Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
Examples: @three-ws/vision-mcp
Two runnable examples. Both spawn this package's own MCP server over stdio (the
same node src/index.js entry point the README documents), speak real MCP
JSON-RPC to it, and hit the live three.ws vision pipeline. Neither one needs a
key and neither one costs anything.
| File | What it does | Run |
|---|---|---|
list-tools.mjs |
Prints all 3 tools with titles, annotation hints, and full input schemas. | node examples/list-tools.mjs |
read-an-image.mjs |
Probes the lane with get_vision_status, gets alt text from describe_image, then reads the text off the same image with analyze_image. |
node examples/read-an-image.mjs |
Run them from the package directory:
cd packages/vision-mcp
node examples/list-tools.mjs
node examples/read-an-image.mjs
Nothing to install and nothing to configure: the vision lane serves anonymous
callers on the free NVIDIA NIM models. The server prints a one-line banner to
stderr on connect ([[email protected]] connected over stdio with <n> tools),
which is normal.
list-tools.mjs
Runs the MCP initialize handshake, then tools/list, and formats every tool.
Expected output (abridged):
server: vision-mcp v0.1.1 (stdio)
capabilities: tools
tools: 3
1. analyze_image
title: Analyze an image against a prompt
hints: read-only, open-world
params:
- prompt (required; string, maxLength 2000)
- imageUrl (optional; string)
...
Every tool is read-only, so the hint line never says destructive.
read-an-image.mjs
The end-to-end path an agent takes when it needs to see something:
get_vision_status: is the lane live?
configured: true
image types: image/jpeg, image/png, image/webp, image/gif
max size: 12 MB
describe_image: alt text for https://three.ws/og-image.png
served: nvidia/nemotron-nano-12b-v2-vl (nvidia lane)
alt: An advertisement for a website called three.ws that allows users to create a 3D AI agent.
analyze_image: read the text in that same image (OCR through a VLM)
served: nvidia/nemotron-nano-12b-v2-vl (nvidia lane)
text: three.ws
PLATFORM
The 3D agent layer of the
internet.
...
A VLM is non-deterministic, so the exact wording changes between runs; the shape does not.
Point it at your own image with IMAGE_URL:
IMAGE_URL=https://example.com/screenshot.png node examples/read-an-image.mjs
The image must be a public https URL serving JPEG, PNG, WebP, or GIF (the vision
server fetches it, so private and loopback hosts are rejected). To read a local
file instead, base64 it and pass image + imageType to the same tools.
A busy free NIM lane can exceed its deadline; the example retries an
upstream_error / timeout / network_error up to three attempts with a short
backoff (you will see (upstream_error on attempt 1, retrying) when that
happens), then reports the tool's own error code rather than a stack trace.