Honest status (as of May 2026)

Surface	Status	Notes
MCP server (17 native tools + passthrough)	Production	Running on Azure Container Apps (East US). Available to connect now.
Local stdio transport (Claude Code, Cursor, VS Code)	Production	Verified end-to-end. Clone raku-runtime, add the JSON block, restart — tools appear.
Claude Desktop (macOS / Windows)	Connect via MCP	Same config block as Claude Code. Client-level e2e verification still in progress; connection method is standard and expected to work.
ChatGPT (OpenAI)	Connect via MCP	ChatGPT MCP plugin connector available in ChatGPT Plus / Team. Point it at the hosted MCP URL + your bearer token. Client-level e2e verification in progress.
Gemini / Google AI Studio	Connect via MCP	MCP tooling available via Google AI Studio extensions. Client-level e2e verification in progress.
GitHub Copilot (VS Code)	Connect via MCP	VS Code MCP support added in VS Code 1.99+. Config block documented below. End-to-end verification in progress.
Raku Capture live viewer	Early access	Splat reconstruction pipeline gated on GPU credit approvals (Azure A10, AWS G5). The scene query tools work against existing captures now.

What your assistant gets access to.

The headline contract is 17 native tools — a versioned, stable API surface integrators can pin to. A tool picker may show many more (147 is expected, not a bug): the registry also carries 7 universal passthrough tools and 123 subsystem-catalog wrappers. The 17 native tools are the ones to build on.

5 read-only tools — zero mutation risk

scene_query — spatial distance, occupancy, free-space extents
scene_describe — object list, pose deltas, area summaries
capture_list — enumerate captures attached to your account
capture_status — check reconstruction progress
capture_open — load a capture into the session context

12 mutation tools — audited, validated

capture_create — start a new capture session
xr_create_anchor — place a named spatial anchor
xr_list_anchors — enumerate anchors in the scene
xr_remove_anchor — remove an anchor by ID
xr_get_session — inspect the active XR session
load_world_model — attach an adapter to the runtime
ingest_frame — feed a world-model frame
set_render_target — configure the output surface
start_simulation — begin the simulation loop (sandbox / dev only)
get_scene_state — full scene-graph snapshot
get_metrics — runtime performance snapshot
+ 1 additional (see mcp.html for the full list)

Universal passthrough: api_call / api_search reach the full ~9,500-endpoint raku-api surface; engine_call / engine_search reach the C++ engine API. These passthrough tools are available on every tier and let your assistant reach anything in the platform that isn’t yet wrapped as a native tool. They are the “escape hatch” — use the 17 native tools first for stability.

Connect your assistant.

All four clients connect over MCP. The two transports are stdio (local — your editor spawns the server as a child process) and Streamable HTTP (remote — point at the hosted endpoint with a bearer token). The steps below use stdio for local clients and HTTP for hosted ones.

◄

Claude (Desktop & Code)

Verified (Code) Connect via MCP (Desktop)

Both clients use the same mcpServers JSON block. Claude Code reads ~/.claude.json; Claude Desktop reads claude_desktop_config.json.

Clone raku-runtime: git clone https://github.com/RakuXR/raku-runtime.git
Install dependencies: pip install 'mcp>=1.2.0' fastapi pydantic jsonschema requests httpx
Add this block to ~/.claude.json (Claude Code) or claude_desktop_config.json (Claude Desktop — macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

// ~/.claude.json or claude_desktop_config.json { "mcpServers": { "raku": { "command": "python", "args": ["-m", "src.mcp.stdio_server"], "cwd": "/absolute/path/to/raku-runtime", "env": { "PYTHONPATH": "/absolute/path/to/raku-runtime", "RAKU_MCP_MODE": "sandbox", "RAKU_API_BASE": "https://api.rakuai.com" } } } }

Restart the client. The Raku tools appear in the tool picker. Verify with: "Use raku get_metrics to show me what’s running."

●

ChatGPT

Connect via MCP — e2e verification in progress

ChatGPT Plus and Team support MCP server connectors. You’ll need a bearer token from your RakuAI account and the hosted MCP endpoint.

Log in to chatgpt.com (Plus or Team plan required for MCP).
Open Settings → Connected apps / MCP.
Add a new MCP server: URL https://mcp.rakuai.com, transport Streamable HTTP.
Paste your RakuAI bearer token when prompted.
Save and confirm. The Raku tools appear in the tool selection panel.

Note: ChatGPT’s MCP UI is evolving. If the UI has changed since this was written, the underlying protocol is stable — point any Streamable HTTP MCP client at https://mcp.rakuai.com with your bearer token.

▲

Gemini / Google AI Studio

Connect via MCP — e2e verification in progress

Google AI Studio supports MCP tool servers via its extensions / function-calling surface. Gemini Advanced in Google Workspace also supports MCP.

Open aistudio.google.com.
In the model settings panel, find Tools → Add MCP server (or equivalent in your Gemini client).
Enter the hosted endpoint: https://mcp.rakuai.com, transport Streamable HTTP.
Authenticate with your RakuAI bearer token.
Enable the server. The 17 native tools appear as callable functions in your session.

Note: Gemini’s MCP client is under active development. The hosted endpoint (https://mcp.rakuai.com) is a standard Streamable HTTP MCP server and will work with any conformant Gemini MCP client.

⬢

GitHub Copilot (VS Code)

Connect via MCP (VS Code 1.99+)

VS Code 1.99+ added MCP server support in Copilot Chat (Agent mode). Add a workspace or global MCP config file:

Clone raku-runtime and install dependencies (same as Claude above).
Create or edit .vscode/mcp.json in your project:

// .vscode/mcp.json { "servers": { "raku": { "type": "stdio", "command": "python", "args": ["-m", "src.mcp.stdio_server"], "cwd": "/absolute/path/to/raku-runtime", "env": { "PYTHONPATH": "/absolute/path/to/raku-runtime", "RAKU_MCP_MODE": "sandbox" } } } }

Open Copilot Chat in Agent mode. The Raku tools are available for selection. Ask: "Use the raku get_scene_state tool and tell me how many nodes are in the scene."

Common troubleshooting

Symptom	Cause / fix
No Raku tools appear	Check `cwd` and `PYTHONPATH` point at the repo root. Run `python -m src.mcp.stdio_server` by hand — it should log `tools registered (stdio): 147`.
Client shows ~147 tools, not 17	Expected, not a bug. The 17 native tools are the stable contract; the registry also carries 7 universal passthrough + 123 subsystem-catalog wrappers.
Tool calls return `{"status": "denied"}`	Server is in `production` mode (read-only). Set `RAKU_MCP_MODE=sandbox` in the `env` block to enable mutation tools for local dev.
`ModuleNotFoundError: src`	`cwd` / `PYTHONPATH` not set to the repo root. Both must point at the directory containing the `src/` package.
Garbled or no responses (stdio)	Something is writing to stdout and corrupting the JSON-RPC stream. The server keeps stdout clean (all logs go to stderr). Check for any debug `print()` calls in local code.

What you can ask once connected.

These prompts work against any Raku capture once the MCP connection is live. Start with the read-only tools — they carry zero mutation risk and return structured spatial data your assistant can reason about directly.

"What objects are within 2 metres of the front door?"

Uses scene_query. Returns a list of objects with distances and confidence scores. Good first test after connecting.

"Measure the distance between the sofa and the TV."

Uses scene_query with two anchor targets. Returns distance in metres as structured data, not a prose estimate.

"What’s in this room? Give me a count by category."

Uses scene_describe. Returns object inventory grouped by type — furniture, fixtures, equipment.

"List all my captures and tell me which ones are fully processed."

Uses capture_list + capture_status. No mutation — pure inventory query.

"What changed in this space since last Tuesday?"

Uses scene_describe with a time filter. Returns a delta record: moved, added, and removed objects since the reference timestamp.

"Place a named anchor at the kitchen counter and tell me its pose."

Uses xr_create_anchor (mutation tool). The runtime validates the pose against physics before applying. The anchor persists in the scene.

These are illustrative. The 17 native tools + passthrough surface cover the full scene management lifecycle. For the complete tool reference and JSON schema, see mcp.html.

Authority model: assistants propose, the runtime decides.

Nothing silently rewrites your scene. The RakuAI runtime is the authority over physics, collision volumes, transforms, and canonical scene state. An AI assistant’s mutation tool calls are proposals; the runtime validates every one before applying it.

What the runtime owns (non-negotiable)

Physics simulation and collision geometry — no caller bypasses the collision system without an explicit override flag
The canonical scene-state record — all queries answer against this; only the runtime writes to it
Anchor lifecycle and pose authority — created, updated, and removed only through audited MCP calls
Session permissions — what an MCP client can read vs. mutate is set at session creation, not by the caller

What an AI assistant proposes

Mutation tool calls are proposals; the runtime validates against physics and permissions before applying
If a proposed transform collides, the runtime rejects it and returns an error — no silent success
The 5 read-only tools carry zero mutation risk by construction
Multi-step action sequences are validated step by step, not batched

What is audited

Every MCP tool call is logged: caller identity, timestamp, input payload, outcome
Scene-state changes carry a delta record: what changed, when, which session
Logging happens at the runtime layer — it cannot be suppressed by any LLM adapter or passthrough path
The audit log is accessible to account holders for their captures

Practical note: start with the 5 read-only tools. Your assistant asks questions; the runtime answers with structured data; nothing in your scene changes. When you’re ready to test mutation (anchors, session control), those calls go through the same validated, audited path.