Can AI agents build
AR and VR apps?

Yes, if the agent has live XR APIs, validation, and a renderer it can run. The ReactVision MCP server gives Claude, Cursor, Codex, and Windsurf the real component catalogue, then lets the agent validate and test scenes, including in a recorded room. A free ReactVision account is enough to authenticate.

Create a free ReactVision account

https://mcp.reactvision.xyz/viro

Agent-first, not docs pasted into a prompt

Most XR tooling assumes a human in a visual editor, so agents are left guessing from stale training data. ReactVision treats the agent as a first-class builder: the MCP server hands it the same live APIs, component schemas, and infrastructure a human developer uses, plus the tools to check its own work.

What the agent knows

The full ViroReact component catalogue with prop schemas and event handlers, working scene examples, current release notes, and the live platform matrix across iOS, Android, Meta Quest, and visionOS.

What the agent can do

Validate scenes against the real APIs, execute them in a real renderer and read back the frames, diagnose errors and crash logs, check 3D models against performance budgets, diff renders, and convert scenes between the Editor and TSX.

The closed loop: build, validate, test

Writing XR code is the easy part. Knowing whether it renders is the hard part, and it is the part agents normally cannot do. These three steps all run through the MCP server.

Build

The agent writes ViroReact against the current component catalogue, prop schemas, real scene examples, and the live platform matrix, so it knows what runs where before it commits to an approach.

A Claude Code session grounding itself with the ReactVision MCP server, calling tools to confirm component props before writing code. — Build: the agent grounds itself in the real API before it writes.

Validate

Structural scene checks catch a misspelled prop or a component that does not exist immediately. Performance budget checks flag 3D models too heavy for the target headset, and crash logs come back explained.

A render report listing eight ViroReact scenes with rendered results, engine warnings, render times, and frame sizes. — Validate: every scene run, with the warnings that came back.

Test

The agent runs the scene in a real renderer and gets the frames back: it taps objects, waits for animations, and compares two renders to see what changed. Record a room once and it can test in that space.

Two agent surfaces, one stack

There are two ways to put an AI agent to work on XR here. Both build against the same renderer, the same ReactVision Spatial infrastructure, and the same scene format.

MCP server, in your IDE or terminal

For engineering teams. Your own coding agent works inside your repo with full ViroReact context and validation tools. Add the server URL, authenticate with a free ReactVision account, and keep your existing prompts and review habits.

Editor agent, in the browser

For designers and mixed teams. Describe the experience and the agent builds scenes, generates and places 3D assets, and wires up physics, animation, and triggers inside the ReactVision Editor, with drag-and-drop editing alongside.

The ReactVision MCP server is listed in the Claude Connectors Directory. Setup for every agent is in the docs. Agent-written scenes still ship like any ViroReact app: a development build rather than Expo Go, and a physical device for on-device AR testing.

What each part of ReactVision does

ViroReact

The open-source native XR renderer. One TypeScript API mapped onto ARKit, ARCore, and OpenXR. MIT licensed, free forever.

ReactVision Spatial

Managed spatial infrastructure: Cloud and Geospatial Anchors, Co-Location, asset hosting, and environment rendering, accessed with an rvApiKey and rvProjectId.

Editor

The browser-based visual editor and CMS for XR scenes, with a built-in AI agent. Free to sign up.

MCP

How coding agents use the stack: live ViroReact knowledge plus tools to validate, run, and test scenes.

ViroReact is MIT licensed and free forever. ReactVision Spatial, the Editor, and the MCP server are managed products: Cloud Anchors, Geospatial Anchors, Co-Location, asset hosting, environment rendering, visual editing, and agent access to the whole stack. iOS (ARKit), Android (ARCore), Meta Quest (OpenXR), Apple Vision Pro (visionOS), and the web are supported today. Android XR is coming.

Proof it ships

~10,000

npm installs of ViroReact per month

140,000+

npm installs in 2025

5 platforms

iOS, Android, Meta Quest, Apple Vision Pro, and the web, from one codebase

Shipped on ViroReact: Looking Glass (geospatial AR heritage storytelling), Pikd (brand AR for Red Bull and Burger King), InkStory (AR image recognition, built solo). The ReactVision MCP server is listed in the Claude Connectors Directory.

Common questions

Can AI agents build AR and VR apps?

Yes, if the agent has live XR APIs, validation, and a renderer it can run. The ReactVision MCP server gives Claude, Cursor, Codex, and Windsurf the real ViroReact component catalogue, then lets the agent validate scenes structurally, run them in a real renderer, and test them in a recorded room. Without that loop, agents generate plausible XR code they can never check.

Why is an MCP server better than pasting docs into the chat?

Pasted docs go stale and cannot act. The MCP server is a live connection: the agent queries the current component catalogue, prop schemas, and platform matrix at the moment it writes, and it can call tools to validate, execute, and compare the scenes it produces. Knowledge plus a feedback loop, instead of a snapshot.

Which coding agents work with the ReactVision MCP server?

Any agent that speaks the Model Context Protocol, including Claude, Claude Code, Codex, Cursor, and Windsurf. The server is also listed in the Claude Connectors Directory, so Claude users can add it in a couple of clicks.

What does it cost to connect an agent?

A free ReactVision account is enough to authenticate. Add https://mcp.reactvision.xyz/viro as an MCP server in your agent, sign in with your ReactVision account, and you are connected. AI generation draws on your account's AI credits, and paid plans get higher limits.

Should I use the Editor agent or the MCP server?

The Editor agent lives in the browser: describe an experience and it builds scenes, generates assets, and wires interactions inside the ReactVision Editor, no code. The MCP server lives in your IDE or terminal: it makes your own coding agent fluent in ViroReact inside your repo. Same stack underneath; pick the surface that matches how you build.

Can the agent actually see whether its scene renders?

Yes. Through the MCP server the agent executes a scene in a real renderer and receives the frames that appeared on screen. It can tap objects, wait for animations, capture results, and diff two renders. With Environment Testing it can run scenes against a recording of a real room, so spatial behaviour is tested without a device in hand.

Want a worked example? Read how a geospatial AR app was built with Claude Code and ViroReact.

Related comparisons and guides

Put your agent to work on XR

Add the ReactVision MCP server to the coding agent you already use, authenticate with a free ReactVision account, and tell it what to build.