 native + 7 universal passthrough + 123 subsystem-catalog wrappers — and the passthrough reaches the ~9,500 endpoints on api.rakuai.com. No custom integration, no per-vendor fork.
The contract is identical for Claude, ChatGPT, Gemini, and Copilot — identical by construction. No exclusivity. Per-vendor end-to-end conformance testing is in progress; we will publish results before claiming verified parity for any vendor.
RakuAI is a spatial runtime your model calls at inference time. We hold no model weights, run no training jobs against your data, and fine-tune nothing. Your data stays yours. This is a hard line.
Read tools (get_scene_state, capture_scene, get_metrics, xr_get_session, xr_list_anchors) are live in production. Mutation tools (set_render_target, start_simulation, and 10 others) are sandboxed and audit-logged by default.
We are in Phase 1 of our lab partnership strategy. The goal is to land a first-class MCP connector listing in one lab's tool directory and run a reproducible end-to-end demo together. No signed lab partner today — that's the target, not the status.
Any MCP-speaking assistant connects to api.rakuai.com today. The 17-tool contract is stable and versioned. Get listed in your connector directory; we maintain it.
Scan a real space with Raku Capture, then query and act on the scene through your model over MCP. The demo is reproducible and runs without proprietary tooling on either side.
Joint blog post, shared connector listing, and an "LLM Integration Guide" on-site. We do the distribution work; you get a spatial capability story to point to.
Anthropic Claude Startups membership (accepted May 2026) is the warmest first path. NVIDIA Inception (accepted May 2026) adds credibility and compute relationships. We are actively opening conversations with OpenAI, Google, and Microsoft ecosystems.
The first-lab connector path, end to end. Production is deny-by-default: the 5 read tools are live, the 12 mutation tools run sandboxed and audit-logged until a connection is explicitly granted higher tier. Nothing below is faked — it is the contract any MCP client can pin to today.
{
"mcpServers": {
"rakuai": { "url": "https://mcp.rakuai.com/mcp" }
}
}
17 native tools (5 read + 12 mutation), drift-protected in CI. Your tool picker will surface all 147 registry entries (17 native + 7 universal passthrough + 123 subsystem-catalog wrappers) — that is expected, not a bug. Pin to the 17; the rest are conveniences layered on top.
get_scene_state → reason → act. The "scan a space, then talk to it" demo is reproducible with no proprietary tooling on either side — see the LLM Integration Guide and the Claude Desktop guide.
Unity just released Native MCP. It's good for design-time scene manipulation: Claude Desktop builds a scene, Unity renders it. RakuAI's MCP runs at simulation time. Your model isn't generating a scene file — it's driving the world frame by frame. Different category, both valid.
The planned extension surface: partners ship adapters that translate model-specific outputs into the RakuAI scene contract. The adapter interface is being designed now; no live adapters exist yet.
Adapter contract design is in progress. Any model that produces frames + scene deltas will be able to become an adapter. We will publish the spec publicly before requesting implementations.
Our dev loop uses four AI vendors in parallel — every day:
We are not loyal to any single model. Our runtime is not either.
Honest status: The MCP contract is identical by construction for every vendor. Per-vendor end-to-end conformance testing — load world, read state, act, verify round-trip — is a Phase 2 deliverable that is not yet complete. We will publish results publicly. Do not take "multi-vendor" as a claim of verified parity today.
We are actively looking for a first lab connector partner.
These are the labs we are targeting — not existing signed partners. No exclusivity: the multi-vendor contract is the product. If you are building a foundation model or assistant product and want to give it a real, measurable scene over MCP — without staffing a spatial runtime team — we want to talk.
Concretely: we will list your connector, ship the LLM Integration Guide for your platform, and run a reproducible reference demo together. MCP-connection takes an afternoon; the spatial story is yours to tell from day one.
Get in touch →Use our contact form. We respond to every lab inquiry.
Microsoft for Startups
RakuAI is an NVIDIA Inception member. The same NVIDIA-accelerated spatial runtime any model drives over MCP powers Raku Capture — a phone scan of any room turns into an explorable splat in minutes. Read more →
NVIDIA Inception is a free program for AI startups; RakuAI's membership does not imply that NVIDIA endorses RakuAI's products. Learn more at nvidia.com/startups.
RakuAI is built on AWS and an AWS Activate participant. Production reconstruction runs on Azure today; AWS GPU capacity (G5 / G6e) is on our roadmap, honestly gated until quota lands.
“AWS” and “AWS Activate” are trademarks of Amazon.com, Inc. or its affiliates. RakuAI’s participation does not imply that AWS endorses RakuAI’s products.