AI-Assisted Development¶

SceneView is the first 3D/AR library designed for AI-assisted development. Every API is documented in a machine-readable format that AI tools understand natively.

Why this matters¶

When you ask an AI to help you build a 3D scene, it needs to know the exact API — function names, parameter types, threading rules, common patterns. Most 3D libraries have large, complex APIs that AI tools hallucinate about.

SceneView solves this with three layers:

llms.txt — a machine-readable API reference at the repo root
sceneview-mcp — an MCP server that gives AI tools full API context
Skills — the same API contract packaged for Claude Code and for the ChatGPT / Codex plugin directory

For app developers¶

Use with Claude Code¶

Install Claude Code, then either install the official plugin (recommended — bundles MCP + 11 contributor commands + cross-platform reminder hooks):

/plugin marketplace add sceneview/claude-marketplace
/plugin install sceneview@sceneview

Or add just the MCP server directly, from your project directory:

claude mcp add --scope project sceneview -- npx -y sceneview-mcp

That writes .mcp.json at the project root — the only project-scoped MCP file Claude Code reads. Claude Code does not read .claude/mcp.json or ~/.claude/mcp.json: a config placed there is silently ignored, and the server never appears. To write the file by hand instead:

{
  "mcpServers": {
    "sceneview": { "type": "stdio", "command": "npx", "args": ["-y", "sceneview-mcp"] }
  }
}

Claude Code reads .mcp.json at session start and asks you to approve the server the first time. Verify with claude mcp list — sceneview should show ✔ Connected. Use --scope user on the add command instead to enable it in all your projects (that one is stored in ~/.claude.json).

Now Claude has the full SceneView API. Ask it to:

"Add a 3D model viewer to my product detail screen"
"Add AR tap-to-place with pinch-to-scale"
"Add a dynamic sky with fog that changes based on a slider"
"Show a loading indicator while the model loads"

The AI will generate correct SceneView code — no hallucinated methods, no outdated patterns.

Use with Cursor / Windsurf / other editors¶

Copy llms.txt from the SceneView repo into your project root, or add the MCP server to your editor's MCP config. The AI tools will pick it up automatically.

Use with ChatGPT / Codex¶

The SceneView repository is an OpenAI plugin: .codex-plugin/plugin.json at its root points at the three skills under agents/ — sceneview (Compose), sceneview-ios (SwiftUI) and sceneview-web (Filament.js / WebXR). From a checkout:

codex plugin marketplace add "$PWD"    # an absolute path — a relative one does not resolve
codex plugin add sceneview@sceneview-local
codex plugin list

Codex also picks the skills up on its own from .agents/skills/ inside a checkout.

For ChatGPT, the same MCP server speaks the Streamable HTTP transport — a hosted process, not a spawned one:

npx sceneview-mcp --http        # MCP at /mcp, health at /health

It exposes the free tools plus view_3d_model, an MCP Apps widget that renders a public GLB/glTF URL inline in the conversation. Package, listing copy and submission notes live in agents/OPENAI-PLUGIN.md.

Use with Claude web / any chat¶

Paste the contents of llms.txt into your conversation, then ask your question. The AI will use the correct API.

For SceneView contributors¶

Slash commands¶

Inside the SceneView repo with Claude Code (commands shown unprefixed work locally; once you install the SceneView plugin, they're available everywhere as /sceneview:*):

Command	What it does
`/contribute`	Full guided workflow — understand the codebase, make changes, prepare a PR
`/review`	Threading, Compose API, style, module boundaries — plus `--score` (weighted eval), `--coverage` (test gaps), `high` (multi-agent triptych)
`/document`	Generate/update KDoc for changed public APIs, update `llms.txt`
`/release`, `/quality-gate`, `/sync-check`, `/store-status`, `/version-bump`, `/maintain`	Pre-PR + release lifecycle

Tip — namespace conflict: the bare /review command shadows a Claude Code built-in. With the plugin installed, prefer the prefixed form /sceneview:review to disambiguate.

Example workflow¶

cd sceneview
claude

# Then in Claude Code:
> /contribute
# Claude walks you through understanding the codebase,
# making changes, running checks, and preparing a PR.

What's in `llms.txt`¶

A 500-line, machine-readable API reference covering:

All composable signatures with parameter types and defaults
Code examples for every node type
Threading rules and common pitfalls
Resource loading patterns
Gesture and interaction APIs
Math types and coordinate system
AR-specific APIs (anchors, image tracking, face mesh, cloud anchors)

The file is maintained alongside the source code and updated with every release.

What's in the MCP server¶

The sceneview-mcp package provides tools that AI assistants can call:

get_api_reference — returns the full llms.txt content
get_node_reference — look up a specific node type's API
get_sample_code — get working example code for a use case
get_threading_rules — threading and lifecycle rules

Setup¶

{
  "mcpServers": {
    "sceneview": {
      "command": "npx",
      "args": ["-y", "sceneview-mcp"]
    }
  }
}

Works with Claude Code, Claude Desktop, Cursor, Windsurf, and any MCP-compatible tool.

Why no other 3D library has this¶

Library	AI support
SceneView	`llms.txt` + MCP server + Claude Code skills
Unity	Generic docs, frequent hallucinations on API
Sceneform	Archived, AI trained on outdated code
Raw ARCore	Low-level API, AI struggles with GL/Vulkan boilerplate
Rajawali	Minimal docs, AI has no training data

SceneView's AI tooling means faster development, fewer bugs, and correct code on the first try. This is a competitive advantage that compounds — the more developers use AI tools, the more SceneView's AI-first approach matters.