@three-ws/avatar-mcp

@three-ws/avatar-mcp

An MCP server that drops a live, rotatable 3D avatar into any chat — interactive in the conversation, embeddable anywhere. Free, no API key.

Install · Setup · Quick start · Tools · Requirements · three.ws

A thin, zero-config, read-only Model Context Protocol server that brings three.ws 3D avatars into any MCP client — Claude Desktop, Claude Code, Cursor, or any other host. Render a live, rotatable avatar inline, get a paste-anywhere embed iframe, or fetch avatar metadata. Every tool reads live from the real three.ws endpoints — no mock data. Public and unlisted avatars need no API key. Registry name: io.github.nirholas/threews-avatar. Built by three.ws.

Need wallets, voice, generation, or pump.fun powers? See the sibling package @three-ws/avatar-agent, a full 3D AI agent in a box.

Install

npm install @three-ws/avatar-mcp

Run it directly with npx (no install needed) or install globally for the avatar-mcp CLI:

npx -y @three-ws/avatar-mcp           # MCP stdio server
npm install -g @three-ws/avatar-mcp   # exposes `avatar-mcp`

Setup

Claude Code, one line:

claude mcp add threews-avatar -- npx -y @three-ws/avatar-mcp

Claude Desktop / Cursor (JSON config):

{
	"mcpServers": {
		"threews-avatar": {
			"command": "npx",
			"args": ["-y", "@three-ws/avatar-mcp"]
		}
	}
}

No environment variables are required. To read avatars from a different host (e.g. a preview deployment), set THREEWS_BASE_URL. Restart your client after editing the config.

Inspect the tool surface in a GUI:

npx -y @modelcontextprotocol/inspector npx -y @three-ws/avatar-mcp

Quick start

Once connected, ask your client in plain language:

Render avatar bf58f7fc-64b3-4a92-8c27-dd472fc04b58 in the chat, dark background, auto-rotating.

(That UUID is a real public avatar. Any avatar's id is the last path segment of its three.ws URL, and GET https://three.ws/api/avatars/public lists more. A @handle works too, for any three.ws user who has published a public avatar.)

render_avatar returns three things so it looks great in every client:

• a preview image that renders inline everywhere,
• an interactive text/html resource — a real <model-viewer> you can orbit and zoom, for hosts that render HTML resources, and
• the embed URL + iframe to drop the live avatar into any page.

Tools

All three tools are free, read-only, and annotated (readOnlyHint, idempotentHint, openWorldHint) so hosts can run them without confirmation prompts. There is no x402 charge. Identify an avatar by id (UUID), handle (a three.ws username, with or without the @, that has a public avatar), or a raw model GLB URL.

Tool	Selectors	What it does
render_avatar	id · handle · model	Renders the avatar inline. On MCP Apps-capable hosts it shows a live, rotatable 3D model right in the chat; other clients get a preview image + embed URL. Params: background (transparent/dark/light), scene (portrait/headshot/upper-body/full-body), auto_rotate, height (160–1080).
avatar_embed_code	id · handle · model	Returns a ready-to-paste <iframe> that embeds the live avatar anywhere — as easy as a YouTube embed. Params: background, width (CSS), height (160–1080), idle (loop the idle animation), overlay (chrome-free mode for OBS/overlays).
get_avatar	id · handle	Fetches avatar metadata: name, GLB model_url, owner, visibility.

Interactive 3D in the chat — the differentiator

render_avatar is an MCP App (SEP-1865): it declares a ui:// resource that supporting hosts render in a sandboxed iframe — a real, orbit-and-zoom <model-viewer>, not a static image. The avatar is live in the conversation: rotate it, zoom it, watch it idle, without leaving the chat. Hosts without MCP Apps support still get a rendered preview image and a one-tap live embed, so the tool degrades gracefully everywhere.

Example calls

// render_avatar — live, rotatable avatar in chat
{ "id": "bf58f7fc-64b3-4a92-8c27-dd472fc04b58", "background": "dark", "auto_rotate": true }

// avatar_embed_code — paste into any website
{ "id": "bf58f7fc-64b3-4a92-8c27-dd472fc04b58", "height": 560 }

// get_avatar: metadata, by @handle instead of id
{ "handle": "@your-handle" }

get_avatar on that id returns, live:

{
	"id": "bf58f7fc-64b3-4a92-8c27-dd472fc04b58",
	"name": "Jade",
	"slug": "jade-fe1a51",
	"model_url": "https://pub-2534e921bf9c4314addcd4d8a6e98b7b.r2.dev/u/299bdfa7-d848-4865-bdb1-0db23fc3b431/jade-fe1a51.glb",
	"thumbnail": null,
	"visibility": "public"
}

Every tool also advertises an outputSchema describing its structuredContent, so typed clients can consume results without re-parsing the text blocks.

Prompts

Prompt	Arguments	What it does
showcase-avatar	avatar — id (UUID) or @handle	One-step showcase: renders the live, rotatable avatar inline and produces the embed iframe, ending with a copy-paste summary (name, viewer URL, embed URL, <iframe> snippet).

How it works

Each selector maps to a real three.ws endpoint. Raw model URLs must be https:// (or http://localhost for dev); other schemes are rejected.

Selector / asset	Endpoint
id	GET https://three.ws/api/avatars/:id
handle	GET https://three.ws/api/users/:handle/avatar
preview image	GET https://three.ws/api/avatar/render?avatar=:id
live embed	https://three.ws/avatar-embed.html?...
viewer	https://three.ws/viewer?src=:glb

Requirements

• Node >=20.
• No credentials. Public and unlisted avatars need no API key.

Variable	Required	Notes
THREEWS_BASE_URL	Optional	three.ws host to read from. Defaults to https://three.ws.
THREEWS_TIMEOUT_MS	Optional	Per-request network timeout. Defaults to 30000.

Links

• Homepage: https://three.ws
• Sibling package: @three-ws/avatar-agent — full 3D AI agent (wallet, voice, pump.fun)
• Changelog: https://three.ws/changelog
• Issues: https://github.com/nirholas/three.ws/issues
• License: proprietary, see LICENSE

_{Part of the three.ws SDK suite — 3D AI agents, on-chain identity, and agent payments.
Website · Changelog · GitHub}